From plots to region: Machine learning-based UAV-satellite integration for mapping fractional coverage of peanut southern blight

Peanut southern blight (PSB), a soil-borne fungal disease, threatens peanut production. Because PSB originates at the root collar, canopy symptoms develop gradually and subtly, limiting satellite monitoring. UAV observations can capture these symptoms but are spatially restricted. To address this, we developed a machine learning-based UAV-satellite integration framework using fractional coverage (FCover) as a severity and scale-bridging metric for regional-scale PSB monitoring. First, Sentinel-2 composites from Google Earth Engine were used to extract temporal features and construct NDVI time series, from which a novel Autumn Crop Index was derived and applied within the Dynamic World crop mask to delineate autumn crop extent. Within this extent, stable temporal features were identified via fully nested 10-fold cross-validation (CV), combining Random Forest (RF) importance with redundancy filtering (Separability Index and Pearson correlation). An RF classifier trained on these features with a grid-optimized hyperparameter set, chosen based on validation performance across all folds, achieved a mean overall accuracy of 0.98 and was subsequently applied for peanut mapping. Next, UAV-derived spectral, texture, and structural features were selected via ReliefF, correlation analysis, and recursive elimination to train RF classifiers. The resulting classification maps were aggregated to Sentinel pixels to generate weighted FCover samples. Finally, stable, temporally matched Sentinel-2-derived spectral and texture features were identified via fully nested 5-fold CV using weighted correlation. These features were then used to model FCover with weighted Bayesian-optimized RF, Extra Trees, Gradient Boosting, and eXtreme Gradient Boosting (XGB). XGB achieved the highest CV performance (overall R 2 = 0.718, RMSE = 0.123) and, using its optimal hyperparameters, was trained with FCover weights to generate a regional FCover map strongly agreeing with field-surveyed incidence ( R 2 = 0.89). This study demonstrates, for the first time, that UAV-satellite integration enables effective PSB monitoring, providing a scalable approach for precision disease management in peanut-producing regions. • First regional-scale monitoring of peanut southern blight (PSB) severity using a machine learning-based UAV-satellite integration framework. • Introduced fractional coverage (FCover) as a sub-pixel disease severity metric to bridge UAV and satellite observations. • Refined the peanut mapping area using an NDVI-based Autumn Crop Index (ACI). • Achieved high cross-validation performance for FCover prediction using eXtreme Gradient Boosting (XGB) (overall R 2 = 0.718, RMSE = 0.123). • Generated a regional FCover map strongly agreeing with field-surveyed PSB incidence ( R 2 = 0.89).