Improving crop type mapping by integrating LSTM with temporal random masking and pixel-set spatial information

This is an official implementation of the "Improving crop type mapping by integrating LSTM with temporal random masking and pixel-set spatial information". The overall structure of Mask-PSTIN.

The temporal random masking technique.

The architecture of pixel-set aggregation encoder (PSAE)

Requirement

PyTorch
Numpy
tqdm

Usage

The ground truth data of Auvergne, France can be downloaded in https://geoservices.ign.fr/rpg The specific class information in this region is listed as follows:

0: Others
1: Winter wheat
2: Corn
3: Winter rye
4: Winter barley
5: Sunflower
6: Rapeseed

Data Format of Auvergne, France

data
└── <train>
    ├── data.npy # time-series satellite image patches with size of (n,t*c,h,w).
    ├── lbl.npy # ground truth 
└── <val>
    ├── data.npy
    ├── lbl.npy