slf_segmentation

Code supplement: Mapping small landscape features in agricultural lands using CNN-based semantic segmentation

Environment setup

• Install micromamba. See https://mamba.readthedocs.io/en/latest/installation/micromamba-installation.html
• Clone the repository
• Navigate to the repository folder, and create a new micromamba environment using micromamba env create --file environment.yml

Usage

Model training

• Specify input data and hyperparameters in training/train.py and run
• Use the training logs (*.csv) to plot the model behavior during training with notebook/plot_loss.ipynb

Model testing

• Run evaluation/test_inference.py on one or more large images (e.g. 5x5km2) to obtain mosaicked probability prediction raster(s)
• Calculate AUC/ROC to define optimal probability threshold and accuracy metrics with evaluation/test_accuracy.py.

Inference on mosaic

• Run inference/patches_inference to obtain probability prediction at patches level
• Batch resample each probability patch into coarser resolution virtual rasters (vrt) with inference/resample_patches.sh
• Group each vrt into processing tiles using inference/group_patches_to_tile.py.
• Run inference/mosaic_tile.py to create mosaic based on patches

Post-processing

• To apply non-arable land mask, run postprocessing/mask.sh
• To remove sieve pixels, run postprocessing/sieve_removal.py
• Convert probability mosaic raster(s)/tiles into polygon (geopackage) using postprocessing/polygonize.py
• Simplify and smoothen the vertices of each polygon features using postprocessing/smooth_polygon.py.