Discriminating winter wheat yellow rust and Fusarium head blight using Sentinel-2 imagery at a regional scale

Le résumé fourni par la source

Yellow rust ( Puccinia striiformis f. sp. Tritici , YR) and Fusarium head blight ( Fusarium graminearum , FHB) are two major wheat diseases. These two diseases frequently pose concurrent risks to grain security, particularly in high-yielding wheat regions of eastern China. Accurate regional-scale discrimination of wheat YR and FHB is essential for developing effective green and intelligent disease management strategies. While satellite remote sensing shows potential for regional crop disease monitoring, conventional machine learning modeling approaches widely employed often fail to exploit the spectral-spatial information inherent in imagery. Meanwhile, the scarcity of ground-based disease survey samples limits the application of emerging sample-driven deep learning methods. This study evaluated the effectiveness of 27 sample-feature-algorithm combinatorial modeling strategies for discriminating regional-scale wheat YR and FHB using Sentinel-2 imagery. We augmented disease samples using a stepwise approach that combines marking diseased field vector boundaries with sliding window segmentation (SWS), horizontal-vertical flipping (HVF), and multi-angle rotation (MAR). Recursive feature elimination with cross-validation (RFECV) was employed to optimize spectral and textural features, yielding in two distinct feature sets: disease-sensitive spectral features (SFs) and spectral-textural combined features (STCFs). The original spectral bands (OSBs) served as a third feature set. These sample sets and feature sets were input into several fundamentally distinct algorithms to construct wheat YR and FHB discrimination models. These include three commonly used machine learning (ML) methods, namely, support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost). Additionally, include two deep learning methods, namely, the two-dimensional convolutional neural network (2D-CNN) and the spectral-spatial attention network (SSAN). The results indicated that three ML algorithms exhibited stable performance across all three feature sets under SWS-based sample augmentation. SVM yielded the best overall accuracy, but texture features provided only limited improvement over the SVM model compared with RF and XGBoost. The OSBs outperformed SFs and STCFs in 2D-CNN and SSAN modeling, achieving an overall accuracy (OA) comparable to that of SVM under SWS+HVF+MAR-based sample augmentation. Specifically, the SWS+HVF+MAR-OSBs-SSAN model demonstrated superior performance metrics. This model achieved an average accuracy of 81.8%, a Kappa coefficient of 0.704, a G-means of 0.892, and an F1-score of 81.1%. These accuracy results surpassed those of the SWS-STCFS-SVM model, even though the latter achieved the highest OA of 82.8%. Sample augmentation yielded limited gains in modeling for the 2D-CNN but demonstrated more significant gains for the SSAN. Overall, the STCFs-based SVM modeling strategy remains preferable under sample constraints, whereas the OSBs-based SSAN modeling strategy is more competitive with further sample augmentation. Our findings contribute valuable insights towards improved regional-scale crop biological stress discrimination.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.