This repository contains two Python notebooks aimed at training a deep learning segmentation model for road detection on satellite imagery.
File: data_acquisition.ipynb
This notebook contains code to download and preprocess satellite and cartographic data. The prepared data is used in the second notebook.
File: model_training.ipynb
This notebook includes code to prepare the input data, train the model, track experiments, visualize results, and reconstruct road geometry from the model output.
These notebooks can be run locally or on Google Colab.
If the notebooks are run locally, the output data of each notebook is written to local subfolders.
It is possible to run only the model_training notebook, using already existing data. You can download the data from the link below. Unzip the packaged file to the data/ folder in the notebook's location.
When run in Google Colab, it will export the results to a folder in Google Drive. Please make sure to update the desired path in your Google Drive beforehand.
The model_training notebook uses MLFlow and Dagshub for experiment tracking. You need to provide valid URL and credentials to an MLFlow server you have access to to run the "Training with experiment tracking" and "Explore results" sections.
It's also possible to modify the code to use a local model in the "Explore results" section to use a model available locally or from a different remote source.