This repository is the official implementation of the paper "An interpretation pipeline towards understanding multi-temporal deep learning approaches for crop mapping"

• torch
• numpy
• pandas
• scikit-learn
• jupyter
• matplotlib
• seaborn

The code has been tested in the following environment: Ubuntu 16.04.4 LTS, Python 3.5.2, PyTorch 1.2.0

The preprocessed data (.npy files) for model training and evaluation is not directly provided here due to the large data volume. You can download raw Landsat Analysis Ready Data (ARD) from EarthExplore and raw Cropland Data Layer (CDL) from CropScape, then follow the code in the preprocessing folder to generate the .npy files. Specifically, run preprocess_ARD.ipynb at first, then preprocess_CDL.ipynb, and finally category_binarization.ipynb. The raw Landsat ARD and CDL data should be stored in a new data folder that has the following structure (specific downloaded file names may change):

data
├── Site_1
│   ├── ARD
│   │   ├── 2015
│   │   │   ├── LC08_CU_018007_20150424_20181206_C01_V01_PIXELQA.tif
│   │   │   ├── LC08_CU_018007_20150424_20181206_C01_V01_SRB2.tif
│   │   │   └── . . .
│   │   ├── . . .
│   │   └── 2018
│   └── CDL
│       ├── CDL_2015_clip_20190409130240_375669680.tif
│       ├── . . .
│       └── CDL_2018_clip_20190409125506_12566268.tif
├── Site_2
└── Site_3

The preprocessed data should be stored in the preprocessing/out folder that has the following structure:

preprocessing/out
├── Site_1
│   ├── x-2015.npy
│   ├── y-2015.npy
│   ├── x-corn_soybean-2015.npy
│   ├── y-corn_soybean-2015.npy
│   ├── . . .
│   ├── x-2018.npy
│   ├── y-2018.npy
│   ├── x-corn_soybean-2018.npy
│   └── y-corn_soybean-2018.npy
├── Site_2
└── Site_3

• The PyTorch implementation of AtLSTM and Transformer models is located in the models folder. For random forest, the built-in implementation provided by scikit-learn is used.
• The utils folder contains some utilities that are used for data loading, normalization, training and evaluation.

The specific training and evaluation process can be executed by running the .ipynb files in the experiments folder.

The implementation of interpretation approaches is located in the understanding folder. Note that interpretation is based on the outputs of training and evaluation.