Estimating wheat lodging fraction from Sentinel-1/2 imagery using UAV-derived reference labels

Severe wheat lodging requires rapid quantitative assessment, but satellite-based estimation is often limited by scarce spatially compatible reference labels. This study developed an unmanned aerial vehicle (UAV)-derived reference-label framework for estimating wheat lodging fraction (WLF) from Sentinel-1/2 imagery in Xinxiang City, China. UAV multispectral imagery was used to classify lodged and non-lodged wheat; a multi-layer perceptron (MLP) was selected, and its binary outputs were aggregated within Sentinel-aligned 10 × 10 m grids to generate continuous WLF labels. After matching, 5,380 grids from 46 UAV source orthomosaics were partitioned by complete source groups into model-development (4,337 grids; 36 groups) and sealed independent test (1,043 grids; 10 groups) subsets. Six regressors were compared using four Sentinel configurations representing post-event optical data, post-event optical plus synthetic aperture radar (SAR) data, optical pre–post changes, and combined optical/SAR pre–post changes. Adding post-event SAR features produced limited and model-dependent benefits, whereas temporal-change features substantially improved several nonlinear models. The Feature Tokenizer Transformer (FT-Transformer) with combined Sentinel-1/2 pre–post changes achieved the best grouped cross-validation performance (coefficient of determination (R 2 ) = 0.825; normalized root mean square error (NRMSE) = 12.60%). On the spatially independent test subset, it achieved R 2 = 0.770, root mean square error (RMSE) = 0.117, and NRMSE = 12.11%. The model enabled 10 m WLF mapping across winter wheat-growing areas of Xinxiang. Together, the UAV-derived reference-label construction, source-grouped spatial validation, and controlled comparison of Sentinel configurations provide a reproducible framework for extending spatially distributed UAV observations to event-scale satellite WLF assessment.