Cover page

Image Name: The Moon.

The airis4D is setting up a Science Centre and Science Park to promote research interests in kids of all ages.

The program is supported by Ninan Philipose Foundation. Ninan Philipose was a primary school headmaster

in Thelliyoor school where airis4D is located. The foundation set up by his family and friends wants to carry

on his legacy. The image was taken when a 10 inch telescope that is being installed at the science park received

its ﬁrst light. The facility will be open to public by the end of March 2024.

Managing Editor Chief Editor Editorial Board Correspondence

Ninan Sajeeth Philip Abraham Mulamoottil K Babu Joseph The Chief Editor

Ajit K Kembhavi airis4D

Geetha Paul Thelliyoor - 689544

Arun Kumar Aniyan India

Jorunal Publisher Details

Publisher : airis4D, Thelliyoor 689544, India

Website : www.airis4d.com

Email : nsp@airis4d.com

Phone : +919497552476

Editorial

by Fr Dr Abraham Mulamoottil

airis4D, Vol.2, No.3, 2024

www.airis4d.com

In this issue of airis4D, we joyfully document Pro-

fessor Ajit Kembhavi’s visit to the AIRIS Lab on Febru-

ary 21, 2024. Professor Kembhavi shared valuable in-

sights into the Pune Knowledge Cluster (PKC), show-

casing its innovative approach to tackling regional chal-

lenges through collaborative eﬀorts among academia,

R&D institutions, and industry. The engaging discus-

sion with Professor Kembhavi’s team has ignited our

enthusiasm, and we look forward to potential collabo-

rations as we aim to establish a transformative Knowl-

edge Cluster in our own context and region.

Additionally, we are pleased to introduce Dr. Bal-

amuralidhar Purushothaman, former Chief Scientist at

TCS Research Bangalore, who continues to contribute

as a research advisor at TCS. His notable interest in

Neuroarts and Neurophilosophy deserves special men-

tion in this edition.

The edition starts with Arun Aniyan’s article ”Build-

ing Applications with Large Language Models”. The

article discusses the recent developments in Large Lan-

guage Models (LLMs) and their applications in natu-

ral language processing (NLP). LLMs, such as Ope-

nAI’s ChatGPT and Google’s Gemini, are deep learn-

ing models with massive trainable parameters, enabling

them to solve general-purpose language generation and

understanding tasks. They are trained on large datasets

from various sources, allowing them to learn and un-

derstand tasks across diﬀerent domains. The article in-

troduces the Retrieval Augmented Generation (RAG)

methodology to address the issue of hallucination in

LLMs. Hallucination occurs when the model gener-

ates responses unrelated to the input data, and RAG

involves retrieval, augmentation, and generation steps

to reduce this eﬀect. Despite its eﬀectiveness, RAG is

not fail-safe, and other methods like memory addition

and prompt templating are employed to enhance LLM

performance.

In Blesson George’s article ”Building blocks of

Attention Network-Part 1”, the author explores fun-

damental concepts of attention networks, focusing on

sequence-to-sequence networks and encoder-decoder

architectures. The article emphasizes the role of re-

current neural networks (RNNs) and Long Short-Term

Memory (LSTM) networks in sequence-to-sequence

learning, illustrating their importance in tasks like lan-

guage translation. The encoder-decoder framework is

dissected, highlighting its role in transforming input

sequences into coherent output sequences. The article

concludes by emphasizing the signiﬁcance of attention

networks in contemporary computational models.

Linn Abraham’s article ”Introduction to Solar As-

tronomy” explores fundamental aspects of solar sci-

ence. The article covers topics such as sunspots, the so-

lar cycle, and the solar corona. It highlights the signiﬁ-

cance of the Sun as a laboratory for understanding stars

and processes like controlled fusion reactions. The

challenges of studying the Sun, including its magnetic

ﬁelds and complex atmosphere, are discussed. The au-

thor introduces features like active regions, ﬂares, and

coronal mass ejections, emphasizing the importance of

studying the Sun in ultraviolet wavelengths. The article

concludes by mentioning space-based solar observato-

ries like Aditya-L1, contributing to our understanding

of the Sun.

Sindhu G’s article ”Color - Magnitude Diagram,

Part-2” discusses the signiﬁcance of the color-magnitude

diagram (CMD) in astronomy, focusing on the princi-

ples of magnitudes and color. The Johnson system

with ﬁlters U, B, V, R, and I is explained, emphasizing

their role in studying diﬀerent parts of the electromag-

netic spectrum. The article explores the concept of

color in astronomy, detailing the spectral response of

ﬁlters and the understanding of color indices. It high-

lights the role of color in providing insights into stel-

lar properties such as temperature, composition, and

evolutionary stage. The factors inﬂuencing star color,

including temperature, composition, and evolutionary

stage, are discussed. The article also touches on ob-

servational methods and visual astronomy’s limitations

in perceiving color. The next part is teased, promis-

ing an exploration of Color-Magnitude diagrams in the

upcoming article.

Geetha Paul’s article ”Odonates: Sensitive Indi-

cators of Aquatic Ecosystem Health and Pollution” ex-

plores the signiﬁcance of dragonﬂies and damselﬂies

(Odonates) as indicators of pollution, habitat qual-

ity, and landscape disturbance. The article empha-

sizes their taxonomic description, ecological adaptabil-

ity, and responsiveness to physicochemical parameters,

such as pH and dissolved oxygen. Three species—Brachythemis

contaminata, Brachydiplax chalybea, and Rhodothemis

rufa—are detailed, including their appearances, habi-

tats, and behaviors. The article underscores the im-

portance of taxonomic classiﬁcation in understanding

biological diversity and its role in biodiversity man-

agement. Odonates’ value as bioindicators in assess-

ing freshwater ecosystem health and pollution is high-

lighted, setting the stage for a follow-up article on

clean-water indicator species of odonates.

In the article ”Introduction to Aging Clocks” by

Jinsu Ann Mathew, the focus is on DNA methylation as

a crucial aspect of the aging clock. Drawing an anal-

ogy to a musical score, the article explains that DNA

methylation functions like a dimmer switch, inﬂuenc-

ing the activation or deactivation of genes. It delves

into the process, involving DNA methyltransferases

(DNMTs) and CpG sites, where methyl groups are

added to cytosine bases. The article details the roles of

maintenance DNMTs (DNMT1) and de novo DNMTs

(DNMT3A and DNMT3B) and emphasizes the signif-

icance of CpG islands in this process. Additionally,

the article discusses the enzyme S-adenosyl methion-

ine (SAM) as a methyl group donor. The complexity of

DNA methylation, its impact on gene expression, and

its reversible nature are highlighted, concluding with

insights into its role in regulating biological processes

and contributing to genomic stability. The article ”AI

for Remote Sensing” by Balamuralidhar P. a Senior

Scientist from TCS, Bangalore, explores how artiﬁcial

intelligence (AI) is transforming space-based remote

sensing, enhancing Earth observation and sustainabil-

ity eﬀorts. It discusses the shift from large satellites to

smaller, more agile constellations with powerful sen-

sors. The applications of AI in remote sensing include

image classiﬁcation, land cover mapping, and disaster

response. The article addresses challenges like train-

ing complexity and emphasizes impactful applications,

such as wildﬁre detection and precision forestry. It con-

cludes by highlighting challenges and the increasing

importance of AI in handling large volumes of satellite

data and the deployment for satellite image analysis.

Overall, the article provides insights into the growing

role of AI in satellite technology.

iii

News in Brief

by News Desk

airis4D, Vol.2, No.3, 2024

www.airis4d.com

A photographic session at airis4D. Professor Ajit Kembhavi (IUCAA, Pune), Professor Sudhanshu Barway

(IISc, Bangalore), Dr Disha Sawant and Atharva Pathak from the Pune Knowledge Cluster visited airis4D

on February 21, 2024. Professor Dr Fr Alexander Abraham (formerly at Marthoma College, Thiruvalla),

Professor Dr Vinu Vikram (Central University, Kasaragod) with Dr Fr Abraham Mulamoottil and other

members of airis4D.

Contents

Editorial ii

News in Brief v

I Artiﬁcial Intelligence and Machine Learning 1

1 Building Applications with Large Language Models 2

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Large Language Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.3 Design methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Building Blocks of Attention Network-Part 1 6

2.1 Sequence to Sequence Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

II Astronomy and Astrophysics 9

1 Introduction to Solar Astronomy 10

1.1 The Exciting Science in the Sun . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.2 Challenges of Learning about the Sun . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.3 Shedding More Light on the Sun . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.4 Features and Events on the Sun . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.5 How Best to Study the Sun? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Color - Magnitude Diagram, Part-2 14

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2 Filters in Astronomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.3 Understanding Color in Astronomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

III Biosciences 18

1 Odonates: Sensitive Indicators of Aquatic Ecosystem Health and Pollution- Part 2 19

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.2 Species Description of Odonata Species Commonly Found in Polluted Waters . . . . . . . . . . 19

2 Introduction to Aging Clocks - Part 2 24

2.1 The Enzyme: DNA Methyltransferases (DNMTs) . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.2 The Target: CpG Sites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.3 The Chemical Tag: Adding the Methyl Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

CONTENTS

IV Remote Sensing 27

1 AI for Remote Sensing 28

vii

Part I

Artiﬁcial Intelligence and Machine Learning

Building Applications with Large Language

Models

by Arun Aniyan

airis4D, Vol.2, No.3, 2024

www.airis4d.com

1.1 Introduction

The recent developments in Large Language Mod-

els (LLM) are extremely exciting, oﬀering solutions to

solve niche tasks that were earlier tricky and time-

consuming. The general class of machine learning

(ML) solutions, is widely used to solve problems that

need automation and work at scale. They are ﬁne-

tuned for a speciﬁc task and have no access to other

data or data sources apart from what it is trained for.

Such models are generally lightweight and are easy to

deploy. Model size and performance are often high

priority during the design phase, which enables easy

scaling of the application that hosts it.

These ML models are mostly simple and are trained

for one task. For a slightly diﬀerent task, a diﬀerent

model with relevant data needs to be trained. The

model retraining frequency depends on the rate of data

drift of the system. This is largely small across all

bands of applications. The “world” that these models

know is limited in terms of knowledge.

Most of the areas where general machine learning

methods are used do not involve context, except those

required to process natural language and speech data.

Applications such as chatbots require context both in

the long term as well the short term. Many natural

language processing (NLP) problems that require con-

textual information are designed to access information

from knowledge graphs. A knowledge graph allows for

updating the contextual information but is generally an

expensive and time-consuming process. Traditional

ML methods used in NLP applications such as LSTMs

enable the building real-time contextual memory for a

limited length. But as the data ﬂow continues, the ac-

curacy and memory of such methods drop. LLMs oﬀer

solutions to resolve some issues mentioned above.

1.2 Large Language Models

LLMs just like the name mentions are deep learn-

ing models that are used to solve general-purpose lan-

guage generation and understanding issues. They are

large in the sense that they have massive numbers of

trainable parameters (in the order of billions) and hence

require huge amounts of data to train. Under the hood,

they are an encoder-decoder model mostly known as

the transformer model. They are used for text genera-

tion, which takes in a set of input tokens and generates

the next token or word. LLMs are a class of foundation

models that can solve multiple use cases when trained

on a massive dataset. They are generally trained on

text data that comes from various sources, which en-

compass diﬀerent domains. This enables the model

to learn and understand diﬀerent tasks that are spread

across domains. Such models can generate general

text in diﬀerent languages and even computer code.

OpenAI’s ChatGPT, Google’s Gemini, and Facebook’s

Llama are a few examples of LLMs. In addition to

general text generation, recent advances have made it

possible to generate and understand image data using

LLMs. They are known as multimodal LLMs. Ope-

nAI’s GPT-Vision is a well-known example. Multi-

1.3 Design methodology

modal models are powerful enough to understand an

image and produce insights about it. They are eﬃcient

enough to even perform basic computer vision tasks

when provided with the right input prompt. Some

models can even extract text from ﬁles with tables and

analyze them. Foundational model abilities along with

multimodal capacity make LLMs powerful enough to

build applications without special training. The multi-

domain “world” knowledge allows one to use the same

LLM model for a multitude of applications without the

need of additional training in numerous instances.

1.3 Design methodology

LLMs in most cases generate human-like responses

and times tend to generate garbage responses. This is

largely due to an eﬀect called “hallucination”. This hal-

lucination is mainly due to inconsistent training data,

overﬁtting, and model noises. Controlling hallucina-

tion is the ﬁrst and top priority when designing ap-

plications with LLMs. General ML models which do

not hallucinate are always consistent in terms of out-

put. They either generate correct or incorrect predic-

tions. Software applications require consistent results

in terms of format and data structure.

1.3.1 Retrieval Augmented Generation

Controlling hallucinations is the top priority and

there are diﬀerent methods to solve them which in-

volve ﬁne-tuning, data cleaning, and human feedback.

One form of hallucination eﬀect seen with LLMs is

the tendency to generate output that has no relation

to the input data. For example, when asked to gener-

ate a SQL query to fetch speciﬁc data points from a

table given its database schema, some models put in

columns and join statements from other tables that do

not exist. This form of hallucination happens when the

model has no clear context about the problem. LLMs

are mostly trained on a public set of data and may not

have information on closed-domain data such as com-

pany data. If the predictions generated from an LLM

can be compared with the internal data and corrected

based on the diﬀerences, this will enable to generation

Figure 1: Generic RAG architecture used with LLMs

for grounding and reducing hallucinations.

of more consistent and valid results. This is the design

principle of the method called Retrieval Augmented

Generation or RAG. It is a form of method applied to

reduce LLM hallucination with domain-speciﬁc inter-

nal data as shown in Figure 1.

With RAG, there are three major steps. (1) Re-

trieval, (2) Augmentation, and (3) Generation. Re-

trieval involves fetching user-speciﬁc information from

a knowledge base. For example, the model may be

trained with data dating back to the year 2022. To

answer a question based on an event that may have

happened in 2023, the model needs to retrieve the rel-

evant information from a speciﬁc recent source. The

augmentation steps will add or augment the user query

with the retrieved information. This happens before

the LLM takes in the query to generate the response,

essentially adding the retrieved knowledge to the query

the user posed.

The ﬁnal step of generation is the LLM gener-

ating a prediction that is more rich and contextual to

the user application. This mechanism controls LLM

hallucination to generate more factual and grounded

responses.

There is a standard architecture for designing RAG

using a vector store of data and query strategies. A doc-

ument embedding is generated and stored in a vector

database for the application to retrieve and add to the

model query.

Even though RAGs oﬀer a robust solution to re-

duce hallucination, they are not fail-safe at all times.

At times, they also generate empty results even when

provided with grounding data.

1.3 Design methodology

1.3.2 Memory

LLMs have a good amount of memory that is

larger than traditional models like LSTMS. At the same

time, LLMs are notorious for forgetting certain tasks

they just did. For example, the model will generate a

correct prediction for a speciﬁc query and if the same

query is repeated afterward, the LLM may generate a

garbage response saying it is unable to respond or does

not understand the query.

Adding memory to the chat prompt can help re-

solve this. This is not a direct feature provided by the

model, but an external application-level ability. Pack-

ages like LangChain are experimenting with memory

methodologies to embed into LLM applications. The

core principle of inducing memory revolves around two

aspects.

1. Storing a state

2. Querying a state

Storing a state involves saving a list of chat messages

that can be reused when prompting the model. Chat

messages with positive results are stored in a list and

later made available when similar prompts are input.

Querying a state is storing data structures and algo-

rithms on top of the chat.

1.3.3 API response structure

For general purpose applications, the results from

an LLM are passed onto a piece of code which requires

the data to be in a speciﬁc data structure which is

mostly JSON in nature. LLMs are good are generating

general text, but can also conform to speciﬁc structure.

”Prompting template” is a methodology used to restrict

the output of an LLM to a speciﬁc data structure. As

an example, if a code requires the model to generate an

output in the form of a python dictionary which has the

keys ‘exist’ and ‘value’ , the prompt template would

like the following.

Instructions:

Use only the provided schema to

generate the response.

{

exist: <True> or <False>,

value : 0-100

}

This form of template allows the model to only generate

a response within a limited data structure and is mostly

a key requirement of application code.

1.3.4 Retry Methods

Model hallucination comes into play, even meth-

ods like RAG, memory, and prompt templating are

employed. In such cases, a strategy employed to gener-

ate correct responses is retrying the prompt few times

until a correct response is received. One retry strat-

egy is to retry specifying that the model response is

incorrect and respond with a prompt mentioning the

output structure is wrong. Another strategy is to retry

the whole prompt again as a fresh call. This is often

done when the ﬁrst method fails.

1.3.5 Model Deployment

LLMs are very computationally expensive and re-

quire massive resources to train as well as deploy. At

times the deployment strategy fails to generate an API

response from an LLM due to various reasons such

as memory load, large token context and insuﬃcient

resources. LLMs are relatively new technology and

their deployment is done by large enterprises who have

massive compute capabilities. These are specialised

deployment methods which may not be available in

public domain. The diﬃculty in deploying LLMs at

scale is not just compute limitation, but the model ar-

chitecture itself. The model architecture is complex

and is not simple output like the logit activations of

conventional models. Multimodal models are even

more complex because they need to deal with not just

text input, but also image data. This heterogeneity

of model input and output makes the model export

and deployment more diﬃcult than usual. General de-

ployment software packages like Nvidia Triton oﬀers

decent capabilities to host a speciﬁc set of LLM mod-

els, but still limited in terms of model variety that it

1.4 Conclusion

can support. Llama.cpp and Ollama are another option

which supports large variety of models but has limited

scaling ability with load.

1.4 Conclusion

LLM application development is still at its in-

fancy, even though large number of developments have

been made in productising applications. Eﬀorts are

made on both model training to reduce hallucination

and applications side to provide more context as well

generate consistent output. The solution variety as well

power that LLMs oﬀer has massive potential to develop

AI assisted technology.

Reference

What are Large Language Models

Introduction to Large Language Models

What is retrieved augmented generation

Langchain Memory

Prompt Templating

Retry Parser

Triton LLM

Ollama Server

About the Author

Dr.Arun Aniyan is leading the R&D for Arti-

ﬁcial intelligence at DeepAlert Ltd,UK. He comes from

an academic background and has experience in design-

ing machine learning products for diﬀerent domains.

His major interest is knowledge representation and com-

puter vision.

Building Blocks of Attention Network-Part 1

by Blesson George

airis4D, Vol.2, No.3, 2024

www.airis4d.com

In this discourse, we delve into the fundamental

concepts underpinning the attention network, which

serves as the cornerstone of modern computational

models. Our exploration encompasses a thorough ex-

amination of pivotal constructs such as the sequence-

to-sequence network, the intricacies of encoder-decoder

architectures, and a discerning analysis of the con-

straints inherent in recurrent neural networks (RNNs).

Central to our discussion is the sequence-to-sequence

network, a seminal innovation that revolutionized the

ﬁeld of natural language processing and paved the way

for transformative advancements in various domains.

We dissect its inner workings, elucidating how it facil-

itates the seamless transformation of input sequences

into output sequences, thereby enabling a myriad of

applications ranging from machine translation to text

summarization.

2.1 Sequence to Sequence Networks

Sequence-to-sequence learning, often abbreviated

as Seq2Seq, represents a pivotal paradigm in the do-

main of machine learning, speciﬁcally tailored for tasks

involving the transformation of sequences from one do-

main to another. A quintessential example of Seq2Seq

learning is language translation, where the objective

is to convert a sequence of words from one language

(e.g., English) into another language (e.g., French).

At the heart of Seq2Seq models lies the utiliza-

tion of recurrent neural networks (RNNs) and their

variant, Long Short-Term Memory (LSTM) networks,

both adept at processing sequential data. RNNs and

LSTMs are uniquely tailored to handle input sequences

Figure 1: Figure showing a seq2seq model that

presents a two-LSTM network architecture tailored

for general sequence-to-sequence mappings, accom-

modating tasks like language translation where input

and output sequences can diﬀer in length. In this

architecture, the encoder LSTM sequentially processes

the input sequence token-by-token, generating a

ﬁxed-dimension context vector. Subsequently, the

decoder LSTM utilizes this context vector to produce

the output sequence token-by-token. Image Courtesy:

https://medium.com/one-minute-machine-learning/sequence-

to-sequence-learning-with-neural-networks-2014-one-minute-

summary-bce5e24c5e0c

of arbitrary lengths by maintaining a hidden state that

retains crucial information from preceding elements in

the sequence. This inherent capability renders them in-

dispensable for both the encoding and decoding phases

of Seq2Seq models.

In the realm of Seq2Seq learning, RNNs and

LSTMs ﬁnd application in two distinct sections: the

encoder and the decoder.

2.1.1 Recurrent Neurel Networks

A recurrent neural network (RNN) stands as a

cornerstone in the domain of deep learning, renowned

for its proﬁciency in processing and transforming se-

quential data inputs into corresponding sequential data

outputs. Sequential data, characterized by intercon-

nected components such as words, sentences, or time-

series data, embodies complex semantics and syntax

2.1 Sequence to Sequence Networks

rules, thereby demanding specialized processing mech-

anisms.

RNNs are made of neurons: data-processing nodes

that work together to perform complex tasks. The neu-

rons are organized as input, output, and hidden layers.

The input layer receives the information to process, and

the output layer provides the result. Data processing,

analysis, and prediction take place in the hidden layer.

RNNs work by passing the sequential data that they re-

ceive to the hidden layers one step at a time. However,

they also have a self-looping or recurrent workﬂow: the

hidden layer can remember and use previous inputs for

future predictions in a short-term memory component.

It uses the current input and the stored memory to pre-

dict the next sequence. Expanding upon the concept of

recurrent neural networks (RNNs), at every time step,

the network unfolds, essentially extending its architec-

ture to accommodate processing across multiple time

steps. This unfolding allows us to visualize the net-

work’s operations and outputs at each time step, akin

to a temporal sequence of computations. Remarkably,

this unfolded representation bears a striking resem-

blance to a feedforward neural network, where each

time step corresponds to a layer in the network. Within

this unfolded network, each rectangle denotes an oper-

ation occurring at a speciﬁc time step, illustrating the

sequential nature of the computations performed by

the RNN over the entire input sequence. This unfold-

ing mechanism enables the RNN to capture temporal

dependencies and process sequential data eﬀectively,

making it a powerful tool for tasks such as time-series

prediction, language modeling, and sequence genera-

tion.

The mathematics underlying recurrent neural net-

works (RNNs) involve the propagation of information

through time, where the network’s architecture allows

it to maintain memory of past inputs by incorporat-

ing feedback loops. To understand the mathematical

formulation of RNNs, let’s break it down into its key

components:

1. Input at Time Step t: Let x

(t)

represent the

input vector at time step t. This input could be a

word embedding in natural language processing

tasks or a feature vector in time-series analysis.

2. Hidden State at Time Step t: The hidden state

(t)

at time step t is computed based on the

current input x

(t)

and the previous hidden state

(t−1)

. It captures the network’s memory or

context at a given time step.

3. Mathematical Formulation: The hidden state

(t)

at time step t is calculated using the follow-

ing formula:

(t)

= f(W

(t)

+ W

(t−1)

+ b

)

where: - W

is the weight matrix connecting the

input x

(t)

to the hidden state h

(t)

. - W

is the

weight matrix connecting the previous hidden

state h

(t−1)

to the current hidden state h

(t)

. - b

is the bias vector. - f is the activation function,

often a non-linear function like the hyperbolic

tangent (tanh) or the rectiﬁed linear unit (ReLU).

4. Output at Time Step t: The output y

(t)

at time

step t is typically computed based on the current

hidden state h

(t)

using a separate output layer:

(t)

= g(W

(t)

+ b

)

where: - W

is the weight matrix connecting

the hidden state h

(t)

to the output y

(t)

. - b

is the

bias vector. - g is an activation function, often a

softmax function for classiﬁcation tasks or linear

activation for regression tasks.

5. Training: During training, the network learns

the parameters (weights and biases) W

, W

, b

, and b

by minimizing a loss function,

typically using backpropagation through time

(BPTT) or a variant of it.

By iteratively applying the above equations at each time

step, RNNs can process sequential data and capture

temporal dependencies, making them powerful tools

for tasks such as natural language processing, time-

series analysis, and sequence prediction. However,

traditional RNNs suﬀer from the vanishing gradient

problem, which limits their ability to capture long-term

dependencies. This limitation led to the development

of more sophisticated architectures such as Long Short-

Term Memory (LSTM) networks and Gated Recurrent

Units (GRUs), which address this issue and are widely

used in practice.

2.2 Conclusion

2.1.2 Encoder-Decoder Network

The seq2seq model consists of 3 parts: encoder,

intermediate (encoder) vector and decoder.

In the encoder phase of sequence-to-sequence mod-

els, a stack of recurrent units, typically LSTM or GRU

cells for improved performance, processes the input

sequence element by element. Each recurrent unit re-

ceives a single element of the input sequence, aggre-

gates information speciﬁc to that element, and passes

it forward. By iterating through the entire input se-

quence, the encoder gathers contextual information

from preceding elements, gradually updating its in-

ternal states to form a comprehensive representation of

the input. This representation, often termed the context

vector or latent representation, captures the essence of

the input sequence and serves as the foundation for

subsequent phases in the sequence-to-sequence model,

such as decoding and generating the output sequence.

The Encoder Vector, also known as the context

vector, serves as the culmination of the encoder’s pro-

cessing of the input sequence. Computed based on

the ﬁnal hidden state of the encoder, it encapsulates the

collective information gleaned from all input elements.

This vector plays a pivotal role in facilitating accu-

rate predictions by the decoder. It serves as the initial

hidden state of the decoder, providing crucial contex-

tual information to guide the generation of the output

sequence. By condensing the entirety of the input se-

quence into a ﬁxed-dimensional representation, the En-

coder Vector enables the decoder to eﬀectively under-

stand and interpret the input, thus enhancing the overall

performance of the sequence-to-sequence model.

In sequence-to-sequence models, the decoder com-

ponent, composed of recurrent units, generates the out-

put sequence by iteratively predicting elements at each

time step. Each decoder unit accepts the previous hid-

den state and produces both an output and its own hid-

den state, a process continuing until the entire sequence

is generated. For tasks like question-answering, where

the output comprises words forming the answer, each

word is represented as an output y

. The decoder’s

hidden states are computed based on speciﬁc formulas,

typically involving nonlinear activations applied to lin-

ear transformations of input and previous hidden states.

Ultimately, the decoder leverages contextual informa-

tion from the encoder’s ﬁnal hidden state to produce

coherent and contextually relevant outputs, such as an-

swers to questions.

2.2 Conclusion

In this discourse, we’ve navigated through the

foundational elements of attention networks, funda-

mental to contemporary computational models. Our

journey encompassed a thorough exploration of piv-

otal concepts like sequence-to-sequence networks and

encoder-decoder architectures, alongside an insightful

analysis of the challenges posed by recurrent neural net-

works (RNNs). With a focus on sequence-to-sequence

learning, we uncovered its pivotal role in transforma-

tive applications, particularly in language translation

tasks. We’ve unraveled the inner workings of RNNs

and LSTMs, illuminating their indispensable functions

in processing sequential data. Moreover, we’ve dis-

sected the encoder-decoder framework, elucidating its

role in facilitating the seamless transformation of input

sequences into output sequences

References

1. Understanding Encoder-Decoder Sequence to Se-

quence Model

2. Introduction to Sequence to Sequence Models

3. Sequence to Sequence Learning with Neural Net-

works

About the Author

Dr. Blesson George presently serves as an As-

sistant Professor of Physics at CMS College Kottayam,

Kerala. His research pursuits encompass the develop-

ment of machine learning algorithms, along with the uti-

lization of machine learning techniques across diverse

domains.

Part II

Astronomy and Astrophysics

Introduction to Solar Astronomy

by Linn Abraham

airis4D, Vol.2, No.3, 2024

www.airis4d.com

1.1 The Exciting Science in the Sun

The Sun is the star closest to us and hence is

the best laboratory to expand our understanding about

stars and stellar processes in general. It is what drives

our ambitious goal of unlocking the ultimate source of

power, a.k.a controlled fusion reaction. And which is

capable of providing clean energy for the entire known

future of humanity. It can help us learn about neutri-

nos and several more fascinating things. It is also the

heavenly body that has the greatest inﬂuence on Earth.

This impact is seen manifested in many forms. From

the harmless aurorae to geomagnetic storms that are

capable of deorbiting satellites and temporary power

grid failures. However there are a lot of questions that

remains answered about the Sun as well. What heats

up the corona to such high temperatures? What trig-

gers ﬂaring activities in the Sun? and so on. India

now has its own space based solar observatory with the

Aditya-L1 satellite. Several of the instruments that are

present onboard open a never before seen window on

the Sun.

1.1.1 Sunspots

The surface of the Sun reveals that it is not a per-

fect orb as expected of a heavenly body acoording to a

divine picture of creation. This fact was discovered by

Galileo amongst others as soon as he trained the lat-

est invention of the day, the telescopes, towards these

heavenly bodies (see, Figure 1). Note that these obser-

vations are not made directly but by projecting it onto a

screen and making drawings of what was screen. With

the limited knowledge about the Sun they had at the

time, many considered these spots to be other planets

or as slag of the burning Sun or even opaque smoke

clouds. It was soon discovered that there is a cyclic pat-

tern to the appearance and disappearance of Sunspots.

An observation that has ﬁnally lead to the coinage “so-

lar cycle” denoting the periodic magnetic activity in

the Sun. The Sun being crucial to life on Earth, this

cycle can also be seen as one of the fundamental cycles

of nature. Interested readers may explore the popular

science book “Nature’s Third Cycle” by Dr. Arnab Rai

Choudhury.

Figure 1: Original drawing of sunspots made by

Galileo.

1.1.2 Corona

The solar corona was photographed for the ﬁrst

time during a solar eclipse. It is the halo of light with

1.2 Challenges of Learning about the Sun

long strands and intricate pattern seen during a total

solar eclipse of the Sun when the moon blocks the light

coming from the Sun’s disk (see Figure 2). The corona

expands into the interplanetary space and beyond in the

form of the solar wind. You may be familiar with the

grand display of colours that happens in the skies above

the northern and southern poles known as Aurorae.

This happens as a result of the interaction of the solar

wind with the Earth’s own magnetic ﬁeld.

Figure 2: Corona as observed during a total solar

eclipse.

1.2 Challenges of Learning about the

Sun

1.2.1 Interior of the Sun vs. the Solar

Atmosphere

It can be seen from theoretical calculations that

the matter in the interior of the Sun should exist in

the plasma state of matter. However by calculating

the mean free path inside the Sun and the sizes of

the constituent particles it can be seen that the plasma

in the Sun behaves like a perfect gas. Considering the

densities and temperatures inside the Sun it can be seen

that most of the Sun’s interior is opaque. This means

direct probes of the Sun’s interior is very limited.

The part of the Sun from which photons can di-

rectly escape into space is called the atmosphere of

the Sun. It consists of the Photosphere, Chromosphere

and Corona. The part of the Sun that is directly vis-

ible to us is called the Photosphere. It is interesting

to note that the high energy gamma rays generated by

nuclear fusion in the Sun’s core takes 1,70,000 years

to reach the photosphere. Where the collisions brings

the wavelength into the visible regime.

1.3 Shedding More Light on the Sun

We obviously know about the Sun by observations

in the visible spectrum. Or in other words observations

made in the visible region of the spectrum are looking

into the photosphere of the Sun. Being the blackbody

that the Sun is, we also expect it to emit in other wave-

lengths.

1.3.1 How Magnetic Fields Complicates

Observations

The presence of high magnetic ﬁelds on the Sun’s

surface were ﬁrst observed by spectroscopic analysis.

Speciﬁcally the Zeeman splitting of spectral lines al-

lows us to gauge the presence and strength of magnetic

ﬁeld through which light has passed before reaching

the detector. Using the Zeeman eﬀect it was also found

that Sunspots are regions of higher magnetic ﬁeld than

surrounding regions. The presence of magnetic ﬁeld

is unavoidable once we understand that plasma which

the Sun is made up of is essentially a soup of charged

ions and electrons. And we know that moving charges

always create magnetic ﬁeld. However since plasma is

also neutral if you consider both charges together, it is

not that simple to explain. Our best understanding of

the mechanism behind the magnetic ﬁeld generation is

called the Dynamo model.

The presence of magnetic ﬁelds means that the

Sun’s emission can no longer be understood just as a

black body emission. One observation that shows how

magentic ﬁelds can complicate things is the nature of

temperature distribution in the Solar atmosphere. The

temperature in the solar atmosphere instead of decreas-

ing with height, stop decreasing when it reaches the

region known as the transition regions and then rapdily

1.4 Features and Events on the Sun

shoots up. Reaching 1 million in the corona. This is

called the coronal heating problem. The presence of

such high temperature leads to the presence of several

emission lines in the UV and X-ray regions.

1.4 Features and Events on the Sun

1.4.1 Active Regions

The other way in which magnetic ﬁeld compli-

cates things is by producing active regions that are

probable to produce prominences, ﬂares, CMEs etc.

An active region by deﬁnition is a region of complex

magnetic ﬁeld and is produced as a result of the twisting

of magnetic ﬁeld due to the solar diﬀerential rotation

as well the convection currents and other dynamics in

the Sun’s interior. When an active region matures we

ﬁnd within it darker regions of more intense magnetic

ﬁeld that are the Sunspots. Near to these active regions

one can often ﬁnd cool dark ribbons called ﬁlaments

(when observed on disc) and prominences (when ob-

served against the limb). Such prominence are prone

to eruption at some point in their lifetime resulting at

times in a CME or sometimes in a solar ﬂare. CMEs

involve mass being ejected into the space surround-

ing the Sun whereas ﬂares are intense violet release of

energy as radiation.

1.5 How Best to Study the Sun?

There are several methods to study the Sun de-

pending on your speciﬁc interest. For example if you

are interested in learning about ﬂares, there are some

ﬂares that are observable in the visible region of the

EM spectrum, that are called white light ﬂares. There

are ground and space based instruments that may be

able to detect the X-ray particles emitted as part of

these ﬂares. Usually the detections are done separately

in the soft and hard X-ray channels. Such instruments

are often used to detect whether a ﬂare has happened or

not and since they mostly integrate the x-ray emission

over the whole disc of the Sun, the location informa-

tion would be lost for where the ﬂare has occurred. We

mentioned the possibility of using the Zeeman eﬀect to

Figure 3: Post-eruptive loops in the wake of a solar

ﬂare, image taken by the TRACE satellite (photo by

NASA)

quantify the magnetic ﬁeld in the Sun. In fact it is pos-

sible to build a map of the Sun using this magnetic ﬁeld

measurement. Such a map in which the areas of strong

positive magnetic ﬁeld are colored white, the strong

negative magnetic ﬁeld regions are colored black and

grey regions show areas with an absence of magnetic

ﬁelds. Coronographs are used to continously observe

the Corona by artiﬁcially eclisping the Sun with the

help of an occulting disc. The light obtained using this

can be further subjected to spectroscopic analysis.

1.5.1 Imaging in the Ultraviolet

We had already mentioned that at diﬀerent heights

in the solar atmosphere you have diﬀerent tempera-

tures. By imaging the Sun in narrow passbands cen-

tered over wavelengths in the Ultraviolet region, we are

eﬀectively looking into the Solar atmosphere at diﬀer-

ent heights. This is because the temperatures present

at the diﬀerent layers leads to speciﬁc ionizations that

emit in speciﬁc wavelengths in the Ultraviolet region.

Spectroscopy is often time consuming and leads to low

cadence data. However we can obtain a very high

cadence when using imaging and this is specially rel-

evant for changes in the solar surface that happen at

very small time scales. Ultraviolet being abosrbed to

REFERENCES

Figure 4: NISP/SOLIS magnetogram from 18 June

2014 during solar maximum.

a high extent by the Earth’s atmosphere we need space

based telescope for such observations. Existing space

based solar observatories like the Solar Dynamic Ob-

servatory (SDO) have given us valuable data over the

year and is still observing the Sun. A we have al-

ready seen, India now has it’s own space based solar

observatory called Aditya-L1 which carries the Solar

Ultraviolet Imaging Telescope (SUIT) developed by

the Inter University Center for Astronomy and Astro-

physics (IUCAA) along with many other institutions

across India. In comparison to the Atmospheric Imag-

ing Assembly (AIA) telescope onboard SDO which im-

ages the Sun primarily in 7 Extreme Ultraviolet (EUV)

channels, SUIT observes the Sun in 8 Near Ultravio-

let (NUV) channels. Which means new and exciting

science awaits those who are interested to ﬁnd it out.

References

[1] Arnab Rai Choudhuri. Nature’s Third Cycle: A

Story of Sunspots. Oxford University Press, Oxford

; New York, 2015. ISBN 978-0-19-967475-6.

[2] Frank H. Shu. The Physical Universe: An In-

troduction to Astronomy. A Series of Books in

Astronomy. Univ. Sience Books, Sausalito, Calif,

9. print edition, 1982. ISBN 978-0-935702-05-7.

Figure 5: The SDO/AIA images observed in multi-

wavelength channels targeting the AR 12804 observed

on 2021 February 25 at ﬂare peak time (14:16 UT).

[3] Eric Priest. Magnetohydrodynamics of The Sun.

About the Author

Linn Abraham is a researcher in Physics,

specializing in A.I. applications to astronomy. He is

currently involved in the development of CNN based

Computer Vision tools for prediction of solar ﬂares

from images of the Sun, morphological classiﬁca-

tions of galaxies from optical images surveys and ra-

dio galaxy source extraction from radio observations.

Color - Magnitude Diagram, Part-2

by Sindhu G

airis4D, Vol.2, No.3, 2024

www.airis4d.com

2.1 Introduction

The color-magnitude diagram (CMD) functions

as a powerful tool in astronomy, enabling the examina-

tion and comprehension of the characteristics of stel-

lar populations in distinct celestial regions. Before

embarking on an exploration of color-magnitude dia-

grams (CMDs), it is essential to familiarize ourselves

with the principles of magnitudes and color in the ﬁeld

of astronomy. In the previous article, we explored the

concept of magnitudes. Now, we will direct our focus

toward exploring the concept of color in astronomy.

2.2 Filters in Astronomy

Filters play a crucial role in astronomical obser-

vations, helping astronomers study celestial objects in

diﬀerent parts of the electromagnetic spectrum. Each

ﬁlter is speciﬁcally crafted to admit light within a par-

ticular wavelength range. The functionality of these

ﬁlters involves obstructing light across all wavelengths

except those centered around their designated wave-

length. The sensitivity of each ﬁlter decreases gradu-

ally towards both shorter and longer wavelengths. Fil-

ters utilized in ultraviolet, visible, and infrared obser-

vations are typically crafted from colored glass, dyed

plastic, gelatin, or similar materials. These ﬁlters se-

lectively permit a narrow waveband, usually around

100 nm in width, of radiation to pass through. Al-

ternatively, certain optical ﬁlters leverage interference

to create highly specialized narrow band ﬁlters, with

the waveband sometimes being only a few nanometers

Figure 1: Normalised intensity plot showing the spec-

tral response of the ﬁve ﬁlters in the Johnson-Cousins

system modiﬁed by Bessell. Credit: M. Bessell

wide.

The Johnson system is a widely adopted standard

employing ﬁve ﬁlters: U, B, V, R, and I. These ﬁlters

exhibit peak responses in the ultraviolet, blue, yellow-

green, red, and near-infrared segments of the spectrum,

respectively. The speciﬁc combinations of glass ﬁlters

and photo multiplier tubes were outlined to deﬁne these

ﬁlters. M. S. Bessell detailed a set of ﬁlter transmis-

sions for a detector with a ﬂat response, establishing a

basis for calculating color indices. Precision in ﬁlter

selection is vital, with B-V, for objects with mid-range

temperatures, U-V for hotter objects, and R-I for cooler

ones. The spectral response of the Johnson ﬁlter series

is illustrated in the Figure: 1 , depicting a modiﬁed

version developed by M. Bessel.

The Figure: 2 provides the peak wavelengths for

the ﬁlters in the standard Johnson UBVRI system.

2.3 Understanding Color in Astronomy

Figure 2: Peak wavelengths for the ﬁlters in the stan-

dard Johnson UBVRI system. Credit: ATNF

Figure 3: Comparison of a cool carbon star, R CMi

through blue and red ﬁlters. Credit: Adapted from

Digital Sky Survey data obtained from SkyView

The Sloan Digital Sky Survey utilizes ﬁve ﬁlters:

Ultraviolet (u), Green (g), Red (r), Near Infrared (i),

and Infrared (z), with corresponding wavelengths ex-

pressed in Angstroms as 3543, 4770, 6231, 7625, and

9134, respectively.

In Figure: 3, a comparative analysis of stellar im-

ages through various ﬁlters is presented. The central

star in each image is R Canis Minor (R CMi), identiﬁed

as a highly red carbon star with a C7 spectral class. The

observation reveals a noticeable diﬀerence in bright-

ness, particularly with the star appearing considerably

brighter through the red ﬁlter than the blue one. The

spikes surrounding it are not genuine components of

the star; rather, they are diﬀraction spikes resulting

from the bright light of point source stars diﬀracting

on the telescope optics. Similarly, the halo around R

CMi in the red frame stems from photochemical pro-

cesses occurring when the original photographic plate

is exposed to intense point sources. Both halos and

diﬀraction spikes serve as examples of artifacts.

Within the highlighted ﬁelds, two additional stars

are emphasized. The visibility of the hot star in the

blue plate contrasts with its fainter appearance in the

red. Conversely, the cool star stands out more promi-

nently in the red plate compared to the blue.

The set of passbands or ﬁlters is called a photo-

metric system. The use of ﬁlters extends to various

wavebands in the electromagnetic spectrum. Radio

telescopes employ electrical ﬁlters, where signal atten-

uation varies with frequency. Similar to optical ﬁlters,

these radio ﬁlters can be optimized to either permit a

broad or limited range of frequencies.

2.3 Understanding Color in

Astronomy

The role of color in astronomy is of utmost im-

portance, providing crucial insights into the charac-

teristics of celestial entities, especially stars. Color,

in this context, goes beyond visual appeal and acts as

a fundamental indicator of various properties such as

temperature, composition, and evolutionary stage. In

the realm of astronomy, color is deﬁned as the diﬀer-

ence in magnitude between two ﬁlters. Magnitude is a

measure of brightness, with higher numbers indicating

fainter objects. The color index, a numerical expres-

sion, determines the color of an object and is closely

related to its temperature. For example, a hot, young

star emits a larger portion of its light in the blue end of

the spectrum, giving it a lower color index. Conversely,

a cooler, older star emits more red light, resulting in a

higher color index.

In the context of SDSS data, astronomers examine

the ﬁve magnitudes of a star and calculate the diﬀerence

between any two. While the g-r value serves as one

method to ascertain color, astronomers possess addi-

tional choices due to the implementation of ﬁve ﬁlters

in the SDSS. Alternative color measurement options

2.3 Understanding Color in Astronomy

include u-r, r-i, or i-z. Therefore, when astronomers

discuss a star’s ”color,” they speciﬁcally refer to these

magnitude diﬀerence measurements, such as g-r, u-r,

i-z, and so forth. Inquiring about the color of a star

wouldn’t prompt an answer like ”red” or ”white” from

an astronomer; instead, they would provide informa-

tion such as ”this star exhibits a g-r color of 1.3.”

2.3.1 Factors Inﬂuencing Star Color

Composition

The constituents of stars emit distinct wavelengths

of electromagnetic radiation when heated, contributing

to the overall observed color.

Temperature

The emitted color of a star is strongly inﬂuenced

by its temperature. Hotter stars radiate bluer light,

while cooler stars emit light with a redder hue. Changes

in a star’s temperature correspondingly alter its emitted

light spectrum.

Evolutionary Stage

Stars undergo evolutionary changes, impacting

their sizes, temperatures, and colors. For example, as

a star progresses through stages such as the Red Giant

Phase, characterized by hydrogen depletion, it under-

goes expansion and exhibits a shift towards a deep red

color.

2.3.2 Observing Color in Astronomy

Spectroanalysis

Spectrometers are employed by scientists to an-

alyze the wavelengths emitted by stars, aiding in the

determination of their composition and other essential

properties.

Modern Classiﬁcation

Stars are categorized based on critical attributes

such as spectral class (color), temperature, size, and

brightness. The Morgan–Keenan system organizes

Figure 4: Color indices of Solar System bodies.

Credit: Wikipedia

stars from the hottest (O) to the coolest (M). The long-

standing tradition of the Johnson-Morgan system es-

tablishes that designations like UBVRIJK are generally

interpreted in the context of the Vega system, unless

speciﬁcally mentioned otherwise. The V band serves

as a commonly used reference, making colors such as

B-V, V-I, and V-K prevalent. The conventional UBV

system, rely on color indices, which are the recorded

distinctions in three or more color magnitudes. These

numerical values are assigned designations such as U-

B or B-V representing the colors transmitted through

two standard ﬁlters (e.g., Ultraviolet, Blue, and Vi-

sual). Objects with cooler temperatures, emitting min-

imal light below 8000

A, are frequently described by

terms like I-K, J-K, H-K, and so forth. The Sun, with

its whitish appearance, exhibits a B-V index of 0.656

± 0.005. Conventionally, Vega serves as the zero point

for the color index.

2.3.3 Visual Astronomy and Color Perception

Visual Limitations

Visual astronomy predominantly relies on black

and white observations due to the human eye’s limita-

tions in discerning color in low-light conditions.

Photographic Techniques

2.3 Understanding Color in Astronomy

Colorful images of celestial objects are often gen-

erated by combining multiple black and white expo-

sures captured through various colored ﬁlters.

In the forthcoming article, ’Color-Magnitude Di-

agram, Part-3’, we will delve into the exploration of

Color-Magnitude diagrams.

References:

The Colour of Stars

SDSS Filters

UBVRI passbands.

The Deﬁnition of Color in Astronomy

Cluster Colour-Magnitude Diagrams

Why Are Stars Diﬀerent Colors?

Visual Astronomy and Seeing Colours

MAGNITUDE AND COLOR SYSTEMS

Astronomical Toolkit

The SIMBAD astronomical database

Color index

About the Author

Sindhu G is a research scholar in Physics

doing research in Astronomy & Astrophysics. Her re-

search mainly focuses on classiﬁcation of variable stars

using diﬀerent machine learning algorithms. She is also

doing the period prediction of diﬀerent types of variable

stars, especially eclipsing binaries and on the study of

optical counterparts of X-ray binaries.

Part III

Biosciences

Odonates: Sensitive Indicators of Aquatic

Ecosystem Health and Pollution- Part 2

by Geetha Paul

airis4D, Vol.2, No.3, 2024

www.airis4d.com

1.1 Introduction

Odonates, including dragonﬂies and damselﬂies,

are valuable indicators of pollution, habitat quality,

and landscape disturbance. Their well-resolved taxon-

omy, conspicuous diurnal adults, contact with aquatic

and terrestrial environments, and broad environmen-

tal range make them suitable for monitoring ecosys-

tem health. As they inhabit both habitats, odonates

provide information about the condition of both wa-

ter bodies and adjacent landscapes. Odonates exhibit

speciﬁc responses to various physicochemical parame-

ters, such as pH, electrical conductivity, and dissolved

oxygen, inﬂuencing their distribution, abundance, and

diversity. Studies have shown that conductivity in-

creases correlate with higher pollution levels, while

variations in species composition reﬂect environmental

disturbances. Researches indicate that certain odonate

species are sensitive to pollution, providing insights

into the overall health of freshwater ecosystems. Tax-

onomy is the science of naming, describing, and clas-

sifying organisms, and this process involves writing

descriptions that distinguish them from others and as-

signing Latin names for universal identiﬁcation. Tax-

onomy is crucial for understanding biological diver-

sity and forms the basis for biodiversity management.

In this article, the taxonomic description of odonates,

which are bioindicators of aquatic health and pollution,

is described.

1.2 Species Description of Odonata

Species Commonly Found in

Polluted Waters

1.2.1 Brachythemis contaminata: (Fabricius,

1793)

(a) Brachythemis contaminata (b) Brachythemis contaminata

Figure 1: Brachythemis contaminata male(a) and fe-

male(b)

(a) (b)

Figure 2: A polluted wetland site of Kuttanad, with

Brachythemis contaminata (male), Kerala, India.

Taxonomic Classiﬁcation:

Kingdom: Animalia (animals), Phylum: Arthro-

poda (arthropods), Class: Insecta (insects), Order:

Odonata (dragonﬂies and damselﬂies), Family: Li-

bellulidae (true dragonﬂies), Genus: Brachythemis,

1.2 Species Description of Odonata Species Commonly Found in Polluted Waters

Species: Brachythemis contaminata.

The male Brachythemis contaminata presents a

visually striking appearance with orange coloration

on its wings. The hindwings of males measure be-

tween 21 to 23 mm in length, while the total body

length ranges from 29 to 31 mm. Their wings are

tinted with deep amber, except at the tips, and fea-

ture prominent orange veins and pterostigma. As they

age, the thorax and abdomen of males transition from

light brown to a richer, deeper orange hue. The fe-

males are very similar to the males. They are golden

brown with transparent wings and yellowish-orange

wing spots (pterostigma); the thorax and abdomen are

yellow/brown. The females become darker brown as

they age. The immature male is like the female. Both

sexes have brownish-yellow eyes capped with a darker

shade. Their small size and preference for ﬂitting low

over the water amongst vegetation can sometimes make

them challenging to observe. It breeds in weedy ponds,

lakes, and slowly moving streams, especially in slug-

gish waters. It is widespread along sewage canals,

tanks, ponds and ditches. It tolerates polluted water

well and aggregates around favourable spots where it is

active all day until after sunset. Brachythemis contam-

inata is a widespread species. IUCN Red List accessed

as Least Concern.

1.2.2 Brachydiplax chalybea : (Brauer, 1868)

(a) Brachydiplax chalybea (b) Brachydiplax chalybea

Figure 3: Brachydiplax chalybea male(a) and fe-

male(b)

Taxonomic Classiﬁcation:

Kingdom: Animalia, Phylum: Arthropoda, Class:

Insecta, Order: Odonata, Suborder: Anisoptera (true

dragonﬂies), Family: Libellulidae (skimmer dragon-

ﬂies), Genus: Brachydiplax, Species: Brachydiplax

chalybea.

Brachydiplax chalybea, commonly known as the

Oriental Blue Dasher, is a species of dragonﬂy belong-

ing to the family Libellulidae. Male adults measure

between 33 and 35 millimetres, with a hindwing span-

ning 24 to 27 millimetres. Males are powder blue

with light brown sides and a dark tip to the abdomen.

Wings are hyaline, with a tinted burnt-brown base, fad-

ing to amber. Females are brownish yellow with darker

markings along the dorsal abdomen, lacking the yel-

low tinge in males. The characteristic colour of the

thorax and bases of wings help distinguish this species

from other members of the same genus. This species

is commonly found in diverse wet habitats, including

freshwater ponds, marshes, and even slightly brack-

ish water. They are exceptionally adaptable and thrive

in disturbed (polluted) environments. Reproduction:

Males display territorial behaviour, and females de-

posit egg masses onto nearby vegetation.

(a) Mating (b) Egg Laying

Figure 4: Shows the mating and Egg Laying of fe-

male Brachydiplax chalybea. Image courtsey: https:

//en.wikipedia.org/wiki/Brachydiplax chalybea

Brachydiplax chalybea is a widespread species.

IUCN Red List accessed as Least Concern.

1.2.3 Rhodothemis rufa : (Rambur, 1842)

(a) Rhodothemis rufa (b) Rhodothemis rufa

Figure 5: Rhodothemis rufa male(a) and female(b)

Taxonomic Classiﬁcation:

1.2 Species Description of Odonata Species Commonly Found in Polluted Waters

Kingdom: Animalia, Phylum: Arthropoda, Class:

Insecta, Order: Odonata, Suborder: Anisoptera (true

dragonﬂies), Family: Libellulidae (skimmer dragon-

ﬂies), Genus: Rhodothemis, Species: Rhodothemis

rufa.

The Rhodothemis rufa, also known as the red-

bodied marsh hawk, is a medium-sized dragonﬂy with

vibrant colours, but the appearance diﬀers between

males and females. Males boast a striking, almost

scarlet red colour that covers their thorax, abdomen,

and even the base of their wings. Their eyes combine

reddish brown on top and bright red below. Young

males, however, share a ﬂeeting resemblance to the fe-

males, with a mid-dorsal yellow stripe running along

the thorax and abdomen.

Females, on the other hand, are pretty distinct.

They lack the vibrant red of the males, opting instead

for a rusty brown body. However, they retain the yellow

stripe from their younger stages, extending prominently

from the head down to the fourth segment of the ab-

domen. Their wings are similar to the males but with

less extensive markings at the base.

Both male and female Rhodothemis rufa share a

preference for freshwater habitats, and they can survive

in polluted waters. They are commonly found near

open ponds, weedy marshes, and calm lakes. These

areas provide them with the perfect environment for

hunting small ﬂying insects, their primary food source.

Their IUCN Red List accessed as Least Concern.

1.2.4 Zyxomma petiolatum : (Rambur, 1842)

(a) Zyxomma petiolatum (b) Zyxomma petiolatum

Figure 6: Zyxomma petiolatum male(a) and female(b)

Taxonomic Classiﬁcation:

Kingdom: Animalia, Phylum: Arthropoda, Class:

Insecta, Order: Odonata, Suborder: Anisoptera (true

dragonﬂies), Family: Libellulidae (true skimmers),

Genus: Zyxomma Species: Zyxomma petiolatum.

The Zyxomma petiolatum, also known as the dusk

ﬂier, is a medium-sized dragonﬂy with a distinctive

appearance. Both males and females share a chocolate-

brown body and striking emerald-green eyes. Their

wings are generally brown, with a blackish spot near

the tip.

While both sexes are similar in appearance, there

are subtle diﬀerences. Males tend to have a slightly

thinner abdomen compared to females.

Additionally, the males’ eyes are brighter, more

vibrant emerald green, while the females’ eyes appear

duller.

These dragonﬂies primarily inhabit areas near

freshwater sources less polluted with organic content

and leaf litter than others in polluted waters. They

can be found in small pools, ponds, swamps, and even

slow-ﬂowing rivers. Interestingly, they are most active

during dusk and dawn, making them crepuscular crea-

tures. During the day, they are much harder to spot as

they roost amongst dense vegetation near their habitat.

When active, Zyxomma petiolatum is known for

its extremely rapid and erratic ﬂight patterns. They

often ﬂy low over water bodies, preying on small in-

sects like midges and mosquitoes. They are also at-

tracted to lights, especially during the summer, making

them occasional visitors to porches and verandas in the

evenings.

1.2.5 Ceriagrion cerinorubellum : (Brauer,

1865)

(a) Ceriagrion cerinorubellum (b) Ceriagrion cerinorubellum

Figure 7: Ceriagrion cerinorubellum male(a) and

female(b)

Taxonomic Classiﬁcation:

Kingdom: Animalia, Phylum: Arthropoda,

Class: Insecta, Order: Odonata, Suborder: Zy-

1.2 Species Description of Odonata Species Commonly Found in Polluted Waters

goptera, Family: Coenagrionidae, Genus: Ceriagrion,

Species: Ceriagrion cerinorubellum.

The Ceriagrion cerinorubellum, also known as

the orange-tailed damselﬂy, is a vibrant damselﬂy with

distinct features between males and females. Males

boast a captivating appearance with a pale bluish-

green head and thorax, contrasting beautifully with

their bright orange base and tail tip on the abdomen.

The remaining abdominal segments showcase a unique

combination of black and pale blue. Females, though

similar, are generally less eye-catching with duller colours.

Their bodies are predominantly greenish with a slightly

yellowish tinge, and their abdomens are typically golden

yellow to brown.

This captivating damselﬂy prefers freshwater habi-

tats like weeded ponds, marshes, and other polluted

stagnant water bodies. Here, they thrive amongst the

aquatic vegetation, readily available for hunting and

breeding. Despite being primarily active during the

day, their small size and subtle colours can make them

easily overlooked. So, keep a keen eye out for these

colourful creatures ﬂitting amidst the reeds and grasses

if you ever ﬁnd yourself exploring their preferred wet-

land habitats.

1.2.6 Pseudagrion microcephalum :

(Rambur, 1842)

(a) (b)

Figure 8: Pseudagrion microcephalum male and

Pseudagrion microcephalum female Image courtesy:

https://indiabiodiversity.org/species/show/228180

Taxonomic Classiﬁcation:

Kingdom: Animalia, Phylum: Arthropoda, Class:

Insecta, Order: Odonata, Suborder: Zygoptera (dam-

selﬂies), Family: Coenagrionidae (narrow-winged dam-

selﬂies), Genus: Pseudagrion, Species: Pseudagrion

microcephalum.

Pseudagrion microcephalum, commonly known

as the blue riverdamsel, is a species of damselﬂy with

distinct diﬀerences between the sexes. The males are

visually striking with their sky-blue abdomens tipped

by vibrant black markings. Their heads and eyes are

also blue but with a slightly darker shade.

Conversely, females are less colourful, featuring

orange-tinted heads and thoraxes with a single, thin

black line running down the back. Their abdomens are

yellowish-green with black detailing. Pseudagrion mi-

crocephalum primarily inhabits areas with still or slow-

moving and polluted water sources, such as ponds,

canals, streams, and rivers. They are frequently found

resting on vegetation near the water’s edge, making

them a common sight on the leaf edges of their suit-

able habitats. Their IUCN Red List accessed as Least

Concern.

The above described are the common odonates

seen in polluted waters.

In conclusion, odonates serve as valuable bioindi-

cators for assessing the health of aquatic ecosystems

due to their sensitivity to environmental changes and

well-established taxonomy. Their diverse ecological

requirements and distinct responses to various water

quality parameters make them crucial tools for moni-

toring pollution levels, habitat integrity, and landscape

disturbances. By understanding their taxonomic clas-

siﬁcation and ecological relationships, we can eﬀec-

tively utilise odonates to identify environmental issues

and implement appropriate conservation measures to

safeguard the health of freshwater ecosystems.

The upcoming article will include the taxonomic

description of the Clean-water (unpolluted water) in-

dicator species of odonates. Continued. . . . . .

References

https://en.wikipedia.org/wiki/Brachydiplax

chalybea

https://uk.inaturalist.org/taxa/111765-Rhodothemis-rufa

https://indiabiodiversity.org/species/show/228180

1.2 Species Description of Odonata Species Commonly Found in Polluted Waters

About the Author

Geetha Paul is one of the directors of

airis4D. She leads the Biosciences Division. Her

research interests extends from Cell & Molecular Bi-

ology to Environmental Sciences, Odonatology, and

Aquatic Biology.

Introduction to Aging Clocks - Part 2

by Jinsu Ann Mathew

airis4D, Vol.2, No.3, 2024

www.airis4d.com

In our previous exploration of the aging clock, we

delved into various biomarkers like telomere length,

protein levels, DNA methylation and more. Today,

we’ll narrow our focus to the fascinating world of DNA

methylation.

Imagine our DNA as a musical score. The notes

represent the genes, dictating cellular functions. But

what if there were subtle markings on the sheet mu-

sic, inﬂuencing how loudly or softly each note plays?

That’s precisely what DNA methylation does. It adds

methyl groups to speciﬁc DNA regions, acting like a

dimmer switch, turning genes on or oﬀ.

At its core, DNA methylation is a fundamental

epigenetic mechanism that regulates gene expression

without altering the underlying DNA sequence. This

process involves the addition of methyl groups to cy-

tosine bases, primarily occurring within CpG dinu-

cleotides across the genome.

2.1 The Enzyme: DNA

Methyltransferases (DNMTs)

DNA methyltransferases (DNMTs) are a group of

enzymes responsible for adding methyl groups to DNA

molecules, thereby catalyzing the process of DNA methy-

lation. These enzymes play a crucial role in regulat-

ing gene expression and maintaining genomic stability

by modifying the epigenetic landscape of the genome.

There are several types of DNMTs, each with distinct

functions:

DNMT1: This enzyme is often referred to as

the maintenance methyltransferase because it primar-

ily maintains existing DNA methylation patterns during

(image courtesy:https://www.researchgate.net/ﬁgure/Members-of-mammalian-DNMTs-family-

DNMT1-DNMT3A-and-DNMT3B-consist-of-N-terminal ﬁg2 6815995)

Figure 1: Maintenance DNMTs (DNMT1)

(image courtesy:https://www.researchgate.net/ﬁgure/Members-of-mammalian-DNMTs-family-

DNMT1-DNMT3A-and-DNMT3B-consist-of-N-terminal ﬁg2 6815995)

Figure 2: De Novo DNMTs (DNMT3A and

DNMT3B)

DNA replication. DNMT1 ensures the faithful inheri-

tance of DNA methylation patterns to daughter strands

following DNA replication(Figure 1). It recognizes

hemimethylated DNA, where one strand is methylated

and the other is not, and methylates the unmethylated

strand to maintain methylation patterns.

DNMT3A and DNMT3B: These enzymes are

responsible for de novo DNA methylation, which in-

volves adding methyl groups to previously unmethy-

lated DNA regions(Figure 2). DNMT3A and DNMT3B

establish new DNA methylation patterns during early

development, embryogenesis, and cellular diﬀerentia-

tion. They play crucial roles in shaping the epigenetic

landscape of the genome and regulating gene expres-

sion during these processes.

2.2 The Target: CpG Sites

(image courtesy:https://en.wikipedia.org/wiki/CpG site)

Figure 3: a CpG site, i.e., the ”

5’—C—phosphate—G—3’ ” sequence of nu-

cleotides, is indicated on one DNA strand (in yellow).

On the reverse DNA strand (in blue), the complemen-

tary 5’—CpG—3’ site is shown.

2.2 The Target: CpG Sites

While we’ve explored the ”engineers” of DNA

methylation, namely the DNA methyltransferases (DN-

MTs), they wouldn’t be able to do their job without a

speciﬁc target. This target is a unique sequence within

the DNA code known as a CpG site.

Understanding these sites is crucial for compre-

hending how DNA methylation works. Simply put,

CpG refers to a location where a cytosine (C) nu-

cleotide sits next to a guanine (G) nucleotide in the

DNA strand(Figure 3). Interestingly, these sites often

cluster together in speciﬁc regions called CpG islands.

These islands, especially prevalent near the beginning

(promoter) of genes, are particularly important for reg-

ulating their activity.

But why are CpG sites the target of choice for

DNA methylation? Initially, such sites were much

more abundant in the genome. However, over time,

they became susceptible to mutations, where the cy-

tosine (C) could spontaneously change to a thymine

(T), potentially leading to harmful consequences. As

a result, CpG sites became scarce in most areas of the

genome, oﬀering a layer of protection against such mu-

tations. However, in speciﬁc regions like CpG islands,

where mutations are less detrimental, CpG sites are re-

tained and serve as valuable targets for the methylation

process.

The presence of a methyl group on a CpG site,

particularly within a CpG island, often leads to gene

silencing. This is because the methyl group acts as a

physical barrier or attracts proteins that condense the

DNA, making it inaccessible for the machinery needed

to read and express the gene. It’s important to note that

not all CpG sites are created equal. While methylation

in CpG islands often leads to gene silencing, methyla-

tion in other regions can have diﬀerent, and sometimes

even opposite, eﬀects. Understanding the nuances of

these diﬀerences is crucial for unraveling the intricate

dance between DNA methylation and various biologi-

cal processes.

2.3 The Chemical Tag: Adding the

Methyl Group

We’ve already explored the dynamic duo of DNA

methylation: the DNA methyltransferases (DNMTs),

acting as the engineers, and the CpG sites, serving as

their targets. Now, let’s delve deeper into the fascinat-

ing process of adding a methyl group to the DNA, a

crucial step in regulating gene expression.

Imagine these DNMTs as skilled engineers, need-

ing a speciﬁc tool for the job. This tool is called S-

adenosyl methionine (SAM), a molecule that plays a

vital role by acting as a methyl group donor. Think of

SAM as a small container holding the essential chem-

ical tags (methyl groups) needed for the process.

Here’s a breakdown of the remarkable steps in-

volved in this intricate process:

Targeting the Site: The ﬁrst step involves the

DNMT identifying its target – a speciﬁc sequence on

the DNA strand known as a CpG site. Recall that this

site consists of a cytosine (C) nucleotide followed by a

guanine (G) nucleotide.

Extracting the Tag: Once the DNMT ﬁnds its

target, it binds tightly to the DNA at the CpG site. Then,

like reaching into the SAM ”container,” it extracts a

methyl group (-CH3) from the molecule.

Precise Attachment: With the methyl group in

hand, the DNMT acts with remarkable precision. It

meticulously attaches the methyl group to the 5th car-

bon atom of the cytosine base within the CpG site. This

speciﬁc location is crucial for the regulatory eﬀects of

methylation.

Transformation and Silencing: This action trans-

forms the original cytosine (C) into a modiﬁed version

2.4 Conclusion

(image

courtesy:https://e-trd.org/journal/view.php?number=18&viewtype=pubreader#!po=19.5652)

Figure 4: Addition of a methyl group to a cytosine

residue in a CpG dinucleotide

called 5-methylcytosine(Figure 4). This modiﬁed base,

sporting its newly acquired methyl group, often leads

to gene silencing. The methyl group acts as a physical

barrier or attracts repressor proteins, making it diﬃcult

for the cellular machinery responsible for gene expres-

sion to access and read the gene.

It’s important to remember that this process isn’t

a one-way street. While adding a methyl group often

leads to gene silencing, there are mechanisms for re-

moving them, allowing for dynamic regulation of gene

expression. Think of it like putting a lock on a door:

when you add a methyl group, it’s like locking the door,

preventing the gene from being used. But, just as you

can unlock the door when you need to get in, cells have

ways to remove the methyl group, allowing the gene to

be active again. This process gives cells control over

which genes are turned on or oﬀ, kind of like having

a key to open or lock diﬀerent doors in a house. So,

while adding a methyl group can silence a gene, cells

also have the ability to unlock it, providing ﬂexibility

in how they respond to their environment.

2.4 Conclusion

In conclusion, DNA methylation, facilitated by

DNA methyltransferases (DNMTs), represents a fun-

damental mechanism for regulating gene expression

and chromatin structure. By targeting speciﬁc DNA

sequences known as CpG sites, DNMTs add methyl

groups to cytosine bases, inﬂuencing the activity of

nearby genes. CpG islands, regions of DNA rich in

CpG dinucleotides, serve as primary targets for DNA

methylation and play crucial roles in gene regulation

and genomic stability. The addition of methyl groups

to CpG sites can lead to gene silencing and chro-

matin compaction, aﬀecting diverse biological pro-

cesses such as development, diﬀerentiation, and dis-

ease. Importantly, DNA methylation is not a one-

way street; there are mechanisms for removing methyl

groups, allowing for dynamic regulation of gene ex-

pression. This interplay between DNA methylation and

demethylation underscores the plasticity of the genome

and its ability to respond to environmental cues and de-

velopmental signals. Overall, understanding the roles

of DNA methyltransferases, CpG sites, and the addition

of methyl groups provides insights into the complex-

ities of epigenetic regulation and its implications for

health and disease.

References

DNA Methylation and Its Basic Function

DNA Methylation

A Deep Dive Into DNA Methylation

DNA methylation and DNA methyltransferases

Factors underlying variable DNA methylation in

a human community cohort

The Genomic Impact of DNA CpG Methylation

on Gene Expression

About the Author

Jinsu Ann Mathew is a research scholar

in Natural Language Processing and Chemical Infor-

matics. Her interests include applying basic scientiﬁc

research on computational linguistics, practical appli-

cations of human language technology, and interdisci-

plinary work in computational physics.

Part IV

Remote Sensing

AI for Remote Sensing

by Balamuralidhar P

airis4D, Vol.2, No.3, 2024

www.airis4d.com

Space based remote sensing is already proven as

an impactful technology for earth observation at vary-

ing quality and granularity. It is all the more important

contributor to the sustainability initiatives across the

world. Opening up of space sector for private players

has contributed to the expansion of space industry that

made space access much easier and cheaper. Several

startups in this ﬁeld are creating headlines with their

achievements these days. Large monolithic satellites

are giving way to smaller and cheaper satellite con-

stellations that can be managed with lot of agility and

ﬂexibility. With powerful and compact sensors and

embedded technologies on these platforms triggering

a disruption in space based remote sensing.

This is the era of Artiﬁcial Intelligence and its

application in remote sensing is revolutionizing data

analysis and applications in many domains of Earth

sciences. The diverse applications of AI, speciﬁcally

machine learning (ML), in remote sensing includes

image classiﬁcation, land cover mapping, object de-

tection, change detection, hyperspectral and radar data

analysis, and data fusion.

In addition to data processing and analysis, AI

has made impact on other aspects of remote sensing

that include satellite coordination & control, onboard

processing, ground systems processing and integration.

In this article we would focus on the data analysis aspect

only.

Remote sensing is a technology that enables data

collection at a distance from the subject. It generally

uses sensors that detects emitted, reﬂected or scattered

electromagnetic energy from the target revealing its

characteristics. Multiple sensors and platforms have

Figure 1: Optical (a) and SAR(b) imagery of same

location

been developed and used for remote sensing that in-

clude optical cameras in visible, multi-spectral and

hyper spectral band structure, RADAR, LIDAR and

others. Some of the powerful satellites currently in or-

bit are capable of capturing images with resolutions as

ﬁne as 30 centimeters and even better. Rise of satellite

constellations providing near-real-time data at a global

scale, allowing for more accurate and timely decision-

making in industries such as agriculture, forestry, util-

ities and disaster response.

Remote sensing can be achieved with sensor pay-

loads mounted on multiple platforms including Un-

manned Aerial Vehicles, Unmanned ground Vehicles,

aircrafts and satellites. Suitable one should be chosen

based on the target application, resolution, latency, cov-

erage and cost. A summary of various sensor modal-

ities and their target applications are given in Table

1.1.

There are many impactful applications of remote

sensing. One major application of satellite-based re-

mote sensing is in weather forecasting. Satellites equipped

with sensors can monitor weather patterns, track storms,

and gather data on atmospheric conditions, which helps

meteorologists predict and monitor weather events such

as hurricanes, tornadoes, and blizzards.

Another important application is in agriculture.

Satellites can provide valuable information about crop

health, soil moisture levels, and potential areas of pest

infestation. This data can help farmers make informed

decisions about irrigation, fertilization, and pest con-

trol, ultimately improving crop yields and reducing

environmental impact. Satellite remote sensing is also

used in environmental monitoring and conservation ef-

forts. By collecting data on deforestation, urbanization,

pollution levels, and changes in land use, scientists

can better understand and protect natural ecosystems,

wildlife habitats, and endangered species.

Additionally, satellite-based remote sensing is used

in disaster management, urban planning, transporta-

tion, and many other ﬁelds. Overall, satellite technol-

ogy plays a crucial role in providing valuable informa-

tion for decision-making, resource management, and

global monitoring.

Some of the emerging space based remote sensing

applications to exemplify the potential of this technol-

ogy are listed below:

Wildﬁre Detection and Management

Illegal Logging and Deforestation Monitoring

Coastal and Marine Ecosystem Monitoring

Biodiversity Conservation and Habitat Monitor-

ing

Airborne Disease Monitoring and Forecasting

Precision Forestry

Urban Heat Island Mitigation

Precision Water Management

Disaster Resilience Planning

With the increasing data quality and volume from re-

mote sensing platforms, there is a need for computa-

tional platforms and eﬀective tools to handle and ex-

tract valuable information from remote sensing datasets.

AI-powered onboard and ground processing systems

are catching up, speciﬁcally orbital edge computing, to

meet the challenge. Orbital edge computing enables

satellites in orbit to autonomously handle critical tasks

like data capture, calibration, ﬁltering, compression

and other data processing tasks on board. It targets

to optimize the computation, storage and communica-

tion bandwidth to achieve timely and economic satellite

constellation operation with ground stations.

There are two highly useful free satellite imagery

data platform for application developers and users namely

a) Bhuvan – geospatial data platform from ISRO [1] b)

Sentinel Hub – EU data platform [2].

AI techniques for satellite image processing

The integration of AI techniques in remote sens-

ing has emerged as a powerful paradigm with tremen-

dous potential for practical applications. This has

accelerated the advancement of our understanding of

Earth’s dynamics, support decision-making processes,

and foster sustainable development.

Satellite images can be signiﬁcantly diﬀerent from

natural images -- they can be multi-spectral beyond

what we see in visual bands, irregularly sampled across

time -- and many AI models trained on images from

the Web do not support them. Furthermore, remote

sensing data is inherently spatio-temporal, requiring

conditional processing to get best out of it.

There are majorly three primary applications, namely,

a) classiﬁcation that groups similar pixels together, b)

segmentation that involves dividing the image into dif-

ferent regions to detect objects, and c) denoising - mak-

ing an estimate of the obtained image. Application

areas utilizing satellite image data such as change de-

tection, land use land cover, vegetation monitoring, etc.

require classiﬁcation of satellite image under investi-

gation. Whereas, segmentation is required for urban

growth monitoring, road extraction, building extraction

and detection, etc., and denoising is a preprocessing to

improve the quality of images.

Traditionally there are several classical techniques

used for satellite image processing analysis. A quick

summary of prominent techniques for various func-

tionality are listed below:

Gaussian Mixture Models (GMM) for Denois-

ing.

Principal Component Analysis (PCA), Local lin-

ear embedding (LLE), Isometric Mapping (ISOMAP)

for Dimensionality reduction.

Sparse coding for Sparse representation.

HOG, SURF, SIFT, decision trees, random for-

est, genetic algorithm, HMRF, SVM, MRF for

Feature selection.

Decision tree, multilayer perceptron, logistic re-

gression, SVM, K-means, GP, NN, ARIMA for

Segmentation and classiﬁcation.

KNN, SVM, SOM, GMM, K-means, fuzzy clus-

tering, hierarchical clustering, hybrid clustering

for Clustering

Image transformations, correlation analysis for

Change detection.

There several publications in the literature on these

techniques applied to satellite image analysis. The

review papers [3] and [4] may be consulted to get a

good list of references and pointers.

Some of the recent trends in ML algorithms for

satellite image analysis

Manifold learning techniques are successfully be-

ing used for dimensionality reduction and nonlinear

feature extraction. Local linear embedding (LLE) trans-

forms a very high-dimensional space-embedded im-

ages into two dimensional and makes visualization and

processing much simpler.

Semisupervised learning, which makes use of both

labeled data and the wealth of unlabeled data samples

for development of model using manifold data struc-

ture. In perspective of remote sensing data, a variety

of methods have been developed that are either gener-

ative or discriminative. As an example [6] shows how

domain adaptation approach enables semisupervised

learning.

Transfer learning or domain adaptation can be

used when training samples are available for a par-

ticular time, and we need to classify future time series

data , for example, to update land cover maps.

Active learning approach contributes to improve

the quality of training through intelligent selection of

most relevant sample to train the model.

Structured learning is an approach where multi-

ple labels can be predicted simultaneously; computer

vision-based structured SVM (SSVM) is one such al-

gorithm.

Family of techniques emerging under deep neural

networks enables automatic feature extraction from the

data without human intervention. This tremendously

help in the scale up of processing huge volumes of

satellite data.

A summary of prominent AI algorithms being

used in satellite image analysis is given in Table 1.2.

Classiﬁcation of remotely sensed data is inher-

ently a semi-supervised classiﬁcation problem. Often,

the labeled pixels and the unlabeled pixels in the image

may have diﬀerent distribution. Hence classiﬁcation

accuracy of such images is aﬀected. The paper [7]

presents a semi-supervised learning that considers the

domains shifts in labeled and unlabeled pixels (called

Domain Aware Semi-supervised learning- DASSL).

Following the success of vision transformers that

uses attention mechanisms in computer vision, it has

been successfully applied with additional noval tech-

niques. Temporo-Spatial Vision Transformer (TSViT),is

one such fully-attentional model for general Satellite

Image Time Series (SITS) processing based on the Vi-

sion Transformer (ViT). TSViT splits a SITS record

into non-overlapping patches in space and time which

are tokenized and subsequently processed by a factor-

ized temporo-spatial encoder [8].

Recently, the ﬂourishing large language models

(LLM), especially ChatGPT, have shown exceptional

performance in language understanding, reasoning, and

interaction, attracting users and researchers from mul-

tiple ﬁelds and domains. Although LLMs have shown

great capacity to perform human-like task accomplish-

ment in natural language and natural image, their po-

tential in handling remote sensing interpretation tasks

has not yet been fully explored. Moreover, the lack

of automation in remote sensing task planning hin-

ders the accessibility of remote sensing interpretation

techniques, especially to non-remote sensing experts

from multiple research ﬁelds. Recent developments

like Remote Sensing ChatGPT, an LLM-powered agent

that utilizes ChatGPT to connect various AI-based re-

mote sensing models to solve complicated interpreta-

tion tasks are promising [7]. Here, given a user request

and a remote sensing image, ChatGPT is utilized to

understand user requests, perform task planning ac-

cording to the tasks’ functions, execute each subtask

iteratively, and generate the ﬁnal response according to

the output of each subtask.

Foundation models are ﬂexible deep learning al-

gorithms that are designed for general tasks, rather than

being immediately focused on speciﬁc tasks. Trained

on large amounts of unlabeled data, they can be ap-

plied to a variety of downstream tasks with minimal

ﬁne-tuning. In last year, NASA and IBM released

the Geospatial AI Foundation Model for NASA Earth

Observation Data [10]. The model is available open

source on Huggingface under the name of Prithvi, the

Hindu goddess of Mother Earth.

Application of generative AI is also gathering mo-

mentum in satellite image analysis. In one such work

on image classiﬁcation, generative adversarial network

(GAN) is used for data augmentation following by vi-

sion transformer for classiﬁcation. DiﬀusionSat, to

date the largest generative foundation model trained on

a collection of publicly available large, high-resolution

remote sensing datasets.

AI deployment for satellite image analysis has sev-

eral challenges:

Training: Training AI algorithms, especially deep

learning models, requires signiﬁcant computa-

tional resources, making them challenging to de-

velop on resource-constrained shared devices.

Explainability: Many neural network-based mod-

els are often considered black-box models, and

understanding the reasons behind AI predictions

is diﬃcult but critical for gaining trust and en-

suring eﬀective decision making.

Training data: Creating labeled datasets for train-

ing AI models in remote sensing can be labor-

intensive and time consuming, especially for ﬁne-

grained or multi-class tasks.

Domain adaptation: Transferring AI models trained

on one dataset to perform well on diﬀerent datasets

can also require additional resources.

Domain Knowledge integration: Incorporating

domain-speciﬁc knowledge and expertise into AI

models is essential to ensure the representation

of relevant features and relationships.

References

1. Bhuvan – Indian Geo Platform of ISRO

2. Sentinel Hub – Europen satellite data platform

3. A Review of Practical AI for Remote Sensing in

Earth Sciences, Bhargavi Janga,Gokul Prathin

Asamani ,Ziheng Sun,Nicoleta Cristea

4. Next-Generation Artiﬁcial Intelligence Techniques

for Satellite Data Processing, Neha Sisodiya et.al,

Book chapter of the book “Artiﬁcial Intelligence

Techniques for Satellite Image Analysis”

5. Exploiting High Geopositioning Accuracy of SAR

Data to Obtain Accurate Geometric Orientation

of Optical Satellite Images, Zhongli Fan et.al,

Remote Sensing MDPI , 13(17):3535

6. Semi-supervised learning by domain adaptation

for hyperspectral image, Shailesh Deshpande,

Chaman Banolia, Balamuralidhar P., IEEE IGRASS

2023

7. Remote Sensing ChatGPT: Solving Remote Sens-

ing Tasks with ChatGPT and Visual Models,

Haonan Guo, Xin Su, Chen Wu, Bo Du, Liang-

pei Zhang, Deren Li

8. ViTs for SITS: Vision Transformers for Satellite

Image Time Series, Michail Tarasiou et al.

9. Improving satellite image classiﬁcation accuracy

using GAN-based data augmentation and vision

transformers, Ayyub Alzahem et al

10. A Foundation Model for Satellite Images

Table 1.1: Summary of sensor modalities and popular applications

Technique Highlights Limitations Sample Applications

Optical remote

sensing ⍟ Imaging with easily available

camera technology in visible and

infrared spectrum

⍟ Easy to visualize and interpret

⍟ Higher resolution

⍟ Visibility aﬀected by

atmospheric conditions,

rain, cloud, sun angles

etc

⍟ Might-time nonvisi-

bility

⍟ Land-use mapping

⍟ Crop health assess-

ment

⍟ Monitoring vegeta-

tion, forest

⍟ Monitoring climate

change

⍟ Object detection, as-

set monitoring

Radar remote

sensing ⍟ Operates in the microwave re-

gion, providing data on distance,

direction, shape, size, roughness,

and dielectric properties of tar-

gets

⍟ Can penetrate clouds and

other limited visibility conditions

day & night

⍟ Utilizes dual-polarization

technology

⍟ Interferometry for high preci-

sion change detection

⍟ Data processing can

be complex, including

removal of unwanted ar-

tifacts and uncertainities

⍟ Image interpretation

not very easy

⍟ High sensitivity to

surface roughness

⍟ Back scatter has de-

pendency on look angle

⍟ Mapping land

surfaces and monitoring

weather patterns

⍟ Studying ocean cur-

rents, ice

⍟ Detecting buildings,

vehicles, and changes in

forest cover

⍟ Detection of land sub-

sidence

LiDAR

⍟ Provides precise distance

and elevation measurements of

ground objects

⍟ High-resolution 3D data

⍟ Penetration of vegetation

⍟ Day and night operation

⍟ Multiple returns of one sin-

gle laser pulse and reduced atmo-

spheric interference

⍟ Data processing com-

plexity, especially for full

waveform LiDAR sys-

tems

⍟ Accuracy dependent

on elevation and angle

high cost and availability

⍟ Limited penetration

through thick dense veg-

etation

⍟ Create accu-

rate and detailed 3D

maps of trees, buildings,

pipelines, etc

Thermal re-

mote sensing ⍟ Measures radiant ﬂux emitted

by ground objects within speciﬁc

wavelength ranges

⍟ Provides information on the

emissivity, reﬂectivity, and tem-

perature of target objects

⍟ Atmospheric con-

ditions, changes in so-

lar illumination, and tar-

get variations can impact

data accuracy

⍟ Agriculture (e.g., ﬁre

detection, urban heat is-

lands) and environmen-

tal monitoring

Multispectral

and Hy-

perspectral

imaging

⍟ Captures a broad range of

wavelengths, including infrared

and ultraviolet, for comprehen-

sive data collection

⍟ Provides insights into material

composition, structure, and con-

dition

⍟ High-dimensional

and noisy data in HSI

pose analysis challenges

⍟ Trade-oﬀ between

spatial and spectral res-

olution

⍟ Recognition of veg-

etation patterns such as

greenness, vitality, and

biomass

⍟ Material classiﬁca-

tion

⍟ Mineral resource

monitoring

Table 1.2: Summary of major AI algorithms in use for satellite image analysis ( adapted from [1] [2])

Technique Advantages Limitations Applications

Random Forest

⍟ Eﬀectively handles multi-

temporal and multi-sensor re-

mote sensing data

⍟ Provide variable importance

measurements for feature selec-

tion

⍟ Enhances generalization and

reduces computational load and

redundancy

⍟ Superior feature selection by

evaluating interrelationships and

discriminating ability in high-

dimensional remote sensing data

⍟ Can be sensitive

to the choice of hyper-

parameters

⍟ Does not guarantee

that the selected features

will be the best for all

tasks

⍟ Classiﬁcation

⍟ Object detection

XGBoost (eX-

treme Gradient

Boosting)

⍟ The ability to handle cases

where diﬀerent classes exhibit

similar spectral signatures

⍟ Eﬀective diﬀerentiation of

classes with subtle spectral dif-

ferences, enhancing classiﬁcation

performance.

⍟ Utilization of hyper-parameter

tuning techniques to ensure opti-

mal accuracy and prevent overﬁt-

ting

⍟ Hyperparameter sen-

sitivity

⍟ Pone to overﬁtting

⍟ Slower than RF

⍟ Classiﬁcation with

robustness

Deep Convolu-

tional Neural

Networks

(DCNN)

⍟ Eﬃciently handle intricate

patterns and features

⍟ Learn hierarchical representa-

tions of features from convolution

and pooling layers

⍟ Training DCNNs

can be computationally

expensive, especially for

large-scale datasets

⍟ May suﬀer from van-

ishing gradients or over-

ﬁtting if not properly reg-

ularized

⍟ Remote sensing im-

age recognition and clas-

siﬁcation

⍟ Object detection tasks

in remote sensing using

RPN

ResNets

⍟ Alleviate the degradation

problem in deep learning mod-

els, allowing the training of much

deeper networks

⍟ Handling complex high-

dimensional and noisy data in re-

mote sensing

⍟ Implementing very

deep networks may still

require signiﬁcant com-

putational resources

⍟ Image recognition

⍟ Object detection

Self Attention

methods –

Transformer

Models

⍟ Capture long-range depen-

dencies in sequences and handle

spatial and spectral dependencies

in remote sensing data

⍟ Provide access to all elements

in a sequence, enabling a compre-

hensive understanding of depen-

dencies

⍟ Transformer models

can be memory-intensive

due to their self-attention

mechanism

⍟ Properly tuning

the number of attention

heads and layers is es-

sential for optimal per-

formance

⍟ Sequence modeling

and image classiﬁcation

⍟ Time series analysis

of remote sensing data

and capture diverse pixel

relationships regardless

of spatial distance

Recursive Neu-

ral Networks (

LSTM – Long

and Short

Term Memory)

⍟ Eﬀectively captures long-term

dependencies in sequences

⍟ Overcomes the vanishing gra-

dient problem with gate mecha-

nisms

⍟ Training LSTMs can

be time consuming, par-

ticularly for longer se-

quences

⍟ Can struggle

with capturing very long-

term dependencies in se-

quences

⍟ May require careful

tuning of hyperparame-

ters to prevent overﬁtting

⍟ Sequence modeling

and time series analysis

in remote sensing data

Generative Ad-

versarial Net-

works (GAN)

⍟ Capable of handling com-

plex, high-dimensional data dis-

tributions with limited or no an-

notated training data

⍟ Data augmentation method

enhances the performance of

data-reliant deep learning models

⍟ Training GANs can

be challenging and un-

stable, requiring careful

hyper-parameter tuning

⍟ Generating high-

quality, realistic images

may be diﬃcult in some

cases

⍟ May suﬀer from

mode collapse, where the

generator produces lim-

ited variations in images

⍟ Image-to-image

translation tasks like

converting satellite im-

ages with cloud coverage

into cloud-free versions

using CycleGAN

⍟ Enhancing the res-

olution of low-resolution

satellite images with SR-

GAN and similar ap-

proaches

⍟ Image-to-image

translation, data aug-

mentation, and pan-

sharpening

Deep Rein-

forcement

Learning

(DRL)

⍟ Learns from unlabeled data

to improve decision-making pro-

cesses

⍟ Combines reinforcement

learning (RL) with deep neu-

ral networks for solving complex

problems

⍟ Handles redundant spectral

information

⍟ Requires careful de-

sign and tuning of reward

functions to ensure the

desired behav

⍟ Training deep neu-

ral networks in DRL can

be computationally ex-

pensive and time con-

suming

⍟ Exploration vs. ex-

ploitation trade-oﬀ in RL

can impact the learning

process and can be de-

pendent on the sample

⍟ Improving unsu-

pervised band selection

in hyperspectral image

classiﬁcation using DRL

with DQN

⍟ Image processing

applications that analyze

large amounts of data

About the Author

Dr.Balamuralidhar Purushothaman is a former Chief Scientist at TCS Research Bangalore. Currently

he is continuing in TCS Research as a research advisor. He obtained his PhD from Aalborg university Denmark,

MTech from IIT Kanpur and BTech from TKM Engg Quilon. His research works are in IoT, Sensor Informatics,

Remote Sensing and Robotics. He has over 170 publications and 110 patents in these technology and application

areas. Of late he has also picked up interest in Neuroarts and Neurophilosophy.

About airis4D

Artiﬁcial Intelligence Research and Intelligent Systems (airis4D) is an AI and Bio-sciences Research Centre.

The Centre aims to create new knowledge in the ﬁeld of Space Science, Astronomy, Robotics, Agri Science,

Industry, and Biodiversity to bring Progress and Plenitude to the People and the Planet.

Vision

Humanity is in the 4th Industrial Revolution era, which operates on a cyber-physical production system. Cutting-

edge research and development in science and technology to create new knowledge and skills become the key to

the new world economy. Most of the resources for this goal can be harnessed by integrating biological systems

with intelligent computing systems oﬀered by AI. The future survival of humans, animals, and the ecosystem

depends on how eﬃciently the realities and resources are responsibly used for abundance and wellness. Artiﬁcial

intelligence Research and Intelligent Systems pursue this vision and look for the best actions that ensure an

abundant environment and ecosystem for the planet and the people.

Mission Statement

The 4D in airis4D represents the mission to Dream, Design, Develop, and Deploy Knowledge with the ﬁre of

commitment and dedication towards humanity and the ecosystem.

Dream

To promote the unlimited human potential to dream the impossible.

Design

To nurture the human capacity to articulate a dream and logically realise it.

Develop

To assist the talents to materialise a design into a product, a service, a knowledge that beneﬁts the community

and the planet.

Deploy

To realise and educate humanity that a knowledge that is not deployed makes no diﬀerence by its absence.

Campus

Situated in a lush green village campus in Thelliyoor, Kerala, India, airis4D was established under the auspicious

of SEED Foundation (Susthiratha, Environment, Education Development Foundation) a not-for-proﬁt company

for promoting Education, Research. Engineering, Biology, Development, etc.

The whole campus is powered by Solar power and has a rain harvesting facility to provide suﬃcient water supply

for up to three months of drought. The computing facility in the campus is accessible from anywhere through a

dedicated optical ﬁbre internet connectivity 24×7.

There is a freshwater stream that originates from the nearby hills and ﬂows through the middle of the campus.

The campus is a noted habitat for the biodiversity of tropical Fauna and Flora. airis4D carry out periodic and

systematic water quality and species diversity surveys in the region to ensure its richness. It is our pride that

the site has consistently been environment-friendly and rich in biodiversity. airis4D is also growing fruit plants

that can feed birds and provide water bodies to survive the drought.