Precision Agriculture General Agricultural and Biological Sciences
Vegetation index-based hyperspectral assessment of sequential herbicide combinations on maize physiology and weed response under controlled maize–weed system
Abstract
Purpose Conventional visual evaluation methods often fail to detect herbicide-induced physiological responses during the initial post-application period, limiting the early assessment of herbicide efficacy for precision weed management. Therefore, this study investigated the potential of vegetation index-based hyperspectral phenotyping, combined with multivariate and machine-learning approaches, for the rapid diagnosis of the effectiveness of sequential herbicide programs in a maize-weed system. Methods A controlled pot experiment was conducted using four sequential herbicide programs comprising pyroxasulfone fb halosulfuron methyl, pyroxasulfone fb tembotrione, atrazine fb halosulfuron methyl and atrazine fb tembotrione, along with an untreated weedy control. Visible-near infrared (VNIR; 400–1000 nm) hyperspectral imagery was acquired before post emergence herbicide application and at 2, 4, 6, and 9 days after application (DAA). Multiple vegetation indices representing chlorophyll status, canopy structure, pigment composition, stress responses, and water status were derived from hyperspectral reflectance data. Principal component analysis (PCA), correlation analysis, hierarchical clustering, and supervised classification models (logistic regression, random forest, and support vector machine) were employed to evaluate treatment discrimination and early diagnostic separability. Results Tembotrione-based treatments, particularly pyroxasulfone (PE) fb tembotrione (PoE), produced the strongest physiological responses, causing pronounced reductions in chlorophyll-, red-edge-, structural-, stress-, and water-related vegetation indices within four days after application. In contrast, halosulfuron methyl-based treatments exhibited spectral trends similar to the untreated control, indicating limited impact on the existing weed flora. Among the evaluated indices, GDVI, GNDVI, NDRE, mNDVI_705, and mRESR were the most sensitive indicators of early herbicide stress. PCA revealed that the first two principal components explained more than 90% of total variability, clearly separating tembotrione-based treatments from halosulfuron methyl treatments and the weedy check. Hierarchical clustering and correlation analyses further confirmed treatment-specific physiological responses. Supervised classification models successfully classified herbicide treatment-response groups under controlled conditions, with Random Forest achieving the highest performance (ROC-AUC: 0.981 ± 0.020), demonstrating the potential utility of PCA-derived hyperspectral features for treatment-response detection. Conclusion Hyperspectral vegetation indices enabled early, objective detection of herbicide-induced stress before visible symptoms appeared. Chlorophyll- and red-edge-sensitive indices were the most responsive indicators of herbicide activity. The integration of hyperspectral phenomics with multivariate analysis and machine-learning classification provided reliable early discrimination of herbicide efficacy, highlighting its potential as a practical decision-support approach for precision weed management and future sensor-based monitoring systems.