Using Monoscopic Multispectral Earth Observation Images to Predict Terrain Features With Deep Neural Networks - Supplementary material

This is supplementary material to the journal paper, "Using Monoscopic Multispectral Earth Observation Images to Predict Terrain Features With Deep Neural Networks", published in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS) 25 August 2025. --------------------------- Contents: Source_code.zip: Contains the Python source code for training the R2U-Net for predicting terrain slopes in optical Sentinel-2 L2A images, with AW3D30 DSM as targets. The Python files for downloading and processing the raw images have already been used to produce the provided merged L2A and DSM images in Data.zip ready for using as input to the R2U-Net for training and evaluation.requirements.txt contains the Python environment libraries needed to train and run the source codes.The YAML files contain the folder paths to the merged images used as input to the network. Data.zip: Contains the merged L2A and DSM images used to train and evaluate the deep learning models. It also contains supplementary material related to the Sentinel-2 image data used; like the Sentinel-2 image file names and ids for downloading from the Copernicus Browser, for both training and testing with the deep learning model, and image location maps.NB! Full unzipped Data-folder is 57.1GB. Results.zip: Contains some of the results achieved with the trained R2U-Net models. For each of the two trained models, tensorflow event files from training and evaluation are provided to be visualized with tensorboard. A final folder contains images used in the article for this supplementary material, with corresponding names to match the article figures.Provided are also the best model checkpoints from training the multispectral and RGB models, used to evaluate the trained models and produce the images and results in Results.zip.