Dataset

Due to size constraints, the dataset and checkpoints are hosted externally.

1. Download the Multi-Temporal Crop Classification dataset and the corresponding GeoMTL text annotations.
2. Extract both into the same directory. It should contain chips_df.csv, training_data.txt, validation_data.txt, test_data.txt, training_chips/, validation_chips/, and the per-chip .txt / .qual.txt / .clean.txt caption files.
3. Pretrained checkpoints to reproduce paper results: GeoMTL checkpoints on Hugging Face.

Installation

git clone https://github.com/TadieB/geoMTL.git
cd geoMTL
conda env create -f environment.yml
conda activate prithvi_py311

Usage

All entry points use a subcommand for the model, followed by that model's arguments. Run python scripts/train.py <model> --help for the full argument list.

Train

python scripts/train.py geomtl --decoder_mode scratch \
    --data_path /path/to/data --model_weights /path/to/Prithvi_EO_V2_300M.pt

python scripts/train.py resnet_lstm --data_path /path/to/data

python scripts/train.py seg_ablation --finetune_full_model \
    --data_path /path/to/data --model_weights /path/to/Prithvi_EO_V2_300M.pt

Evaluate (quantitative metrics → JSON)

python scripts/evaluate.py geomtl --decoder_mode scratch \
    --checkpoint checkpoints/geomtl_scratch_best.pth \
    --model_weights /path/to/Prithvi_EO_V2_300M.pt --data_path /path/to/data \
    --output_file metrics_geomtl_scratch.json

Inference (qualitative panels + text reports)

python scripts/inference.py geomtl --decoder_mode scratch \
    --checkpoint checkpoints/geomtl_scratch_best.pth \
    --model_weights /path/to/Prithvi_EO_V2_300M.pt --data_path /path/to/data \
    --output_dir sample_outputs/geomtl_scratch

SLURM equivalents (see slurm_jobs/*.sh headers for more examples):

sbatch -J geomtl-scratch slurm_jobs/run_train.sh geomtl --decoder_mode scratch \
    --data_path /path/to/data --model_weights /path/to/Prithvi_EO_V2_300M.pt

Citation

@inproceedings{medimem2026geomtl,
  title     = {GEO-MTL: A Multi-task Framework for Joint Semantic Segmentation and Textual Scene Reporting in Remote Sensing},
  author    = {Medimem, Tadie B. and Melgani, Farid and Fiore, Sandro Luigi and Anantharaj, Valentine G.},
  booktitle = {IEEE International Geoscience and Remote Sensing Symposium (IGARSS)},
  year      = {2026}
}