Dr. Wenyuan Li’s Recent Publication on Remote Sensing of Environment

AgriFM: A multi-source temporal remote sensing foundation model for Agriculture mapping

Congratulations to Dr. Wenyuan Li, Professor Shunlin Liang, and the research team for their recent publication in Remote Sensing of Environment (Vol. 334, 115234). The paper, titled "AgriFM: A multi-source temporal remote sensing foundation model for Agriculture mapping," is available via https://doi.org/10.1016/j.rse.2026.115234. The codes and models are freely accessible at https://glass.hku.hk/casual/AgriFM/.

Precise agricultural mapping is crucial for global food security, yet it faces challenges in modeling multi-scale spatiotemporal patterns from fine field textures to full growing-season dynamics. Existing deep learning methods and general-purpose Remote Sensing Foundation Models (RSFMs) often process spatial and temporal features separately or utilize fixed-window strategies (like standard ViT) that compromise fine-grained details. To address these limitations, the team proposed AgriFM, a multi-source temporal foundation model specifically designed for agriculture. The model introduces a synchronized spatiotemporal downsampling strategy within a modified Video Swin Transformer backbone, enabling efficient handling of variable-length satellite time series while preserving multi-scale phenological information. AgriFM was pre-trained on a globally representative dataset of over 25 million images from MODIS, Landsat-8/9, and Sentinel-2, utilizing land cover fractions as supervision to robustly learn geographic priors.

AgriFM was rigorously evaluated across five diverse agricultural mapping tasks, including agricultural land mapping, field boundary delineation, and specific crop mapping (paddy rice and winter wheat). The results demonstrate that AgriFM consistently outperforms existing state-of-the-art models (including Prithvi, SatMAE, Galileo, and SMARTIES). Specifically, in the challenging agricultural land use/land cover mapping task, AgriFM achieved an F1 score of 60.49%, significantly surpassing other foundation models (which ranged from 40% to 47%). For paddy rice and winter wheat mapping, it achieved F1 scores of 86.97% and 75.85%, respectively. Compared to conventional methods, AgriFM not only delivers higher accuracy but also improves computational efficiency, reducing training time by approximately 40–45% through its optimized downsampling strategy. The model demonstrates exceptional capability in preserving fine field boundaries and capturing long-term temporal patterns critical for crop identification.

Figure 1: The overall framework of AgriFM, illustrating the pre-training phase with multi-source satellite data and the unified mapping framework with a versatile decoder.

Figure 2: Visual comparison of agricultural land use/land cover mapping results. AgriFM (second from right) demonstrates superior performance in identifying diverse crop types and delineating clear field boundaries compared to other baseline models.

References:

Li, W., Liang, S., Chen, K., Chen, Y., Ma, H., Xu, J., Ma, Y., Zhang, Y., Guan, S., Fang, H., & Shi, Z. (2026). AgriFM: A multi-source temporal remote sensing foundation model for Agriculture mapping. Remote Sensing of Environment, 334, 115234. https://doi.org/10.1016/j.rse.2026.115234

Keywords:

Foundation model, Agriculture mapping, Remote sensing, Deep learning