Quantamaster/Earth-Observation

• 📌 Overview
• 🛰️ Project Context
  • Data Sources
• 🧠 System Architecture
• 📁 Directory Structure
• 📦 Dataset
• 🔍 Pipeline Breakdown
  • Phase 1: Spatial Reasoning & Filtering
  • Phase 2: Label Construction
  • Phase 3: Dataset Cleaning & Split
  • Phase 4: CNN Training & Evaluation
• 📊 Results & Metrics
• 🧠 Model Details
• ⚙️ Technical Requirements
• 🚀 Usage
• 🧾 Data Format
• 🛠️ Troubleshooting
• ✅ Best Practices
• 📚 Citation

This repository presents a reproducible Earth Observation (EO) and Deep Learning pipeline for automated land-cover classification using Sentinel-2 RGB image patches and ESA WorldCover 2021 data.

The system enables:

• Automated geospatial data filtering
• Raster-based ground-truth label generation
• Supervised CNN training using ResNet-18
• Quantitative land-use and environmental analysis

The project is framed as an AI-based Geospatial Audit of the Delhi Airshed, suitable for research labs, environmental agencies, and policy analytics.

This work demonstrates how satellite imagery + deep learning can be leveraged for urban land-use monitoring and environmental auditing.

Data Sources

• Sentinel-2 (10m resolution RGB imagery)
• ESA WorldCover 2021 land-cover raster
• Delhi-NCR administrative boundary (EPSG:4326)

┌──────────────────────────────┐
│   Delhi-NCR AOI (GeoJSON)    │
└──────────────┬───────────────┘
↓
┌──────────────────────────────┐
│   Spatial Grid (60×60 km)    │
└──────────────┬───────────────┘
↓
┌──────────────────────────────┐
│ Sentinel-2 RGB Image Chips   │
│  (128×128 @ 10m resolution)  │
└──────────────┬───────────────┘
↓
┌──────────────────────────────┐
│ ESA WorldCover Raster (10m)  │
│  → Mode-based Labeling       │
└──────────────┬───────────────┘
↓
┌──────────────────────────────┐
│ Clean Labeled Dataset (CSV)  │
└──────────────┬───────────────┘
↓
┌──────────────────────────────┐
│ ResNet-18 CNN (PyTorch)      │
└──────────────┬───────────────┘
↓
┌──────────────────────────────┐
│ Metrics: Accuracy, F1, CM    │
└──────────────────────────────┘

/Earth_Observation_Pipeline/
│
├── data/
│   ├── delhi_ncr_region.geojson
│   ├── delhi_ncr_grid.geojson
│   ├── worldcover_bbox_delhi_ncr_2021.tif
│   ├── rgb/                          # Sentinel-2 image patches (128×128)
│   ├── image_coords.csv
│   ├── imgs_within_grid.csv
│   └── labelled_images_clean.csv
│
├── 01_grid_visualization.py
├── 02_label_extract_assignment.py
├── 03_train_test_split.py
├── 04_cnn_train_eval.py
│
├── requirements.txt
└── README.md

Kaggle Dataset (Required): https://www.kaggle.com/datasets/rishabhsnip/earth-observation-delhi-airshed

Includes:

• Sentinel-2 RGB image patches
• Delhi-NCR shapefiles
• Image coordinate metadata

• Define AOI using Delhi-NCR boundary
• Generate a uniform 60 × 60 km grid
• Filter Sentinel-2 images by grid intersection

Script: 01_grid_visualization.py

• Extract raster patches from ESA WorldCover
• Assign land-cover class using mode-based sampling
• Handle missing data and edge effects

Script: 02_label_extract_assignment.py

• Remove invalid labels
• Perform stratified train-test split
• Analyze class distribution

Script: 03_train_test_split.py

• CNN: ResNet-18
• Input: Sentinel-2 RGB chips
• Metrics: Accuracy, F1 Score, Confusion Matrix

Script: 04_cnn_train_eval.py

(Representative – depends on training run)

• Overall Accuracy: ~75–85%
• Macro F1 Score: ~0.72–0.82
• Strong Performance: Urban, Vegetation, Water
• Challenges: Mixed land-cover and boundary regions

• Confusion Matrix
• Class-wise F1 Scores
• Correct vs Incorrect Prediction Visualizations

• Architecture: ResNet-18
• Framework: PyTorch
• Loss Function: Cross-Entropy
• Evaluation Metrics: Accuracy, F1 Score

• Python 3.8+

pip install geopandas numpy pandas rasterio shapely scipy \
            matplotlib seaborn scikit-learn \
            torch torchvision torchmetrics \
            geemap

python 01_grid_visualization.py

python 02_label_extract_assignment.py

python 03_train_test_split.py

python 04_cnn_train_eval.py

| filename | lat | lon | label | class_str |

• Ensure all data paths are correct
• Run scripts in order
• Verify image presence in rgb/

• Preserve intermediate outputs
• Validate class balance before training
• Track experiment configurations

Please cite the following if used in research or applications:

• ESA WorldCover 2021
• Copernicus Sentinel-2
• Relevant geospatial data providers

⭐ If this repository helps your work, consider starring it!