ellaampy/CropTypeMapping — reverse-engineered prompt

Build me a Python deep learning project for crop type mapping from satellite time series.

I want to train a model on parcel level Sentinel 1 radar data, Sentinel 2 optical data, or both together, then classify fields into the 12 crop groups used in the French RPG data, like maize, wheat, barley, rapeseed, meadows, vegetables and the others. The data should be loaded from train, validation, and test folders containing numpy time series shaped like acquisitions, bands, pixels.

Please include the data preparation helper for checking the minimum usable time sequence length, training scripts for single sensor and multi sensor fusion experiments, and clear example commands so I can run the same kind of experiments for different areas. Save the results in an output folder with the usual evaluation files and any plots the code already supports.

Use PyTorch and keep the setup simple enough that I can clone it, install requirements, organize the Sentinel folders, and start training. Look up current docs online if needed.