Field-Scale Estimation of Maize Leaf Area Index from UAV Multispectral Imagery Using Compact Spectral and Texture Predictors

Abstract

Leaf area index (LAI) is an important indicator of canopy development and is widely used for crop diagnosis and field-scale monitoring. We developed a UAV multispectral approach to estimate and map maize LAI in Tumed Right Banner, Inner Mongolia, China. The dataset comprised 488 paired observations from 122 georeferenced locations in 13 fixed plots, measured repeatedly at four growth stages. Vegetation indices and gray-level co-occurrence matrix texture features were derived from the green, red, red-edge, and near-infrared bands. Five models were evaluated using a single 70:30 split. Model robustness was further examined through 100 repeated stage-stratified splits, leave-one-plot-out validation, leave-one-growth-stage-out validation, and RF-based feature ablation. In the single-split benchmark, RF achieved R2 = 0.889, RMSE = 0.249, and RRMSE = 0.082. Across the repeated splits, the corresponding values were R2 = 0.876 ± 0.016, RMSE = 0.258 ± 0.018, and RRMSE = 0.085 ± 0.006. Leave-one-plot-out validation gave R2 = 0.878 ± 0.062, whereas leave-one-growth-stage-out testing revealed weaker, stage-dependent transferability (RF R2, −0.542 to 0.441). Feature ablation showed that vegetation indices provided most of the predictive information; adding the two NIR texture variables increased R2 by 0.010 relative to vegetation indices alone. Because the field platform lacked independent plot replication, the maps are interpreted as descriptions of canopy heterogeneity rather than evidence of treatment effects. The approach is suitable for site-specific maize canopy monitoring, but broader application requires independent validation across years and locations.