On the Generalizability of Foundation Models for Crop Type Mapping

This is the official repository for the paper "On the Generalizability of Foundation Models for Crop Type Mapping", accepted at IEEE IGARSS 2025.

Authors: Yi-Chia Chang, Adam J. Stewart, Favyen Bastani, Piper Wolters, Shreya Kannan, George R. Huber, Jingtong Wang, Arindam Banerjee.

Installation

1. Clone the repository:

   git clone https://github.com/yichiac/crop-type-transfer-learning.git
   cd crop-type-transfer-learning
   

2. Create a virtual environment and activate it:

   python3 -m venv env
   source env/bin/activate
   

3. Install the required packages:

   pip install -r requirements.txt
   

Usage

Download Dataset

The dataset is available on the HuggingFace. Login to download the crop type datasets.

huggingface-cli login

You can run the following Python script to download the dataset:

from huggingface_hub import snapshot_download
snapshot_download(repo_id="torchgeo/harmonized_global_crops", repo_type="dataset", local_dir="harmonized_global_crops")

Sentinel-2 Download Pipeline

Sentinel-2 donwload pipeline can be found in Ai2 rslearn_projects.

Data Split

Follow the files in train-test-splits/ to split the dataset into training and testing sets. For demo, you can run the following command to create train/test data for a CDL OOD experiment:

python3 train-test-splits/create_splits.py

Training

Run the training script for CDL with SSL4EO-S12 pre-trained weights:

python3 -m torchgeo fit --config experiments/fewshot/cdl_100_ood_ssl4eo.yaml \
data.dict_kwargs.sentinel2_paths='./harmonized_global_crops/sentinel2_subsample_cdl_2023_100_ood' \
--seed_everything 0

Testing

After completing training, you can use the checkpoints to test on the data.

python3 -m torchgeo test --config experiments/fewshot/cdl_100_ood_ssl4eo.yaml \
data.dict_kwargs.sentinel2_paths='./harmonized_global_crops/sentinel2_subsample_test' \
--seed_everything 0 \
--ckpt_path=...

Reference

To cite our IGARSS publication, use the following citation:

@INPROCEEDINGS{chang2025croptypetransfer,
    author={Chang, Yi-Chia and Stewart, Adam J. and Bastani, Favyen and Wolters, Piper and Kannan, Shreya and Huber, George R. and Wang, Jingtong and Banerjee, Arindam},
    booktitle={IGARSS 2025 - 2025 IEEE International Geoscience and Remote Sensing Symposium},
    title={On the Generalizability of Foundation Models for Crop Type Mapping},
    year={2025},
    pages={948-953},
    doi={10.1109/IGARSS55030.2025.11242260}
}