🛰️ Deep UNet for Satellite Image Segmentation

---

📌 Overview

A Keras-based Deep UNet implementation for multi-class semantic segmentation of high-resolution satellite imagery. The model leverages spectral data from 8-band commercial satellite images to identify and classify land-use features with high accuracy.

---

📂 Dataset

The training data is sourced from the SpaceNet Dataset and consists of commercial-grade 8-band multispectral satellite imagery.

Property	Details
Source	SpaceNet (commercial-grade)
Locations	24 geographic locations
Image Type	8-channel multispectral TIFF
Image Resolution	16-bit satellite images
Mask Resolution	8-bit segmentation masks
Classes	Buildings, Roads, Trees, Crops, Water

🌈 Spectral Bands

Band	Name	Wavelength
1	Coastal	~450 nm
2	Blue	~480 nm
3	Green	~560 nm
4	Yellow	~590 nm
5	Red	~625 nm
6	Red Edge	~700 nm
7	Near-IR1	~780 nm
8	Near-IR2	~870 nm

---

⚙️ Implementation Details

• Architecture — Deep UNet with encoder-decoder structure for precise pixel-level segmentation
• Augmentation (Train) — Input images are augmented to significantly expand the effective training dataset
• Augmentation (Test) — Test-time augmentation (TTA) is applied across 7 geometric variants; mean predictions are exported to result.tif
• Hardware — Trained on an NVIDIA Tesla P100-PCIE-16GB GPU

Sample Input & Masks

---

🏗️ Network Architecture

The Deep UNet employs a symmetric encoder-decoder design with skip connections that preserve spatial detail lost during downsampling.

---

🔍 Prediction Example

---

🖥️ Interactive UI

A Streamlit dashboard lets you explore training data, visualise predictions, and inspect the model — no code required.

Dashboard — training loss curves, class weights, spectral band reference

Data Explorer — browse all 24 scenes with false-colour, single-band, and mask views; per-scene class coverage chart

---

🚀 Setup & Usage

1. Clone the repository

git clone https://github.com/devxrachit/UNET-Segmentation.git
cd UNET-Segmentation

2. Install dependencies

Requires Python 3.9+. A virtual environment is recommended.

pip install -r requirements.txt

Key dependencies and version constraints:

Package	Version	Notes
tensorflow / keras	—	Required for training & inference
tifffile	<2024	Pinned for NumPy 2.x compatibility
streamlit	>=1.32	Interactive UI
plotly	>=5.19	Charts in the UI
numpy	>=1.24	Core numerics

3. Prepare data

Place the SpaceNet multispectral images under data/:

data/
├── mband/        ← 8-band input TIFFs  (01.tif … 24.tif + test.tif)
└── gt_mband/     ← segmentation masks  (01.tif … 24.tif)

4. Train the model

python train_unet.py

Weights are saved to weights/unet_weights.hdf5. Training history is logged to log_unet.csv and TensorBoard events to tensorboard_unet/.

# Monitor training with TensorBoard (optional)
tensorboard --logdir tensorboard_unet

5. Run inference

python predict.py

Outputs:

• result.tif — raw probability maps (one channel per class)
• map.tif — colorized segmentation map

6. Launch the interactive UI

Live demo: https://devxrachit-unet-segmentation-app-9xobfy.streamlit.app/

Or run locally:

streamlit run app.py

Open http://localhost:8501 in your browser. The UI works without trained weights (demo mode using ground-truth masks).

Page	Description
Dashboard	Training loss curves, class weights, spectral band overview
Data Explorer	Browse the 24 training scenes; false-color, single-band, and mask views
Inference	Upload a TIFF or select a sample scene to run/preview segmentation
Architecture	U-Net diagram, filter progression, config table
About	Dataset details, spectral band reference, project structure

---

📁 Directory Structure

UNET-Segmentation/
├── app.py                  ← Streamlit interactive UI
├── train_unet.py           ← Training pipeline
├── predict.py              ← Inference with 7-way TTA
├── unet_model.py           ← Standard Deep UNet definition
├── unet_model_deeper.py    ← Deeper UNet variant
├── gen_patches.py          ← Patch sampling & augmentation
├── requirements.txt        ← Python dependencies
├── log_unet.csv            ← Training history (epoch; loss; val_loss)
├── .streamlit/
│   └── config.toml         ← Dark theme configuration
├── data/
│   ├── mband/              ← 8-band multispectral input images
│   └── gt_mband/           ← Ground-truth segmentation masks
├── weights/
│   └── unet_weights.hdf5   ← Saved model weights (after training)
└── tensorboard_unet/       ← TensorBoard event logs

---

📄 License

This project is open-source. See the LICENSE file for details.

---

Made with ❤️ using Keras & SpaceNet Data