Construction of a novel index driven by time-series hyperspectral features and the PROSAIL model for early warning of walnut leaf scorch.

Walnut leaf scorch syndrome is a critical factor hindering the high-quality development of the walnut industry. Its incubation period is characterized by a lack of identifiable external morphological features, resulting in significant delays in early warning through traditional monitoring methods. This study seeks to establish a proactive early identification system for scorch leaf syndrome by integrating temporal spectral features with physical mechanism models. The goal is to provide theoretical support and technical assurance for precise forecasting, as well as proactive prevention and control measures in orchards. The study focused on the 'Wen 185' walnut plot in Aksu Prefecture and employed hyperspectral dynamic monitoring throughout the entire growing season. By applying the PROSAIL radiative transport model to derive physiological indicators of leaves, a novel physically driven index was developed. A comprehensive comparison was conducted on the performance of genetic algorithm, simulated annealing algorithm, and random forest in managing time-series redundant bands and identifying key features. Ultimately, models utilizing random forest, support vector machine, and artificial neural network were established for the early detection of focal leaf syndrome. The findings indicate that, despite the absence of visible lesions on walnut leaves during the latent period of scorch leaf syndrome, the hyperspectral curve has already exhibited a distinct "blue shift" trend at the red edge position. At the feature selection level, genetic algorithms, due to their robust global search capabilities, significantly outperform other algorithms in eliminating temporal redundancy and preserving the essential features of disease progression. Further analysis utilizing SHAP interpretability confirmed that the health index and photochemical reflection index are pivotal in the model's decision-making process. Crucially, the optimized classification framework achieved robust performance, with the random forest model combined with simulated annealing feature selection attaining the highest overall accuracy of 0.793 and a Kappa coefficient of 0.597. Furthermore, models using traditional and newly constructed vegetation indices yielded strong early warning capabilities, achieving overall accuracies up to 0.782 and 0.751 (with Kappa coefficients of 0.564 and 0.556, respectively) in distinguishing healthy, latent, and early symptomatic stages. This study illustrates that the comprehensive integration of temporal logic and physical driving features can markedly enhance the sensitivity and robustness of early warning systems for leaf blight, facilitating a transition from "symptom-driven" to "mechanism-driven" disease monitoring. This advancement offers substantial practical guidance for the intelligent and digital management of the walnut industry.