ISRO - Multi-spectral Image Classification & Segmentation
Land-use/land-cover classification and segmentation on multi-spectral satellite imagery using deep learning.
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
Impact & Results
- Multi-class land-cover classification on ISRO NRSC imagery
A machine learning POC developed at ISRO NRSC applying deep learning for land-use and land-cover classification on multi-spectral satellite imagery.
The Problem
Accurate mapping of land cover types (urban, forest, water bodies, agricultural land) is essential for remote sensing analysis. This requires scalable ML models capable of processing rich multi-spectral satellite data.
The Solution
A deep learning vision pipeline that classifies and segments different land cover types directly from multi-spectral satellite imagery.
Architecture & Implementation
Modeling: Applied CNN-based classifiers and segmentation models.
Data Processing: Handled high-dimensional multi-spectral image data to extract spatial and spectral features for classification.
Data Used: Multi-spectral satellite imagery from ISRO NRSC.
Tech Stack: PyTorch, TensorFlow, Python, OpenCV, NumPy, QGIS.
Challenges & Solutions
High Dimensionality: Handling high-dimensional multi-spectral bands far beyond standard RGB images.
Class Imbalance: Balancing class distributions (e.g., abundant water vs. sparse urban areas) in the training set.
Scale: Processing massive raster files without exhausting memory.
Key Takeaways
Deep learning applied specifically to remote sensing and earth observation.
Handling severe class imbalance in segmentation tasks.
Spectral feature extraction and dimensionality reduction.