Engineers for Exploration - Mangrove Monitoring

Machine learning platform for mangrove ecosystem monitoring from aerial and satellite imagery.

Developed by: Mangrove Monitoring Team, Engineers for Exploration (E4E), UC San Diego

---

Overview

This repository provides tools for:

• Pixel-level mangrove detection from drone imagery (~100 km² labeled data)
• Infrastructure detection (roads, buildings) threatening mangrove ecosystems
• ArcGIS Pro integration for environmental scientists
• Experimental architectures including state-space models (Mamba)

---

Quick Start

# Clone
git clone https://github.com/UCSD-E4E/ml-mangrove.git
cd ml-mangrove

# Setup environment
conda create -n mangrove python=3.11 -y
conda activate mangrove

# Install PyTorch (with CUDA if available)
# Follow https://pytorch.org/ to install torch for your OS
# Example for Windows: pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu126

pip install -r requirements.txt

See Environment Setup for detailed instructions.

---

Project Structure

ml-mangrove/
├── DroneClassification/    # Main ML pipeline
│   ├── data/               # Data processing & loading
│   ├── models/             # Model architectures & losses
│   ├── training_utils/     # Training framework
│   ├── experiments/        # Training outputs & logs
│   └── testing/            # Experimental work (Mamba integration)
├── ARC_Package/            # ArcGIS Pro toolbox
├── docs/                   # Documentation
└── archive/                # Historical approaches (reference only)

---

Models

Model	Use Case	Performance
ResNet18-UNet	Mangrove detection	82-85% IoU
SegFormer B0/B2	Infrastructure detection	Fast / Accurate
DeepLabv3+	Multi-class segmentation	81-84% mIoU
MambaUNet	Experimental (state-space)	25× smaller model

---

Documentation

Guide	Description
Environment Setup	Python, CUDA, WSL2 configuration
ArcGIS Setup	Toolbox installation for ArcGIS Pro
Data Preparation	Processing geospatial imagery
Model Training	Training segmentation models
Inference	Running predictions
Model Architectures	Architecture details
Loss Functions	Available loss functions

Full documentation: docs/index.md

---

Pipeline

1. Data Processing: Tile GeoTIFFs, create train/val/test splits
2. Model Training: Train with custom losses (Jaccard, Boundary IoU)
3. Deployment: Package for ArcGIS Pro or run inference directly

---

Key Files

File	Purpose
DroneClassification/data/utils.py	Geospatial utilities
DroneClassification/models/models.py	Model architectures
DroneClassification/models/loss.py	Loss functions
DroneClassification/training_utils/training_utils.py	Training framework
ARC_Package/SegmentationToolbox.pyt	ArcGIS toolbox

---

Contributing

1. Create a feature branch from master
2. Follow existing code style
3. Add tests for new functionality
4. Submit PR with clear description