Multiple U-Net encoders mapping different resolution images into the same feature space for semantic segmentation of the Forest Tree Line in the Alps. The data used is from Swisstopo.
Original code from: https://github.com/thienanhng/ForestMapping/blob/main/launch.sh
HR = High Resolution encoder LR = Low Resolution encoder
Model (HR and LR encoders mapping to the same decoder):
Confusion matrices (HR top, LR bottom):
Confusion matrices (HR top, LR middle, labels bottom):
The labels in black are No Forest, dark gray Close Forest, light gray Open Forest, beighe Shrub Forest and white woodland(neglected for the type of forest segmentation).
Results (HR top, LR bottom):
conda env create -f myenv.yml
or
conda create -n myenv python=3.10.0
packages: gdal rasterio pandas torch torchvision tqdm numpy psutil opencv-python conda install -n myenv <package_name> conda activate myenv
Train is the training process to train the two encoders and decoder. Infer is the validation process to evaluate the performance of the trained model.
Check in infer.py and train.py the arguments for more help
in order 1946, 2017, 2017 labels:
input_0,input_1,target /media/data/charrez/SwissIMAGE/1946/gray_DOP100_LV95_2570_1099_1946.tif, /media/data/charrez/SwissImage/2017_25cm/DOP25_LV95_2570_1099_2017_1.tif, /media/data/charrez/TLMRaster/5c/TLM5c_2570_1099.tif