Semantic Segmentation with UNet

Overview

This project involves developing a model to segment changes in two remote sensing images of the same geographic area. The goal is to input two images of identical resolution and output a multi-class mask that identifies different types of changes. These changes are classified into several categories: buildings, roads, vegetation, water, land, and areas that remain unlabeled.

Classes

The model will classify changes into the following categories:

• Building: Changes involving buildings.
• Land: Changes related to the land surface without specific structures.
• Road: Changes related to road structures.
• Vegetation: Changes in areas covered by vegetation.
• Water: Changes involving bodies of water.
• Unlabeled: Areas where no specific changes can be classified.

Setup

Prerequisites

This project uses Poetry for package management to avoid dependency conflicts and maintain a clean working environment.

Installation

If you don't have Poetry installed on your system, follow the installation instructions from Poetry's official documentation.

To install the project dependencies, navigate to the project directory and run:

poetry install

Training the Model

To start training the model, execute the following command:

python -m src.train --root_directory "datasets/aerial_image_segmentation" --save_path "models/segmentation_model.h5"

For tracking the training process run

tensorboard --logdir=logs

Running segmentation inference on a single image

python -m src.inference_segment --model_path "models/unet_100_epochs.h5" --image_path "example_images/original_image.jpg"

Running change detection inference

python -m src.change_detection --model_path "models/unet_100_epochs.h5" --original_image "example_images/original_image.jpg" --changed_image "example_images/changed_image.jpg"

Ps. obviously segmentation model's accuracy is not great, since more time is needed for hyperparameter tuning for the model. However, in general the structure is something like this.