Welcome to the PyTorch implementation of the change detection project conducted at the Lastig lab, IGN. This project uses advanced machine learning techniques to detect temporal changes in the landscape of Fréjus, France, from aerial photos throughout the 20th century.
Author: Cédric BARON
./evaluation_models/ stores various models that can be loaded in the script.
frejus_dataset.py is a script used to load the dataset.
ground_truth.py is an archive of the code used to produce the ground truth.
model_evaluation.py loads a model, runs various visualization on it and assesses its performances.
pre_processing.py is the script used to pre_process the raw rasters.
To compute a model and perform a general assesment, the main.py and the other scripts are used.
For the coding environment I used an environment made by my university, which entails various geo-packages. You can run this line of code in your console:
wget https://raw.githubusercontent.com/GeoScripting-WUR/InstallLinuxScript/master/user/environment.yml
conda env create -f environment.yml
To train a model with specific settings use:
python main.py --epochs 25 --lr 0.025
Note: Check main.py for a full list of command-line arguments and their descriptions.
Models are stored in ./evaluation_models/. To evaluate a specific model, enable the evaluation flag in main.py and specify the model's name.