A CNN-Transformer Hybrid Framework for Mapping Annual Wheat Fractional Cover From 2001-2023 Using MODIS Satellite Data Over Asia

Wheat is a staple crop in over 40 countries, with Asia accounting for more than 40% of global cultivation. Long-term mapping of wheat cover is critical for agricultural management and food security. However, existing wheat mapping products face a key limitation in spatiotemporal coverage: they either offer broad spatial coverage for a single or few years, or provide long time series that are confined to specific regions. To address this gap, we propose DeepMapping, a hybrid deep learning framework designed to generate a consistent annual fractional wheat cover product at 250 m resolution from 2001 to 2023 over Asia. Our framework integrates Convolutional Neural Networks (CNNs) and Transformer models to extract complementary spatial and temporal features. It processes multi-resolution data, including 250 m and 500 m MODIS reflectance, alongside GLASS Leaf Area Index and Fractional Vegetation Cover products. The model is trained with fractions derived from the 10 m resolution WorldCereal 2021 map and refined with ancillary information such as coarse-resolution crop products, land cover data, and agricultural statistics to enhance reliability. By learning the relationship between short-term, high-resolution labels and long-term, coarse-resolution MODIS observations, DeepMapping is efficiently used to generates annual wheat fractional cover product over Asia at 250 m from 2001–2023. Wheat is a staple crop in over 40 countries, with Asia contributing more than 40% of global wheat cultivation. Long-term wheat cover is critical for agricultural management and food security, but existing products face a trade-off: they either provide high spatial resolution with short temporal coverage or offer long time series with limited spatial coverage. We propose DeepMapping, a hybrid deep learning framework, to estimate long-term annual wheat fractional cover at 250 m from MODIS data. DeepMapping integrates Convolutional Neural Networks (CNNs) and Transformer models to extract spatial and temporal features from MODIS data at 250 m and 500 m resolutions and Global LAnd Surface Satellite (GLASS) Leaf Area Index and Fractional Vegetation Coverage products. The DeepMapping is trained using WorldCereal 2021 at 10 m and refined with other ancillary information, such as coarse-resolution crop products, land cover data, and statistics to ensure reliability. By learning the relationship between high-resolution labels and long-term coarse-resolution MODIS observations, DeepMapping is efficiently used to generates annual wheat fractional cover product over Asia at 250 m from 2001–2023. Validation with 2,556 samples demonstrates overall accuracy of 84.1%, producer's accuracy of 87.8%, and user's accuracy of 70.07%. Comparison with national statistical data confirms high consistency ($R^{2}$: 0.880–0.943). DeepMapping offers a scalable solution for long-term, reliable agricultural mapping. This product is publicly available athttps://glass.hku.hk/casual/Asia_Wheat/.