Deep learning model training pipeline for delineating land from satellite imagery. Features preprocessing, BsiNet model implementation, and custom loss functions.
- Year
- 2025
- Status
- In development
- Stack
- Python, PyTorch, GDAL, OpenCV, TensorBoard
- Source
- Internal. Source is not public.
What I built
- Developed a deep learning pipeline for land delineation using BsiNet architecture.
- Implemented preprocessing modules using GDAL for satellite imagery (GeoTIFF) handling, including normalization, mask binarization, and resampling.
- Created a custom PyTorch dataset loader handling images, masks, contours, and distance maps.
- Defined and implemented custom loss functions including Dice Loss, Focal Loss, and a combined LossBsiNet.
- Set up a training loop with TensorBoard logging for monitoring model performance.
- Built a testing script to generate binary masks from model predictions and save them as GeoTIFFs.