🛰️ Satellite Water Segmentation

🧠 Overview

This project applies deep learning models (like U-Net and its variants) to perform semantic segmentation on satellite images. The goal is to accurately detect water regions from multi-band input images (e.g., 12-band satellite imagery) and deploy the solution using a Flask web interface.

🚀 Features

• U-Net-based architecture with various backbones (ResNet34, ResNet50, EfficientNetV2B0).
• Accepts 12-channel multispectral input data.
• Pixel-wise segmentation output with water masks.
• Flask-powered web app for uploading images and displaying results.
• Evaluation metrics: Accuracy, IoU, Precision, Recall.
• Download predicted mask as an image.

📊 Dataset Summary

The dataset includes preprocessed satellite images with 12 spectral bands and corresponding binary masks indicating water bodies.

Band Explanation

The multi-band structure enhances water segmentation accuracy.

📷 Below is a sample visualization of the spectral bands:

Band	Name	Min	Max	Use
1	Coastal aerosal	-1393.0	6568.0	Aerosol detection
2	Blue	-1169.0	9659.0	Useful in cloud and snow discrimination
3	Green	-722.0	11368.0	General-purpose true-color rendering
4	Red	-684.0	12041.0	Detecting plant stress, soil exposure
5	NIR	-412.0	15841.0	Differentiating between water and land
6	SWIR1	-335.0	15252.0	Soil moisture, drought stress detection, Burned area detection, Differentiating snow/cloud
7	SWIR2	-258.0	14647.0	Geological mapping, identifying bare soils, rocks, Detecting water content and surface changes
8	QA Band	64.0	255.0	Masking clouds, shadows, saturation, or other invalid pixels
9	Merit DEM	-9999.0	4245.0	Identifying basins, slopes, floodplains
10	Copernicus DEM	8.0	4287.0	Similar to MERIT DEM, but higher resolution
11	ESA world cover map	10.0	100.0	Provides land use/land cover context (e.g., water, forest)
12	Water occurence probability	0.0	111.0	Water body persistence analysis

🛠 Installation & Setup

git clone https://github.com/bassantsherif123/ComputerVision_Satellite_Image_Water_Segmentation.git

Download the model using the following link, then create a Model folder inside the project and add it to it: 📥 Download Model File

You will need to download Flask, if you don't already have it:

pip install flask

🌐 Flask App

• Upload multispectral image (with 12 channels)
• Visualize band combinations
• View the predicted segmentation mask
• Download the bands the output mask

To run the app, use:

python Deployment/app.py

Then open on any browser using:

http://127.0.0.1:5000/

or

http://localhost:5000

📈 Model Evaluation

Model Name	Accuracy	IoU
Custom UNet	0.955315	0.826221
EfficientNetV2B0	0.901890	0.408989
ResNet34	0.958784	0.836132
ResNet50	0.857304	0.408989