Satellite Imagery Binary Classifier: Forest Detection (NDVI)

This project implements a comprehensive pipeline to classify satellite imagery into Forest and Non-Forest categories. By combining cloud-based data acquisition with local Deep Learning, the system monitors environmental changes using Sentinel-2 NDVI data.

📌 Project Overview

The workflow bridges the gap between massive satellite data collection and localized AI model training. We use a Convolutional Neural Network (CNN) to analyze $32 \times 32$ pixel patches, generating high-resolution forest maps to compare land cover between 2020 and 2024.

🛠️ Tech Stack

• Data Acquisition: Google Earth Engine (JavaScript API)
• Deep Learning: PyTorch (CNN Architecture, Autograd)
• Geospatial Data: Rasterio (GeoTIFF management), Zarr (Efficient chunked storage)
• Analysis: NumPy, Matplotlib

---

📂 Hybrid Workflow Structure

Phase 1: Data Acquisition (JavaScript - GEE)

Before the Python pipeline begins, the raw data is prepared using the Google Earth Engine Code Editor:

• Collection Filtering: Sentinel-2 L2A collections are filtered for specific dates (Dry Season) and low cloud cover.
• NDVI Calculation: Normalized Difference Vegetation Index is computed across the entire region.
• Export: Cleaned GeoTIFF composites for 2020 and 2024 are exported to Google Drive for local processing.

Phase 2: Data Extraction & Labeling (Python)

• Automated Labeling: Forest centroids are identified using an NDVI threshold ($\ge 0.6$).
• Sampling: A balanced dataset of 1,000 patches (500 per class) is randomly sampled.
• Zarr Storage: Patches are streamed into a .zarr file, allowing the training loop to access data without overloading system RAM.

Phase 3: CNN Training

A custom SimpleCNN is trained to recognize spatial patterns:

• Layers: Dual-layer convolution with MaxPooling for feature extraction.
• Training Logic: 20 epochs using the Adam optimizer and BCELoss.
• Hardware: Optimized for CUDA/GPU with a CPU fallback.

Phase 4: High-Resolution Inference

The model is deployed on the full-scale GeoTIFFs:

• Sliding Window: Applies the model across the entire image.
• Stride = 1: By moving the window pixel-by-pixel and averaging overlaps, the system creates a smooth, sub-pixel probability map that avoids "blocky" artifacts.

---

🚀 Execution Guide

1. Requirements

pip install torch rasterio zarr numpy matplotlib

2. Steps

1. GEE (JS): Export your NDVI GeoTIFFs from Google Earth Engine.
2. Process: Run the extraction notebook to generate dataset_ndvi.zarr.
3. Train: Run the CNN training loop. The model weights will save to forest_classification_model.pth.
4. Infer: Apply the model to the 2020 and 2024 images using the apply_fine_model function.

🛡️ Data Integrity & Safety

• NaN Handling: Uses np.nan_to_num to handle "NoData" pixels often found in satellite exports.
• Normalization: Accounts for contrast variations across different satellite orbits/dates.
• Shape Consistency: Uses .view(-1) to ensure the 1D label tensor matches the model output, preventing batch-dimension crashes.

---

Developed for environmental monitoring and deforestation analysis.