Wheat stripe rust monitoring based on weighted multi-temporal features from UAV multispectral imagery under field conditions

Le résumé fourni par la source

Wheat stripe rust is a destructive airborne disease that requires timely field-scale monitoring. However, unmanned aerial vehicle (UAV) studies often use single-date imagery or fixed bi-temporal normalised difference (DN) features, overlooking unequal contributions from observation dates during symptom development. This study proposed a feature-level adaptive weighted multi-temporal framework for wheat stripe rust monitoring from UAV multispectral imagery. Vegetation indices (VIs), texture features (TFs), and colour indices (CIs) were extracted on 24 April 2025 (booting), 10 May 2025 (heading), and 19 May 2025 (flowering). For each feature, the median absolute deviation (MAD) was used to estimate date-specific weights before weighted summation. Features were selected separately within VIs, TFs, and CIs using a three-stage procedure: Relief, minimum redundancy maximum relevance (mRMR), and sequential forward selection (SFS). Eight temporal representations were compared using XGBoost under shared stratified five-fold splits. MAD-weighted features achieved the highest mean overall accuracy (OA), Kappa, and balanced accuracy. The selected MAD-weighted VIs+TFs+CIs feature set was then evaluated using seven commonly used classifiers. XGBoost performed best, with mean OA, Kappa, and balanced accuracy of 0.831 ± 0.068, 0.649 ± 0.124, and 0.825 ± 0.052, respectively. Across feature combinations, 19 May showed the best average performance among the three single-date datasets. Out-of-fold SHAP analysis identified blue-band texture second moment, PSRI, and RG as having the largest mean absolute SHapley additive exPlanations (SHAP) values. Overall, these results indicate the potential of the framework for UAV-based stripe rust classification, while further evaluation across environments would clarify its broader applicability.

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