This notebook demonstrates timeseries classification for crop identification on a subset of the MiniTimeMatch dataset by training an LSTM model.
The first two sections of this notebook are dedicated to data exploration and setting up a data preparation pipeline using a custom dataset class. The rest of the notebook deals with the set up and finetuning of the model architecture and the training loop.
Briefly about the data: Each observation in the data consists of a time sequence of 52 observations taken across 10 spectral bands of Sentinel-2. Each observation corresponds to spectral measurements aggregated over a land parcel. The observations are labelled with crop found in the parcel.
Data source: TimeMatch Dataset