Improving maize LAI estimation by integrating multispectral imagery and digital surface models with ensemble learning

To overcome the limitations of single remote-sensing features in estimating maize canopy leaf area index (LAI), this study developed a UAV-based estimation approach by integrating multispectral vegetation indices (VIs) with digital surface model (DSM) features and stacking ensemble learning. Field experiments were conducted in Dehong, Yunnan Province, China, during 2023-2024, and UAV multispectral images and DSM products were acquired for maize grown under three planting-density treatments. Five vegetation indices and three DSM-derived texture/structural features were retained according to their correlation with measured LAI, statistical significance, and complementary spectral or structural information. The VI-based random forest (VI-RF) model achieved an R² of 0.835 and an NRMSE of 9.5%, whereas the DSM-based model showed lower performance (R² = 0.641; NRMSE = 14.2%). Under the same random-forest modeling framework, fusing VIs with DSM features improved the overall model performance to R² = 0.892 and NRMSE = 7.6%, indicating that DSM-derived structural information mainly enhanced the feature representation of maize LAI. Using the same VI-DSM feature set, the stacking model with support vector machine (SVM) as the meta-learner further improved the overall performance to R²= 0.930 and NRMSE = 6.3%. The additional gain from stacking was moderate but consistent, whereas feature fusion contributed the dominant improvement. The combined VI-DSM-Stacking workflow improved prediction stability across planting densities, especially under low- and high-density canopy conditions where soil background interference and spectral saturation were more evident. These results demonstrate that integrating spectral and DSM-derived structural information with stacking ensemble learning can improve the accuracy and robustness of UAV-based maize LAI estimation.