Problem statement list:

1)

2)

Description

Persistent cloud cover is a major challenge in optical remote sensing, particularly over tropical
and mountainous regions such as the North Eastern Region (NER) of India. Clouds and cloud
shadows significantly reduce the usability of optical satellite imagery for applications such as
land use–land cover mapping, disaster monitoring, environmental assessment, and
infrastructure analysis. LISS-IV imagery provides high spatial resolution data that is valuable for
detailed geospatial analysis. However, frequent cloud contamination limits the temporal
availability of usable observations. Traditional cloud masking techniques often lead to loss of
information and incomplete scene interpretation.

This problem statement focuses on developing a Generative AI-based framework for automated
cloud removal and surface reconstruction in LISS-IV imagery. The framework should leverage
spatial, spectral, and temporal information to generate cloud-free imagery while preserving fine-
scale spatial details and spectral consistency. The proposed solution may explore advanced
deep learning approaches such as diffusion models, GANs, transformer-based architectures, or
multi-modal fusion techniques using auxiliary data sources such as Sentinel-1 SAR imagery,
Sentinel-2 optical imagery, or temporal reference observations.

3)

● Air pollution is a major health hazard that’s liable for many annual excess
deaths worldwide. The concept of air quality index (AQI) that transforms
weighted values of individual air pollution related parameters into a single
 number or set of numbers is widely used for air quality communication and
decision making in many countries. Hence the AQI is a rating scale for
reporting the Ambient air pollution (CO, NO2, O3, SO2, and PM2.5 etc)
recorded at monitoring sites on a particular time scale (e.g., daily). The two
main objectives of AQI are (a) to inform and caution the public about the risk of
exposure to daily pollution levels and (b) to enforce required regulatory
measures by competent authority for immediate local impact. The higher the
AQI value, the greater the level of air pollution and health risk. Most of the
global population lives further than 100 km from air quality monitors and devoid
of air quality information in that area. Hence it is important to develop a surface
AQI derived from satellite data to know its spatial distribution on daily basis.

VOC’s can be monitored from Global Ozone Monitoring Experiment (GOME), the Scanning
Imaging Absorption Spectrometer for Atmospheric Cartography (SCIAMACHY), the Ozone
Monitoring Instrument (OMI), and the TROPOspheric Monitoring Instrument (TROPOMI), play
pivotal roles in retrieving HCHO columns. Formaldehyde (HCHO) is a key indicator of volatile
organic compound (VOC) emissions and plays an important role in ozone formation and
atmospheric chemistry. During biomass burning events (such as agricultural residue burning
and forest fires), large amounts of VOCs are released, leading to enhanced HCHO
concentrations. This problem aims to identify and analyze spatio-temporal hotspots of HCHO
over India using satellite observations, with a focus on biomass burning periods.

4)Modern urban centres, particularly rapidly expanding Indian metropolises (e.g.

Bengaluru), face a dual challenge in spatial modelling: fragmentation and stagnation.

Standard satellite-based road extraction often fails due to "spectral blindness" caused
by tree canopies, building shadows and cloud cover. These "broken" masks are useless
for real-world applications like disaster response or traffic simulation because they lack
topological connectivity. This solution aims to bridge this gap by creating an end-to-end
pipeline: first, using context-aware Deep Learning to "see through" occlusions, and
second, transforming those masks into a mathematically continuous, weighted graph to
identify systemic bottlenecks and simulate urban collapse scenarios.
5)

A digital twin of the climate system targeting adaptation is expected to make use of
observations, integrate several climate models to consider uncertainty sources, include
applications for climate-sensitive sectors directly connected to the climate models, and
provide interfaces to configure the simulations, output, and data consumers. An “AI-
powered Digital Twin of India’s Climate using national datasets” refers to the creation of
a high-fidelity, dynamic virtual replica of India’s climate system that continuously evolves
using real-time and historical observations. This digital twin integrates multi-source data
from Indian satellites (e.g., INSAT, Oceansat), ground-based meteorological networks
(IMD), reanalysis products, and hydrological datasets to simulate atmospheric, oceanic,
and land-surface processes at high spatial and temporal resolution. Leveraging
advances in Artificial Intelligence, Machine Learning, and Data Assimilation, the system
fuses heterogeneous datasets to generate near-real-time climate states and predictive
scenarios. The digital twin can capture complex phenomena such as monsoon
variability, extreme, drought evolution with improved accuracy compared to
conventional models. This framework supports scenario testing and decision-making,
enabling policymakers and scientists to evaluate the impacts of climate change,
urbanization, and land-use alterations. It can also enhance early warning systems for
extreme weather, optimize water resource management, and guide agricultural
planning. By utilizing India’s national data infrastructure and computational capabilities,
this digital twin represents a transformative step toward self-reliant, data-driven climate
intelligence.

6)

Accurate, timely identification of crop types and detection of moisture stress across
phenological stages are critical for precision agriculture, drought monitoring, yield
forecasting, and climate-resilient farm management. Satellite remote sensing—
consisting of optical (e.g., LISS-IV. LISS-III, AWiFS, Sentinel-2, Landsat, MODIS etc.,)
and microwave SAR (e.g., Sentinel-1, EOS-4, NISAR upcoming etc.,)—provides
continuous, large-scale monitoring capabilities. Optical data captures vegetation vigour
and biochemical properties, while microwave data penetrates clouds and is sensitive to
soil moisture and crop structure.

This proposal focuses on developing an AI-driven automated methodology for: Crop

type classification, stage-wise phenological mapping and moisture stress detection
using multi-source satellite data fusion and machine learning/deep learning models, and
actionable irrigation decisions.
7)

Develop an AI-based data analysis pipeline capable of automatically detecting

exoplanet transit signals from noisy astronomical light curve data.

Details: Exoplanet detection through transit photometry requires the identification of

extremely small brightness variations in stars. For light curves of astronomical sources
present in crowded fields, there can be significant contaminations arising from effects
such as stellar blending by foreground or background sources in the aperture, and the
intrinsic noise in the data due to the detector’s response, to name a few. Apart from
contamination, the brightness variations in the light curves can be due to a transiting
planet across the host star's disk, an eclipsing stellar companion in binary star systems,
or even starspots. Different phenomena give rise to distinct features in light curves,
which, however, become difficult to disentangle while dealing with noisy datasets in
crowded fields.

8)The discovery and characterization of water-ice in the lunar South Polar Region is a high-
priority scientific and exploration objective, particularly for enabling sustained human presence
on the Moon. Observations from Chandrayaan-2 have opened new avenues to probe
surface/subsurface using high-resolution optical and radar datasets. The ‘doubly shadowed
craters’ in the lunar permanently shadowed regions (PSRs) provide access to some of the
coldest environments on the Moon that are ideal candidates for long-term volatile preservation.
However, identifying subsurface ice unambiguously and translating these detections into
actionable exploration strategies (landing and rover traversal) remains a key challenge.
9)

Turbulence in the atmosphere distorts a plane-parallel wavefront propagating through it.

An SH-WFS samples this distorted wavefront using an array of small lenslets. This
Microlens Array (MLA) creates a pot-field on the detector, and the deviation of thes e s
pots from their reference position is used to derive the reconstructed wavefront and its
as s ociated zernike coefficients. The conjugate of this reconstructed wavefront is
typically used to derive an actuator map in units of the actuator's stroke length, which is
then fed to a deformable mirror (DM) to correct for this distortion in real-time.

10)
Satellite remote sensing frequently relies on infrared (IR) sensors to capture data during night
time or under adverse weather conditions. However, IR images are inherently monochrome,
suffer from low contrast, and lack the rich semantic textures found in visible-spectrum (RGB)
imagery. This makes it difficult for human analysts and computer vision models to accurately
identify objects like vehicles, buildings, roads, and vegetation.

This challenge focuses on developing an end-to-end framework that simultaneously enhances

the structural details of IR satellite images and predicts realistic colorization (RGB translation).
By transforming low-visibility IR data into high-fidelity, colorized images, this solution will
drastically improve situational awareness and automated object interpretation.

11)Satellite remote sensing systems acquire Earth observation data using different sensors
such as optical, multispectral, hyperspectral, Synthetic Aperture Radar (SAR), and elevation-
based sensors. Each sensor captures different physical characteristics of the Earth surface.
Optical and multispectral images provide rich visual and spectral information, while SAR images
provide structural information and can operate under cloud cover and night-time conditions.

With the rapid growth of satellite image archives, content-based retrieval of relevant images has
become an important requirement. Traditional metadata-based search may not be sufficient
when similar land-cover or land-use regions are observed under different seasons, acquisition
conditions, or sensor modalities. Therefore, this problem focuses on developing a multi-modal
satellite image retrieval system for both same-modal and cross-modal retrieval.
12)

Satellite images for geo-stationary satellites are often captured at fixed intervals (e.g., every 30
minutes for geostationary satellites like INSAT, every 10 minutes for Himawari/GOES satellites
or every few days for polar-orbiting satellites). This leads to limitation in temporal resolution that
restricts near real-time monitoring of dynamic phenomena such as fire, cyclones,
thunderstorms, floods and rapid land changes. Traditional optical-flow based temporal
interpolation is limited with incorrect results (blurred images & other artefacts) and they fail to
capture fast moving or non-linear changes in cloud dynamics.
This problem proposes to develop an AI/ML based Optical Flow frame interpolation technique to
generate intermediate frames between two consecutive satellite images to effectively enhance
temporal resolution, enabling more frequent observations without additional satellite resources.

13)

Modern enterprise and government networks increasingly rely on SD-WAN deployments

running over MPLS underlays to deliver resilient, application-aware connectivity across
distributed branches, datacenters, and cloud environments. As these networks grow in
complexity, operational visibility and response speed become critical to maintaining service
quality and security posture.

Conventional NOC tooling remains predominantly reactive — faults are detected only after user-
visible service degradation has occurred. Two compounding challenges define this operational
gap:

● Reactive detection: threshold-based alerts fire only after performance

thresholds are breached, providing no time for pre-emptive intervention.

● Air-gap constraints: regulated and government environments prohibit cloud-

connected AI inference tools, leaving operators without intelligent guidance in
the most security-sensitive deployments.

This problem statement calls for an autonomous, air-gapped offline AI NOC Copilot that predicts
network failures before operational impact, explains reasoning in natural language, and
operates entirely within an air-gapped network

14)
To develop and demonstrate an algorithm to predict energetic particle fluxes of electrons at
geostationary orbit. The algorithm should be able to predict the harsh radiation fluxes at least 30
to 45 minutes in advance, and also give a reasonable forecast for 6 hours and 12 hours ahead.

15)

Solar flares are sudden, intense bursts of radiation originating from the release of magnetic
energy in the solar atmosphere. These events can trigger severe space weather disruptions,
impacting Earth-bound satellite communications, GPS navigation, and power grids. Anticipating
these disruptions is critical for infrastructure safeguarding.

ISRO’s Aditya-L1 mission continuously monitors the Sun from Lagrange Point L1 using two X-
ray payloads: SoLEXS (Solar Low Energy X-ray Spectrometer, capturing soft X-rays) and
HEL1OS (High Energy L1 Orbiting X-ray Spectrometer, capturing hard X-rays).

This challenge requires participants to build an automated algorithmic pipeline that uses the
combined time-series data from SoLEXS and HEL1OS to detect (nowcast) or predict (forecast)
these solar flares.
PROS AND CONS:

1. Optimizing Urban Heat Mitigation and Cooling Strategies via AI/ML

 Core Objective: Developing a city-scale geospatial AI/ML-based platform, backed with

physics-informed decision-making, to identify nighttime urban heat island hotspots,
quantify socioeconomic and meteorological driving factors, and model simulated cooling
scenarios.
 Pros: * Highly impactful for humanitarian and city-policy applications (mitigating health
hazards from lethal heatwaves).
o Direct utility in urban structural design and resource allocation for municipal
bodies.
 Cons: * City-wide environments are intensely chaotic with multiple co-dependent
variables (demographics, air flow, anthropogenic emissions), making general equations
highly fragile.
o Micro-level adjustments (like individual building geometry changes) are lost
when applying the required city-scale macro-modeling approach.

2. Generative AI-Based Cloud Removal and Reconstruction for LISS-IV

Satellite Imagery

 Core Objective: Utilizing dense image-to-image translation networks (e.g., GANs,

Diffusion Models) to reconstruct the optical land features hidden behind thick cloud
covers on LISS-IV satellite imagery.
 Pros: * Maximizes the operational utility of optical imaging over coastal/monsoon zones
where continuous cloud cover often leaves images completely blind.
o Allows leveraging auxiliary open data networks (like Sentinel-1 SAR radar or
DEMs) to feed details into the generator.
 Cons: * Deep learning models are inherently prone to "hallucinating" geographic ground
details (e.g., generating a false patch of forest where a building actually sits beneath the
cloud).
o Ensuring precise spatial, spectral, and physical alignment across missing data
blocks without a real-time optical reference is exceptionally complex.

3. Development of Surface AQI and Identification of HCHO Hotspots Over

India Using Satellite Data

 Core Objective: Bridging the gap between satellite readings and terrestrial metrics by
using deep learning architectures (like CNN-LSTMs) to mathematically project column-
density observations down into point-specific surface Air Quality Index values, alongside
mapping formaldehydes ($HCHO$) during active agricultural burning seasons.
  Pros: * Dramatically expands air quality surveillance capabilities into deep rural and
forest sectors currently lacking physical Central Pollution Control Board ground stations.
o Captures real-time transport maps tracking the direction of toxic industrial and
fire pluming.
 Cons: * Satellite systems naturally measure atmospheric columnar totals; isolating
exactly what amount rests at ground-breathing level involves battling unpredictable
meteorological filters like humidity, wind drift, and boundary layer shifts.
o Ground-truth metrics are heavily skewed towards metropolitan areas, starving the
deep training model of uniform validation inputs for remote regions.

4. Route Resilience: Occlusion-Robust Road Extraction and Graph

Criticality Analysis for Urban Mobility

 Core Objective: Designing vision transformers to continuously detect and map

interconnected road grid networks obscured under cloud, building shadows, or dense tree
canopies, followed by stress-testing the network through graph theory to map critical
traffic bottlenecks.
 Pros: * Invaluable asset for military route allocation or automated first-responder vehicle
dispatches following natural disasters (e.g., identifying isolated sections).
o Sidesteps manual classification pipelines since existing maps (OpenStreetMap
layers) can automate model training labels.
 Cons: * Medium-scale satellites (like Sentinel's 10m bands) are physically blind to
smaller service bypasses or urban alleys, forcing the pipeline to depend on high-cost,
specialized imagery formats.
o Small local extraction gaps can entirely snap the topological graph continuity,
rendering large-scale automated pathfinding engines useless unless robust graph-
healing functions are integrated.

5. AI-Powered Digital Twin of India’s Climate Using National Data

 Core Objective: Creating a virtual, scalable, data-driven replica of the Indian

subcontinent’s atmospheric behavior using deep learning to instantly compute "What-If"
scenarios (e.g., assessing the exact regional rain shifts if average summer temperatures
lift by $1^\circ\text{C}$).
 Pros: * Runs computationally cheaper and magnitudes faster than traditional, hyper-
dense fluid dynamic supercomputer equations.
o Aggregates decades of deep archives (IMD historical rain arrays extending back
to 1951) for training.
 Cons: * Lacks the hard mathematical constraints embedded within strict physics-based
models, risking non-sensical parameter predictions during extreme, historically
unprecedented weather fluctuations.
o Managing data anomalies, cleaning historical entry gaps, and smoothing diverse
coordinate projections across hundreds of grid layers presents a massive data
engineering hurdle.
6. AI-Driven Automated Crop Type, Moisture Stress Detection, and
Irrigation Advisory Across Growth Stages Using Moderate Resolution
Spectral Signatures

 Core Objective: Engineering a multitemporal, automated sensor fusion system (Optical

+ SAR) to identify varying crop variants across shifting growth periods and compute
absolute water deficit advisories for local reservoir command borders.
 Pros: * Directly drives domestic agricultural yield forecasts and serves as a backbone
infrastructure for large-scale economic systems like localized crop insurance programs.
o Utilizing SAR radar signals ensures cloud-penetrating crop checks during rainy
monsoon cycles.
 Cons: * Indian farming landscapes are notoriously fragmented, featuring highly
heterogeneous crop matrices sharing cramped plots, which creates intense pixel blending
at moderate resolutions.
o Computing accurate water deficit advisories demands intricate soil condition
layers and localized weather tracking variables that go far beyond standard
satellite imagery alone.

7. AI-Enabled Detection of Exoplanets from Noisy Astronomical Light

Curves

 Core Objective: Building an automated deep-learning time-series system capable of

parsing extreme detector noise in stellar light curves (from missions like NASA's TESS)
to identify subtle planet transits and compute orbital durations.
 Pros: * Enables high-speed, automated cataloging of deep-space systems, bypassing
manual search pipelines across millions of observed star signals.
o Accelerates space science by offering deep parameter extraction for celestial
bodies.
 Cons: * Space telescope systems are continuously hit by systemic artifacts, instrumental
drift, and stellar activity that mimic actual planet transits.
o Simple astronomical configurations, such as a smaller star passing in front of a
larger one (eclipsing binary system), generate identical light curve dips, making
precise classification difficult without deep multi-channel analysis.

8. Detection and Characterization of Subsurface Lunar Ice and Mission

Planning for Future Robotic Exploration

 Core Objective: Processing Chandrayaan-2 Dual-Frequency SAR (DFSAR) and High-

Resolution Camera (OHRC) files to detect buried ice caches in cold, doubly shadowed
polar craters, assessing terrain hazards, and plotting a navigable rover transit vector.
 Pros: * A crucial foundational element for establishing long-term survival logistics and
fuel-harvesting outposts for human deep-space habitats.
o Pushes the user to think through a practical, comprehensive lens—bridging
remote data processing with active tactical mission deployment.
  Cons: * Lunar radar readings contain high geometric ambiguity; surface irregularities
and fields of dense jagged rocks mimic the exact circular polarization ratios ($CPR$) cast
by subsurface ice blocks.
o True deep-crater floor details must be derived under total dark environments,
severely restricting visual verification and complicating automated safety
evaluations for landing zones.

9. Developing and Optimizing Algorithms for Wavefront Reconstruction and

Turbulence Characterization Using Shack-Hartmann Wavefront Sensor
Time-Series Data

 Core Objective: Creating low-latency phase retrieval math models or deep networks to
compute localized atmospheric disruptions from sensor grid spot arrays, adjusting
deformable mirror systems for sharp ground telescopy.
 Pros: * Extends critical performance enhancements across multiple adjacent high-tech
domains (free-space laser communication arrays and long-range quantum key
distribution).
o Substantially lowers standard image-processing latency compared to archaic,
heavy loop-matrix iterations.
 Cons: * Atmospheric shifting happens at breakneck speeds, requiring models to process
streams and apply surface changes within mere milliseconds to prevent image blur.
o Any minor drift or noise error introduced within the initial spot-center algorithms
quickly compounds into massive structural mirror errors down the line.

10. Infrared Image Colorization and Enhancement for Improved Object

Interpretation

 Core Objective: Engineering a joint neural pipeline that processes low-resolution

thermal infrared bands ($TIR$ at 200m) to reconstruct them into a 100m high-resolution
frame while mapping realistic, structural RGB equivalents.
 Pros: * Provides a powerful night-reconnaissance asset, helping security grids and
civilian analysts instantly identify buildings, vehicles, or hidden assets through crisp
visual color cues.
o Readily takes advantage of perfectly aligned, dual-band satellite platforms (such
as the Landsat constellation archives) for easy training configurations.
 Cons: * Thermal radiation profiles map structural heat signatures, not true surface colors,
meaning objects sharing identical thermal output values can be assigned false color
states.
o Upsampling low-resolution thermal bands to match crisp structural boundaries
requires complex model controls to prevent blurry, distorted shapes.
11. Cross-Modal Satellite Image Retrieval Using Multi-Sensor Remote
Sensing Data

 Core Objective: Designing unified embedding networks capable of mapping matching

topological structures across completely distinct spectrum spaces (e.g., inputting an
optical picture of an airfield and instantly retrieving matching radar profiles from the
archive).
 Pros: * Provides a massive efficiency boost over outdated metadata querying, enabling
operators to find matching geographic regions regardless of active weather cover or
lighting conditions.
o Breaks down structural data silos between multi-agency satellite platforms.
 Cons: * Optical cameras measure reflected light waves while SAR platforms map
material physical texture and metal rebounds, resulting in a fundamentally deep domain
gap between the dataset pairs.
o The trained retrieval model must maintain swift index computation capabilities
across immense, multi-terabyte search galleries to preserve real-world utility.

12. Fill in the Frames Seamlessly: Enhancing Temporal Resolution of

Satellite Imagery Using AI/ML-Based Optical Flow

 Core Objective: Merging classical optical flow motion-tracking with neural networks to
reconstruct artificial intermediate frames between coarse geostationary imaging intervals
(e.g., synthesizing a 15-minute frame step from 30-minute INSAT feeds).
 Pros: * Provides critical early tracking advantages for identifying fast-moving disaster
events, like rapid lightning build-ups, toxic crop fires, or sudden flash flooding.
o Greatly enhances tracking visuals without demanding adjustments to expensive
satellite sensor hardware.
 Cons: * Highly erratic and dynamic tracking subjects, like churning storm clouds, do not
adhere to predictable, uniform vector paths, causing models to produce unnatural
distortions in intermediate frames.
o Processing the massive data payloads generated by high-frequency file
architectures requires substantial computational resources.

13. Air-Gapped Predictive Copilot for Secure MLOps Operations

 Core Objective: Constructing a completely self-contained, offline Local Large Language

Model ($LLM$) ecosystem running a Retrieval-Augmented Generation ($RAG$)
pipeline over localized enterprise network runbooks to forecast internal route failures and
explain issues without an active internet path.
 Pros: * Guarantees total security for sensitive data pipelines, preventing private
operational network logs from leaking onto public cloud systems.
o Shifts corporate technical teams from a reactive position to a proactive
troubleshooting stance through explainable, clear natural language reports.
  Cons: * Eliminating external cloud API access restricts processing power to available
local hardware, forcing a careful trade-off between model speed and reasoning capability.
o The offline assistant must successfully interpret highly complex, raw log strings
(SNMP feeds, system events) to translate them into reliable, accurate summaries
without hallucinating fake network errors.

14. Forecasting Energetic Particle Radiation Environment for ISRO’s

Geostationary Satellites

 Core Objective: Building a long-term and short-term predictive AI framework that

tracks variable solar wind dynamics to forecast sudden fluxes of high-energy "killer
electrons" ($>2\text{ MeV}$) targeting deep geostationary orbit sectors.
 Pros: * Delivers a critical protection window for space operators, allowing them to
proactively safe high-value satellite systems prior to severe space weather impacts.
o Backed by massive historical training catalogs, utilizing decades of open solar
data.
 Cons: * Solar wind behavior exhibits non-linear complexity, meaning small variations in
deep-space monitoring points can produce vastly different radiation impacts at the
satellite's orbital plane.
o Localized longitude validation archives are heavily limited for the specific Indian
space sector, requiring models to transfer insights from generalized datasets.

15. Forecasting and/or Nowcasting of Solar Flares Using Combined Soft

and Hard X-Ray Data from Aditya-L1

 Core Objective: Building an automated, multi-channel parsing framework using

simultaneous high/low energy X-ray spectrometers onboard the Aditya-L1 platform to
detect solar flare events and compute advance warnings.
 Pros: * Safeguards vital ground-based electronics, such as high-altitude communication
arrays and vulnerable power grids, from deep geomagnetic storms.
o Capitalizes on unique data pairs (Soft + Hard X-rays) to capture structural
changes across both rapid spikes and gradual thermal buildups.
 Cons: * Solar flares exhibit unique structural characteristics with no two events evolving
identically, making it highly difficult for generalized deep learning networks to stay
robust across varying scales.
o The system must operate with high sensitivity across massive energy dynamics to
flag faint, emerging anomalies without triggering continuous false alarms across
quiet solar intervals.