Estimation of Grassland Latent Heat Flux in Inner Mongolia from a ConvTransformer Deep Learning Model and MODIS Data

Accurately estimating latent heat flux (LE) across water-limited grassland ecosystems is critically hampered by strong land-surface heterogeneity and pronounced intra-annual variability. Here, we proposed a ConvTransformer framework by integrating MODIS remote sensing products, China Meteorological Forcing Dataset (CMFD) data, and eddy covariance observations from six grassland sites to estimate daily LE across the Inner Mongolia grasslands. The model was evaluated using a leave-one-site-out cross-validation strategy and compared with three widely used machine learning models, including random forest (RF), gradient boosting regression trees (GBRT), and support vector regression (SVR). Across the six validation sites, the ConvTransformer achieved an average R2 of 0.69, an RMSE of 12.56 W m−2, a Bias of 0.67 W m−2, and an average KGE of 0.82. Although RF produced slightly higher R2 values at several individual sites, the ConvTransformer exhibited the highest overall KGE and the most stable performance, indicating superior cross-site generalization. Based on the trained model, a 1 km daily LE dataset for the Inner Mongolia grasslands during 2003–2018 was generated. The estimated LE revealed a distinct decreasing gradient from southeast to northwest and marked seasonality, with summer dominating the annual latent heat exchange. These results suggest that the ConvTransformer constitutes an effective framework for regional LE estimation, while also offering a valuable alternative for ecohydrological studies and regional water-resource assessment in water-limited grassland ecosystems.