This repository contains various collections of work that make up the development and deployment process of my capstone project: building footprint extraction from satellite imagery using semantic segmentation. The project explores the use of deep learning to extract meaningful building footprint data from freely available, Sentinel 2 satellite imagery, in an attempt to contribute to urban planning, humanitarian efforts, and the nonprofit sector.
To see the finalized application in action, visit the deployed version here: Building Footprint Extraction App.
The deployment repository is hosted at: Deployment Repo.
This project was designed with accessibility and cost-efficiency in mind. The semantic segmentation model was initially built through altering the SpaceNet 7 Challenge dataset to fit the spatial resolution necessary for this use case.
The system leverages a UNet-based deep learning model to automatically extract building footprints from satellite images, and the deployment is hosted using no-cost cloud services. While the current deployment meets the project's goals, the architecture is designed to be scalable for future demands, such as higher resolution imagery or additional features.
This repository is organized into several key directories:
- Contains documents and notebooks submitted for each phase of the capstone project, covering the entire workflow from data collection to model deployment.
- Training: Contains notebooks used to train the UNet model, including preprocessing, hyperparameter tuning, and model architecture exploration.
- Evaluation: Includes notebooks for evaluating model performance, such as Dice Coefficient, Jaccard Index, and more, on validation and test datasets.
- Contains a notebook that demonstrates how to use the trained model to make predictions on new satellite imagery. This includes image pre-processing, patching, and stitching operations for efficient model inference.
- Contains the trained model files in various formats (e.g.,
.keras,.h5), including the final UNet model used in the deployed app.
- Contains scripts and notebooks used to clean, preprocess, and format the original SpaceNet 7 Challenge dataset for training the segmentation model.
- Original Dataset: The original, unprocessed dataset can be found through the SpaceNet 7 Challenge.
- Preprocessed Dataset: My cleaned and pre-processed version of the dataset, which was used for model training, can be accessed here: Google Drive Link.
For more specific details on implementation, refer to the documentation in the deployment repository: Deployment Repository.