AgriFM × PASTIS

Reimplementation of AgriFM — A Multi-source Temporal Remote Sensing Foundation Model for Agriculture Mapping — adapted for the PASTIS dataset on AMD GPU (ROCm 7.0).

Results (Fold 1, trained from scratch, 32 minutes)

Per-Class IoU

Model

• Backbone: Video Swin Transformer (synchronized spatiotemporal downsampling)
• Neck: MultiFusionNeck (U-Net style skip connections)
• Head: CropFCNHead
• Parameters: 39.6M (small config)
• Input: Sentinel-2, 32 temporal frames, 128×128 pixels, 10 bands
• Framework: Pure PyTorch, AMD ROCm 7.0 compatible

Training

python train.py \
    --data_root /path/to/PASTIS \
    --fold 1 \
    --small_model \
    --epochs 100 \
    --batch_size 16 \
    --lr 5e-5 \
    --weight_decay 0.05

Citation

@article{li2025agrifm,
  title={AgriFM: A Multi-source Temporal Remote Sensing Foundation Model for Agriculture Mapping},
  author={Li, Wenyuan and others},
  journal={Remote Sensing of Environment},
  year={2025}
}