Deep Representation Learning & Computer Vision for Remote Sensing

Academic Affiliation: Course Projects for GNR 638: Machine Learning for Remote Sensing - II, Centre of Studies in Resources Engineering (CSRE), IIT Bombay
Collaborators: Dheeraj Kumar Maradana (@dheerajkumar2005) & Lohit (@lohit2320)

---

📌 Executive Summary

This repository contains CNN implementation in C++ , transfer learning experiments, and representation probing benchmarks

Due to unique spectral characteristics, spatial resolutions, and domain shifts in aerial imagery, standard ImageNet-pretrained representations often exhibit distinct transferability dynamics. This project systematically investigates feature representations across deep convolutional neural networks (e.g., ResNet-50) using linear probing, layer-wise probing, few-shot adaptation, and out-of-distribution robustness assessments.

---

🏗️ Repository Structure & Modules

Remote-Sensing-Deep-Learning-GNR638/
├── A1/                           # Baseline Classification & Feature Extraction Pipeline
│   ├── src/                      # Model architectures, data loaders, and training utilities
│   ├── train.py                  # End-to-end training harness for satellite image classification
│   ├── test.py                   # Model evaluation, confusion matrices, and precision/recall
│   └── weights/                  # Checkpoint storage
└── A2/                           # Representation Learning & Transferability Probing
    ├── Linear_probe_transfer/    # Linear probing of frozen feature backbones (ResNet-50)
    ├── Layer-wise_Feature_Probing/# Diagnosing representational power across intermediate layers
    ├── Fine_Tuning_Strategies/   # Full fine-tuning vs. backbone-frozen vs. gradual unfreezing
    ├── few_shot/                 # Few-shot learning performance under data-constrained regimes
    └── robust/                   # Robustness evaluation against synthetic corruptions & weather shifts

---

🔬 Core Methodologies & Experiments

0. CNN Implementation in C++ (A1/src/)

• Complete CNN implementation in C++ using only std library

1. Layer-Wise Feature Probing (A2/Layer-wise_Feature_Probing)

• Trains linear classifiers on intermediate feature representations extracted from each residual stage of deep vision backbones.
• Evaluates where geospatial and spectral domain features emerge (early edge detectors vs. late high-level semantic representations).

2. Linear Probing & Transfer Dynamics (A2/Linear_probe_transfer)

• Measures the out-of-the-box representational transferability of self-supervised and supervised pretrained representations to multi-class remote sensing benchmarks.

3. Fine-Tuning Strategies (A2/Fine_Tuning_Strategies)

• Compares convergence rates, compute requirements, and downstream accuracy across:
  • Linear probe only (classifier head updated, backbone completely frozen).
  • Partial unfreezing (last $k$ residual blocks fine-tuned with smaller learning rate).
  • Full fine-tuning (discriminative layer-wise learning rates).

4. Few-Shot Adaptation (A2/few_shot)

• Analyzes sample complexity and classification accuracy under $N$-way $K$-shot conditions ($K \in {1, 5, 10, 20}$) to simulate scarce labeled satellite data scenarios.

5. Out-of-Distribution Robustness (A2/robust)

• Evaluates model sensitivity and calibration under atmospheric perturbations, sensor noise, brightness variations, and contrast degradation.

---

🚀 Getting Started

Environment Setup

# Clone the repository
git clone https://github.com/dheerajkumar2005/Remote-Sensing-Deep-Learning-GNR638.git
cd Remote-Sensing-Deep-Learning-GNR638

# Install dependencies
pip install torch torchvision torchaudio scikit-learn matplotlib seaborn tqdm

Running Experiments

# Train baseline classifier in A1
cd A1
python train.py --epochs 30 --batch_size 64

# Run linear probe transfer evaluation in A2
cd ../A2/Linear_probe_transfer
python linear_probe.py