An end-to-end remote sensing & machine learning system for tracking urban vegetation health, land-use change, and ecological disturbances using multi-spectral satellite imagery.
🚀 Quick Start · 🌍 Study Regions · 🏗️ Architecture · 📊 Results · 🤝 Contributing
- About The Project
- Key Features
- Sandbox Previews
- Tech Stack
- Architecture & Pipeline
- Study Regions
- Project Structure
- Quick Start
- Web Application
- HPC Training Workflow
- Evaluation Metrics
- Roadmap
- Contributing
- License
Urban Leaf Health Monitoring is a comprehensive, production-grade pipeline that leverages multi-spectral Sentinel-2 satellite imagery, deep learning, and an interactive Next.js web application to deliver actionable ecological intelligence for urban planning, conservation, and environmental monitoring.
The system solves a fundamental challenge in geospatial science:
How do we transform raw satellite data into a full machine-learning pipeline for pixel-level vegetation and disturbance segmentation at scale?
By integrating cloud-masked data ingestion from Google Earth Engine, a 14-channel Attention U-Net model, and a browser-based segmentation comparison tool, this project delivers a complete research-to-web workflow for monitoring forests, bushfire recovery, and urban encroachment across diverse geographic regions.
| Feature | Description |
|---|---|
| 🛰️ Multi-Spectral Ingestion | Pulls 10 Sentinel-2 bands + 4 computed indices (NDVI, EVI, SAVI, BSI) via Google Earth Engine |
| ☁️ Cloud-Masked Pipeline | Strict QA60 cloud filtering ensures high-quality, cloud-free imagery |
| 🔬 Attention U-Net | 14-channel input, 5-class dense prediction with attention gates for selective focus |
| 🧪 Domain-Aware Augmentation | Spectral jitter, random band dropout, spatial transforms — purpose-built for satellite imagery |
| ⏱️ Temporal Analysis | Multi-year and event-window comparisons (pre/during/post disturbance) |
| 🌐 Interactive Web App | Side-by-side model comparison lab built with Next.js, WebWorkers, and canvas rendering |
| 🖥️ HPC-Ready | Full PBS job scripts, environment setup, AMP training, and H100-optimized training loop |
| 📦 Dashboard-Ready Artifacts | Inference exports segmentation masks, overlays, JSON summaries, and GeoTIFFs |
Defining geographic bounding boxes for the Hasdeo study area inside Google Earth Engine.
Applying
QA60cloud and cirrus masking ensures downstream vegetation analysis is not polluted by cloud shadows.
NDVI-based degradation visualization exposing deforestation signals in the Hasdeo Forest between 2018–2023.
(Top) Cumulative forest loss chart across the Hasdeo timeline. (Bottom) Class separability matrix validating spectral feature quality.
| Technology | Purpose |
|---|---|
 Next.js 14 |
App Router, API routes, file serving |
 React 18 |
Component model, hooks, transitions |
 TypeScript 5.5 |
Full type safety across the web layer |
| 🕸️ Web Workers | Background thread pixel classification — UI stays responsive |
| 🎨 Canvas API | Client-side mask rendering and overlay compositing |
 Lucide React |
Icon system |
| Technology | Purpose |
|---|---|
 PyTorch 2.2+ |
Attention U-Net training, AMP, GradScaler |
 Earth Engine API |
Sentinel-2 data export, cloud filtering |
| 🗺️ Rasterio + GDAL | GeoTIFF I/O, spatial referencing |
| 🌐 GeoPandas + Shapely | Vector geometry, AOI manipulation |
| 🔬 Albumentations | Augmentation pipeline (spatial + spectral) |
| 🖼️ OpenCV (headless) | Image transforms, mask processing |
| 📐 NumPy + Pandas | Patch arrays, statistics, CSV processing |
| 🤖 Scikit-Learn | Baseline models (SVM, Random Forest) |
| 🗺️ geemap | Earth Engine visualization |
| 📊 Matplotlib | Plot generation and reporting |
The backend follows a clean 6-phase pipeline, each encapsulated in its own module:
┌─────────────────────────────────────────────────────────────────────┐
│ Urban Leaf ML Pipeline │
├───────────┬─────────────────┬───────────────────────────────────── │
│ Phase 1 │ Data Collection │ GEE → 10 bands + 4 indices → GeoTIFF │
│ Phase 2 │ Preprocessing │ Patch extraction (256×256, stride 128)│
│ Phase 3 │ Augmentation │ Spatial + Spectral (SpectralJitter, │
│ │ │ RandomBandDrop, flips, noise) │
│ Phase 4 │ Model │ Attention U-Net (14ch in, 5-class out)│
│ Phase 5 │ Training │ AMP, Cosine LR, Combined Dice+CE Loss │
│ Phase 6 │ Inference │ Tiled prediction → mask/overlay/JSON │
└───────────┴─────────────────┴───────────────────────────────────────┘
Input: (B, 14, 256, 256) ← 10 Sentinel-2 bands + NDVI, EVI, SAVI, BSI
Encoder: DoubleConv → Down → Down → Down → Down (bottleneck)
Decoder: Up + AttentionGate × 4
Output: (B, 5, 256, 256) ← 5 land-cover classes
5 Output Classes:
- 🌳 Vegetation
- 🌾 Sparse Vegetation
- 🟫 Bare Soil / Rock
- 🏙️ Built-up / Urban
- 💧 Water / Shadow
Loss: 50% CrossEntropy + 50% Dice Loss — balances pixel accuracy with region-overlap quality.
| Region | Focus | Time Range |
|---|---|---|
| 🌲 Hasdeo Forest, India | Mining-driven deforestation, NDVI degradation signals | 2018 – 2024 |
| 🏙️ Sydney Blue Mountains Fringe, AU | Urban expansion, ecological boundary shifts | Multi-year |
| 🔥 Kangaroo Island, AU | Black Summer bushfire impact & post-fire recovery | 2020 – 2021 |
Urban-Leaf-Health-Monitoring/
│
├── 📁 app/ # Next.js 14 Web Application
│ ├── segmentation-lab/ # Side-by-side segmentation comparison UI
│ │ ├── page.tsx # Main page with upload, sample selector, KPIs
│ │ └── segmentation.worker.ts # Web Worker — runs pixel classification off main thread
│ ├── hypothesis/ # Hypothesis/analysis pages
│ ├── recommendation/ # Recommendation pages
│ ├── api/ # Next.js API routes (asset serving)
│ ├── globals.css # Global CSS design system
│ └── layout.tsx # Root layout
│
├── 📁 h100_config/ # Core ML Pipeline (HPC-ready)
│ ├── 01_data_collection.py # GEE Sentinel-2 data collection
│ ├── 02_preprocessing.py # Patch extraction, normalization
│ ├── 03_augmentation.py # Spatial + spectral augmentation
│ ├── 04_model.py # Attention U-Net architecture + losses
│ ├── 05_train.py # Training loop (AMP, checkpointing)
│ ├── 06_predict_visualize.py # Inference, mask/overlay export
│ ├── setup_env.sh # HPC environment setup
│ ├── job_cpu.pbs # PBS CPU job script
│ ├── job_gpu.pbs # PBS GPU job script
│ └── LOCAL_TEST.py # Local pipeline test runner
│
├── 📁 scripts/ # Research Notebooks & Analysis
│ ├── 01_area_of_interest_selection/ # ROI selection notebooks
│ ├── 02_comparison_based_on_events/ # Event-window comparison (bushfire, mining)
│ ├── 03_comparison_based_on_years/ # Multi-year temporal analysis
│ ├── 04_data_agumentation/ # Augmentation experiments
│ └── 05_data_modelling/ # Model training notebooks
│
├── 📁 data/ # Earth Engine exports & datasets
│ ├── 01_area_of_interest_selection_using_sampling/
│ │ ├── batch_1/ # Hasdeo Forest dataset
│ │ ├── batch_2/ # Sydney Blue Mountains dataset
│ │ └── batch_3/ # Kangaroo Island dataset
│ └── 02_comparison_based_on_events/
│ ├── event_1/ # Hasdeo event CSVs and TIFF exports
│ └── event_2/ # Additional event data
│
├── 📁 assets/ # Visualizations & project media
│ ├── plots/ # Result plots (plot_1.png … plot_10.png)
│ ├── year_wise_comparison/ # Year-over-year forest change charts
│ ├── presentation_images/ # Key images (NDVI, cloud removal, ROI)
│ └── ppts/ # Presentation PDFs
│
├── 📁 docs/ # Technical documentation
│ ├── Urban_Leaf_Model_Pipeline_Guide.md
│ └── Urban_Leaf_Web_Project_Guide.md
│
├── package.json # Node.js dependencies
├── requirements.txt # Python dependencies
├── next.config.mjs # Next.js configuration
├── tsconfig.json # TypeScript configuration
└── README.md
- Node.js
18+— for the web application - Python
3.9+— for the ML pipeline - pip / conda — for Python dependency management
- Google Earth Engine account — for data collection
git clone https://github.com/vijaysolanki9079/Urban-Leaf-Health-Monitoring.git
cd Urban-Leaf-Health-Monitoring# Install Python dependencies
pip install -r requirements.txt
# Authenticate with Google Earth Engine
earthengine authenticate# Install Node.js dependencies
npm install
# Run the development server
npm run devOpen http://localhost:3000 to view the Segmentation Lab and interactive comparison tools.
The web application provides an interactive Segmentation Comparison Lab that:
- 📤 Accepts user-uploaded PNG / JPEG / WebP images (up to 30 MB)
- 🔄 Runs two segmentation engines simultaneously in a dedicated Web Worker (keeps the UI fully responsive)
- 🎨 Renders colour-coded segmentation masks and overlays using the Canvas API
- 📊 Displays class distribution bars and 4 derived scene indicators — Canopy Cover, Built-up Share, Exposed Surface, Water/Shadow
- 🖼️ Includes 3 preloaded sample scenes from the Hasdeo and Kangaroo Island datasets
npm run dev # Development server (hot reload)
npm run build # Production build
npm run lint # ESLint check
npm run typecheck # TypeScript type checkFor running the model training pipeline on an HPC cluster (PBS/Slurm):
cd h100_config
# 1. Set up the environment
bash setup_env.sh
# 2. Submit CPU preprocessing job
qsub job_cpu.pbs
# 3. Submit GPU training job
qsub job_gpu.pbsManual training and inference:
# Train the Attention U-Net
python h100_config/05_train.py \
--data-dir /path/to/augmented \
--model-dir /path/to/models \
--results-dir /path/to/results
# Run inference & export visualizations
python h100_config/06_predict_visualize.py \
--checkpoint /path/to/models/best_model.pth \
--input /path/to/processed \
--output-dir /path/to/results/inferenceTraining outputs:
best_model.pth— best checkpoint by validation mIoUtraining_history.json/.csv— full loss/metric historypseudo_label_distribution.json— class imbalance report*.pngoverlays,*.tifmasks,*.jsonsummaries from inference
H100 optimizations enabled:
- ✅ TF32 matmul & cuDNN
- ✅ cuDNN benchmark mode
- ✅ AMP with GradScaler
- ✅ Gradient accumulation & clipping
- ✅ Cosine annealing LR schedule
| Metric | Type | Purpose |
|---|---|---|
| mIoU | Segmentation | Mean Intersection over Union — primary quality metric |
| Dice Coefficient | Segmentation | Region-overlap quality across all classes |
| Pixel Accuracy | Classification | Fraction of correctly classified pixels |
| Precision / Recall | Classification | Per-class accuracy analysis |
| Confusion Matrix | Classification | Full inter-class error breakdown |
| Land-cover loss % | Temporal | Quantitative deforestation / recovery tracking |
- GEE-based Sentinel-2 data collection
- Cloud masking & valid-pixel filtering
- Patch extraction & radiometric normalization
- Spectral augmentation pipeline
- Full local sync of 1,000+ Hasdeo GeoTIFF files
- Spectral indices: NDVI, EVI, SAVI, BSI, NBR
- Attention U-Net implementation
- AMP training with checkpointing and resume
- Expert ground-truth masks & final benchmark metrics
- Hasdeo mining event comparison (2022 window)
- Monthly/seasonal time-series exports
- Inference masks, overlays, GeoTIFFs & JSON summaries
- Live checkpoint integration into web segmentation page
Contributions are very welcome! Whether it's selecting new study regions, improving the augmentation pipeline, or enhancing the web UI — feel free to get involved.
- Fork the repository
- Create your feature branch:
git checkout -b feature/your-feature - Commit your changes:
git commit -m 'feat: add your feature' - Push to the branch:
git push origin feature/your-feature - Open a Pull Request
This project is licensed under the MIT License — see the LICENSE file for details.
MIT License
Copyright (c) 2026 Vijay Solanki
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
Made with 🌿 by Vijay Solanki
⭐ Star this repository if you found it useful!