Spectral–Textural Feature Fusion Coupled with Machine Learning Improves UAV-Based Estimation of Maize Leaf SPAD Value in Arid Agroecosystems

Rapid and accurate monitoring of leaf nitrogen status is a key step in achieving precise fertilizer management for maize. However, UAV-based multispectral approaches relying solely on vegetation indices (VIs) may suffer from spectral saturation under dense canopies and provide limited information on canopy structural variation. This study investigated whether integrating VIs and texture features (TFs) extracted from UAV multispectral imagery could improve maize SPAD estimation under different water-nitrogen management conditions and drip irrigation configurations. Time-series UAV images were collected throughout two spring maize growing seasons in a typical arid agricultural region, from which 48 VIs and 40 TFs were derived. Pearson correlation analysis combined with best subset regression was applied to identify optimal feature combinations, and five machine learning models (PLSR, ELM, RBFNN, SVM, and BPNN) were evaluated using different input feature sets. The results showed that spectral–textural feature fusion generally improved SPAD estimation performance, with validation accuracy increasing for four of the five evaluated algorithms. RBFNN achieved the highest accuracy using VIs alone (R2 = 0.807), whereas BPNN performed best when VIs and TFs were combined (R2 = 0.833). Overall, UAV-based spectral–textural fusion coupled with machine learning provides an effective approach for monitoring maize chlorophyll status and supporting precision nitrogen management in arid agricultural regions.