Summary of Problem Statement Explainer Session -2 | Bharatiya Antariksh Hackathon 2025

Summary of "Problem Statement Explainer Session -2 | Bharatiya Antariksh Hackathon 2025"

This video is the second explainer session for the Bharatiya Antariksh Hackathon 2025, aimed exclusively at student participants. The session focuses on detailing problem statements 8 through 14 out of a total of 14 challenges presented in the hackathon. The goal is to provide clarity on each problem, the objectives, datasets, expected outcomes, methodologies, and evaluation criteria to help participants develop innovative solutions.

General Hackathon Overview

Bharatiya Antariksh Hackathon 2025 is a student-only event organized under ISRO’s vision.
There are 14 problem statements in total; the first 7 were covered in the previous session.
Participants must form teams of 3 to 4 members.
Submission deadline is 9th July 2025.
Shortlisted teams will be announced around 24th July, with a 30-hour offline finale on 7th-8th August at NRSC Hyderabad.
Benefits include networking with ISRO scientists, internship opportunities, and national recognition.
Mentorship and discussions will primarily happen on a Discord server, organized by problem statements.
Participants must submit ideas using the provided idea submission template on the hackathon platform.

Detailed Problem Statements Covered

1. Problem Statement 8: Novel Approaches for Optimizing Deep Learning in Earth Observation with Imbalanced Data

Objective: Design new optimizers, loss functions, or model architectures to handle class imbalance in satellite image segmentation.
Context: Satellite images often have dominant classes (e.g., vegetation) overshadowing rare classes (e.g., small water bodies), causing poor detection of rare classes.
Key Concepts:
- Class imbalance in remote sensing datasets.
- Deep learning optimizers (first order like Adam, gradient descent; second order like Hessian-based).
- Challenges like exploding and vanishing gradients.
Datasets: Satellite images with imbalanced class distributions (e.g., one class 88%, others much less).
Expected Outcome: Novel optimizer/loss function improving rare class detection.
Evaluation: Innovation, mathematical soundness, gradient behavior, robustness across datasets.
Tools: Open source tools recommended.

2. Problem Statement 9: AI/ML Algorithm for Identifying Tropical Cloud Clusters (TCC) Using INSAT-3D Satellite Data

Objective: Develop an AI/ML algorithm to detect and track tropical cloud clusters using infrared brightness temperature data.
Importance: TCCs contribute to rainfall and cyclone formation; tracking them aids weather prediction and cyclone modeling.
Methodology:
- Use thresholding on infrared brightness temperature (<218K for North Indian Ocean, <221K for South Indian Ocean).
- Identify cluster size, intensity, position, and persistence.
- Separate multiple clusters based on spatial criteria.
Data: INSAT-3D infrared brightness temperature data, available via MOSDAC.
Tools: MATLAB, Python, or any open-source language.
Expected Outcome: Algorithm capable of real-time TCC detection and tracking.
Evaluation: Algorithm’s accuracy in detecting clusters and physical parameter estimation.

3. Problem Statement 10: Identifying Heliospheric CME Events Using Particle Data from SWIFT Instrument on Aditya-L1

Objective: Develop an algorithm to detect Coronal Mass Ejection (CME) events from solar wind plasma particle data.
Context: CME events impact space weather and geomagnetic storms; currently, no plasma-parameter-only detection algorithm exists.
Data: Time series data of plasma parameters (proton velocity, temperature, etc.) from ISSDC.
Expected Outcome: Algorithm detecting CME events, validated against known catalogs.
Tools: Any open-source tools.
Evaluation: Accuracy in event detection compared to catalogs.

4. Problem Statement 11: Novel Method to Detect Landslides and Boulders on the Moon Using Chandrayaan Images

Objective: Detect and differentiate lunar landslides and boulders using images from Chandrayaan missions.
Challenges: Small size of objects, faint image quality, and distinguishing landslides (on slopes) from boulders (elevated).
Data: Radiometrically corrected images from Chandrayaan 1, 2, and 3 available on ISSDC.
Expected Outcome: Automated detection and localization of lunar landslides and boulders.
Evaluation: Accuracy in detection (true positives), ability to identify landslide origins.

5. Problem Statement 12: Dual Image Super Resolution for High-Resolution Optical Satellite Imagery and Blind Evaluation

Objective:
- Part 1: Generate a super-resolved high-resolution image from two low-resolution images