Repository for the paper: "CropClimateX: A large-scale Multitask, Multisensory Dataset for Crop Monitoring under Climate Extremes" by Adrian Höhl, Stella Ofori-Ampofo, Miguel-Ángel Fernández-Torres, Ridvan Salih Kuzu, and Xiao Xiang Zhu. More information about the dataset can be found on Hugging Face.
Dataset can be download from Hugging Face:
-
website (not recommended for large-scale download):
https://huggingface.co/datasets/torchgeo/CropClimateX -
via git:
git clone https://huggingface.co/datasets/torchgeo/CropClimateX -
or download with a script from Hugging Face API:
python src/datasets/download.pypython src/datasets/download.py -m <modalities> -ids <geoids> --local_dir <download_folder> -nr_workers <number parallel downloads>
Each data source is located in a folder, each folder contains the minicube .zarr files, one .zarr file contains up to 10 minicubes. Each minicubes has an id like this: <GEOID>_<PID>, where the GEOID is the id of the county and PID is the id of the minicube.
The xarray library is recommend to read the files: xr.open_zarr(filename, group=minicube_id). The data is saved as integers to save space, xarray applies the scaling and offset automatically to the data while loading it. Also, all meta data like coordinates and time are loaded correctly.
You can preprocess and change the format of the files into .npy/.zarr by using the script in datasets/dataset_format.py. All the resampling/selection will be taken care of by reusing the dataloader and saving it to new files.
The location of the minicubes were optimzed for each county. Two alogrithms were used a Genetic Algorithm and a Sliding Grid Algorithm, these were compared to a straightforward gridding of the county (baseline). They can be run using the following scripts (see also install environment):
python src/dataset_preparation/minicube_creation/main_gp.py
python src/dataset_preparation/minicube_creation/main_sga_baseline.py
Installing the environment
Run this to install the dependencies (you may need a different pytorch version for your hardware etc.):
python -m venv .venv/venv1
source .venv/venv1/bin/activate
pip install -r requirements.txt
or install with uv and exact python version:
pip install uv # install package manager
uv venv .venv/venv1 --python 3.10.13
source .venv/venv1/bin/activate
uv pip install -r requirements.txt
This repository is based on Pytorch Lightning and Hydra. The exact parameters for the runs can be found in configs/experiments/. All models can be downloaded and evaluated, the script and weights need to be adjusted accordingly. All numbers and figures in the paper can be reproduced by using the weights and configs. For training and testing run:
python src/main.py experiments=<config> train=True test=True paths.data_dir=<use this parameter if the data_dir is not "data/"> trainer.devices=<gpus to use>
For testing with saved weights run:
python src/main.py train=False test=True ckpt_path=<path> experiment=<experiments_file_name>
# or multirun with multiple weights:
python src/main.py -m train=False test=True ckpt_path=<path>,<path>,<path> experiment=<experiments_file_name>
For using different seeds run:
python src/main.py train=True test=True seed=<seed1> experiment=<experiments_file_name>
# or multirun with multiple seeds:
python src/main.py -m train=True test=True seed=<seed1>,<seed2>,<seed3> experiment=<experiments_file_name>
download data:
python src/datasets/download.py -m cdl_500m modis daymet dem soil yield --local_dir data/CropClimateX
preprocess/resample data:
python src/datasets/dataset_format.py --ds_name=prep_yield_modis_corn_float32 --data=yield_modis_data_prep --bands sur_refl_b01 sur_refl_b02 sur_refl_b03 sur_refl_b04 sur_refl_b06 tmax tmin prcp vp srad elevation slope aspect curvature bulk_density cec clay ph sand silt soc --pred_bands yield --bands_channel_dim=-3 --dtype=float32
example for training:
python src/main.py train=True test=True seed=42 experiment=final_yield_cnn_lstm
example for testing with weights:
python src/main.py train=False test=True experiment=final_yield_cnn_lstm ckpt_path=weights/yield/cnn_lstm_seed\=42.ckpt
create 3 fold cross-validation figure in the paper:
python src/create_plots.py experiment=yield_spatial_cv ++ckpt_paths=['weights/yield/3dcnn_cv_data.split.k=0.ckpt', 'weights/yield/3dcnn_cv_data.split.k=1.ckpt', 'weights/yield/3dcnn_cv_data.split.k=2.ckpt'] ++data.split.k=[0,1,2]
download data:
python src/datasets/download.py -m landsat8 dem udsm --local_dir data/CropClimateX
preprocess/resample data:
python src/datasets/dataset_format.py --ds_name=prep_drought_landsat_uint16 --data=drought_landsat_data_prep --bands B02 B03 B04 B05 B06 elevation slope aspect curvature --pred_bands usdm_mode --bands_channel_dim=-3 --dtype=uint16
example for training:
python src/main.py experiment=final_drought_cnn_lstm test=True seed=42
example for testing with weights:
python src/main.py train=False test=True experiment=final_drought_cnn_lstm ckpt_path=weights/drought/cnn_lstm_seed\=42.ckpt
- clone/fork the repo
- implement your model in src/models/
- create/modify a config file configs/models (& configs/data)
- train the model with
python src/main.py ...
- hyperparameter tuning:
python src/main.py --multirun logger=wandb model.optimizer.lr=0.1,00.1 tags=[tune_lr] - hyperparameter tuning with optuna:
python src/main.py -m hparams_search=<config> - cross validation:
python src/main.py --multirun data.split.nr_folds=6 data.split.k=0,1,2,3,4,5 tags=[<tag>] - for more see: https://hydra.cc/docs/intro/
If you use this work or the dataset please consider citing: TBA
The code is licensed under the MIT license.
Code Template: https://github.com/ashleve/lightning-hydra-template