Back to all projects
Computer VisionGeospatial AIResearch / POCResearch / POC

ISRO - Redundant Region Classification for Satellite Image Compression

ML model to classify uninformative regions in satellite images, enabling intelligent data compression and prioritized transmission.

Private repository

Project Snapshot

Role

Data Engineer Research Intern

Organization

ISRO - National Remote Sensing Centre (NRSC), Hyderabad

Timeline

February – June 2025

Status

Research / POC

Deployment

POC / Research

Tech Stack

PyTorchTensorFlowOpenCVPythonNumPy

Impact & Results

  • Demonstrated significant reduction in effective data volume for transmission
  • Prototype integrated into compression pipeline

A machine learning proof-of-concept to classify uninformative regions in massive satellite images, enabling intelligent data compression and prioritized transmission for ISRO NRSC.

The Problem

Satellite images from ISRO are massive, often multiple gigabytes each. Large portions of these images (such as clouds, dark areas, or open ocean) carry no useful information. Classifying these redundant regions would enable intelligent compression of large image archives and optimize data transmission.

The Solution

An ML-based classification pipeline that identifies redundant image patches, allowing uninformative sections to be compressed aggressively without losing scientifically useful data.

Architecture & Implementation

Modeling: A CNN-based patch classifier model to evaluate and categorize image patches as informative vs. redundant.

Pipeline Integration: Integrated the classification model into a prototype image compression pipeline.

Data Used: Multi-gigabyte satellite imagery from ISRO NRSC.

Tech Stack: PyTorch, TensorFlow, OpenCV, Python, NumPy.

Challenges & Solutions

Dataset Creation: Creating a representative labeled dataset of 'redundant' vs 'informative' regions from scratch.

Inference Speed: Optimizing the CNN model for high-speed inference to keep up with the data pipeline.

Resource Constraints: Training models in an air-gapped, resource-constrained environment.

Key Takeaways

  • Applied Convolutional Neural Networks for vision tasks.

  • Handling and preprocessing massive spatial data formats.

  • Optimizing PyTorch models for throughput.

Back to all projects