we do not have a full checkpoint conversion validation, if you encounter pipeline loading failure and unsidered output, please contact me via bili_sakura@zju.edu.cn

AgriFM

Model Description

This repository provides a Hugging Face-compatible checkpoint for AgriFM, a multi-source temporal remote sensing model for agricultural mapping.

The checkpoint was converted from a legacy PyTorch .pth file into standard transformers-style artifacts:

• config.json
• model.safetensors

Model Sources

• Paper: AgriFM: A Multi-source Temporal Remote Sensing Foundation Model for Agriculture Mapping
• Original codebase used for conversion: Bili-Sakura/AgriFM-transformers
• Raw source checkpoint: path/to/raw/AgriFM.pth

Conversion Notes

• The raw checkpoint uses a legacy key layout (encoder.*), which was remapped to Hugging Face PreTrainedModel state dict conventions.
• The raw checkpoint is encoder-only.
• Fusion neck and segmentation head parameters are initialized from model defaults in this converted package.

Intended Use

This checkpoint is intended for:

• feature extraction from multi-source temporal remote sensing inputs
• fine-tuning for segmentation tasks in agricultural or land-cover settings

Input Format

Expected inputs are a dict of tensors with shape (B, T, C, H, W):

• HLSL30: C=6
• S2: C=10
• Modis: C=7

Usage

import torch
from AgriFM.models import AgriFMConfig, AgriFMModel

model_dir = "path/to/AgriFM"
config = AgriFMConfig.from_pretrained(model_dir)
model = AgriFMModel.from_pretrained(model_dir, config=config)
model.eval()

B, T, H, W = 1, 32, 256, 256
pixel_values = {
    "HLSL30": torch.randn(B, T, 6, H, W),
    "S2": torch.randn(B, T, 10, H, W),
    "Modis": torch.randn(B, T, 7, H, W),
}

with torch.no_grad():
    outputs = model(pixel_values=pixel_values)
    features = outputs.features
print(features.shape)

Limitations

• This is a converted checkpoint and not an official upstream release.
• Since the source is encoder-only, task heads may require downstream fine-tuning before production use.
• Performance depends on preprocessing consistency and modality availability.

Training and Evaluation

No additional training or benchmarking was run as part of this conversion step.
For training/evaluation pipelines, refer to Bili-Sakura/AgriFM-transformers.

Dependencies

Recommended environment: environment.yml from Bili-Sakura/AgriFM-transformers
Core libraries: PyTorch, Transformers, timm, einops, scipy, h5py.

Citation

@article{li2026agrifm,
  title={AgriFM: A multi-source temporal remote sensing foundation model for Agriculture mapping},
  author={Li, Wenyuan and Liang, Shunlin and Chen, Keyan and Chen, Yongzhe and Ma, Han and Xu, Jianglei and Ma, Yichuan and Zhang, Yuxiang and Guan, Shikang and Fang, Husheng and others},
  journal={Remote Sensing of Environment},
  volume={334},
  pages={115234},
  year={2026},
  publisher={Elsevier}
}