The repository contains a code that allows training models based on convolutional neural networks for segmenting river areas in satellite images composed of RGB visible bands.
The author's implementation of the vgg_unet model scored IoU=0.90174. Below is a sample data (columns: input, model output, model output, respectively).
- PyTorch - ML framework
- OpenCV - a library for image processing
- NumPy - a library for matrix operations
- neptune - logging tool
Dataset available for download from a separate repository: https://github.com/shocik/sentinel-river-segmentation-dataset
Running the code on your own computer requires the following preparatory steps:
- Neptune configuration in file config.cfg.
- Modify the path to the working directory in the file train.ipynb:
#set workdir os.chdir("/content/drive/MyDrive/RiverSemanticSegmentation/")
- Modifying the path to a dataset in a file train.ipynb:
#dataset configuration dataset_dir = os.path.normpath("/content/drive/MyDrive/SemanticSegmentationV2/dataset/")