Husheng Fang’s recent publication on Remote Sensing of Environment

Generating an annual 30 m rice cover product for monsoon Asia (2018–2023) using harmonized Landsat and Sentinel-2 data and the NASA-IBM geospatial foundation model

Congratulations to Husheng Fang, Professor Shunlin Liang, and the research team for their recent publication in Remote Sensing of Environment (Vol. 335, 115256). The paper, titled "Generating an annual 30 m rice cover product for monsoon Asia (2018–2023) using harmonized Landsat and Sentinel-2 data and the NASA-IBM geospatial foundation model," is available via https://doi.org/10.1016/j.rse.2026.115256.

Rice cultivation in Monsoon Asia is vital for global food security, accounting for over 90% of the world's production. However, accurate monitoring in this region is challenging due to the prevalence of smallholder farming systems and highly fragmented farmlands, which are difficult to map using traditional coarse-resolution satellite imagery. Additionally, the lack of sufficient labeled training data has hindered the application of advanced deep learning models. To overcome these limitations, the team proposed a novel framework that adapts the NASA-IBM geospatial foundation model (Prithvi) for continental-scale agricultural mapping. The study utilized an automatic label generation strategy derived from multiple existing regional products and fine-tuned the Prithvi model using fine-resolution Harmonized Landsat and Sentinel-2 (HLS) data.  A key innovation of the study is proposing a new paradigm illustrates how foundation models can be operationalized for continental-scale agricultural tasks. The fine-tuned Prithvi achieves acceptable accuracy under scenarios with limited training samples. It demonstrates higher precision than UNet and reaches convergence more quickly.

The reliability of the new product was rigorously tested using 35,458 independent validation samples collected through field surveys and visual interpretation. The validation results indicate an overall accuracy (OA) of 84.14% for the entire Monsoon Asia region, with consistent performance across different years (2018–2023). Spatially, the accuracy varies across climatic zones, ranging from 83.07% to 90.06%. Compared to existing mainstream datasets, such as the 500 m MODIS-based product and 20 m SAR-based maps, this new 30 m product demonstrates superior performance. It effectively addresses the mixed-pixel issues inherent in coarse-resolution data and significantly reduces the speckle noise often found in SAR-based products, providing a clearer and more detailed representation of rice distribution.

Figure 1: The workflow for using the geospatial foundation model (Prithvi) in rice mapping.

Figure 2: The architecture of task-specific decoder and added CNN branch.

Figure 3: Spatial distribution of paddy rice in Monsoon Asia (2018–2023).

Table 1: Model performance with different training sample sizes, compared with U-Net.

Table 2: Accuracy assessment.

References:

Fang, H., Liang, S., Li, W., Chen, Y., Ma, H., Xu, J., Ma, Y., He, T., Tian, F., Zhang, F., & Liang, H. (2026). Generating an annual 30 m rice cover product for monsoon Asia (2018–2023) using harmonized Landsat and Sentinel-2 data and the NASA-IBM geospatial foundation model. Remote Sensing of Environment, 335, 115256.

Keywords:

Rice mapping, Geospatial foundation model, Harmonized Landsat and Sentinel-2, Remote sensing, Monsoon Asia