UAV Multi-Source Remote-Sensing Feature Fusion and ES-EfficientNet for Accurate Grading of Maize Curvularia Leaf Spot

Maize Curvularia leaf spot (CLS) severely affects maize yield and quality; the implementation of non-destructive, precise disease grading is crucial for containing the spread of the disease in the field. This study utilised hyperspectral and thermal infrared remote sensing data acquired by unmanned aerial vehicles (UAVs) to extract spectral, textural (TFs) and canopy temperature features. To address the limitations of individual features, the textural-derived indices (CLSTIs) and Normalized Relative Canopy Temperature (NRCT) were developed. Mutual Information-based Recursive Feature Elimination with Cross-Validation (MI_RFECV), Variable Selection Using Random Forests (VSURF), and Variable Combination Population Analysis (VCPA) algorithms were employed to screen for sensitive spectral bands, and these selected bands were fused with the two indices to construct a multidimensional feature set. An ES-EfficientNet model was developed, featuring an MBConv architecture as its backbone with embedded serial multi-scale ECA modules. By utilising depthwise separable convolutions to extract multi-scale spectral features of the disease and employing multi-scale ECA to adaptively optimise channel weights, this model effectively suppresses interference caused by noise and redundant features. The results indicate that the feature combination based on MI_RFECV+CLSTIs+NRCT achieved the optimal classification performance in the ES-EfficientNet model. To ensure the robustness of the results, the model was evaluated across three random seeds (1, 42, and 2024). It maintained a stable performance, achieving an overall accuracy (OA) ranging from 94.17% to 95.33% and a Kappa coefficient up to 94.15%. Its identification performance was superior to that of baseline models using single-source features, and the model maintained stable performance in cross-site transfer within the same provincial-scale study area under the same-year experimental conditions. In summary, this study achieved high-precision, non-destructive grading of maize CLS, providing practical value for field disease monitoring and precise control, while also offering a robust reference framework for remote sensing-based disease identification in other crops.