Inversion of kiwifruit LAI based on fusion of UAV multispectral vegetation indices and texture features

Accurate monitoring of kiwifruit growth parameters is often required in precision agriculture. Among them, Leaf Area Index (LAI) was one of the most important indicators for canopy structure and photosynthesis. Conventional LAI measurement is destructive and inefficient, while unmanned aerial vehicle (UAV) multispectral remote sensing, though effective, can often yield insufficient accuracy when relying solely on Vegetation Indices (VIs), due to kiwifruit's complex canopy characterized by leaf overlap and vine mixture. Single-feature approaches cannot fully meet the requirements of a dynamic orchard. This study aims to achieve high-precision LAI inversion for kiwifruit. The VIs were integrated with Texture Features (TFs) using machine learning models. A series of tests was conducted in a 'Xuxiang' kiwifruit orchard in Shaanxi Province, China. Image data was collected during four growth stages (final flowering, young fruit, fruit expansion, and maturity) in 2025. UAV multispectral imagery was acquired using a DJI Mavic 3, with green, red, red-edge, and near-infrared bands. Ground-truth LAI was measured with an IN-G30 analyzer to calibrate against direct destructive samples. A total of 12 VIs were calculated (e.g., Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index 2 (EVI2)) and 8 TFs (e.g., mean, variance). TFs were derived from a Gray-Level Co-occurrence Matrix (GLCM) with a 9×9 window. Feature selection employed a Random Forest (RF) importance algorithm. The features over 300 iterations were ranked to identify an optimal subset. Three models, RF, Decision Tree (DT), and Support Vector Regression (SVR), were constructed using stepwise feature addition. Coefficient of determination (R²) and root mean square error (RMSE) were evaluated after optimization. Feature selection was performed using the RF importance algorithm. The results showed that the accuracy reached its peak, including 11 features (7 VIs and 4 TFs). The VIs, such as EVI2, NDRE, and mean TFs, were the most important contributors. Dynamic evaluation confirmed that the performance failed to improve beyond feature 11, indicating the high simplicity. In single-feature input, the VIs consistently outperformed TFs over all models. Taking the fruit expansion stage as an example, the RF model with VI input achieved an R² of 0.787 and RMSE of 0.263 on the test set, while the RF model with TF input obtained an R² of 0.462 and RMSE of 0.293. The VIs and TFs were fused to significantly improve inversion accuracy at all growth stages. The RF model with fused features performed the best, especially in the fruit expansion stage (test set R²=0.913, RMSE=0.204). In the whole growth period, the RF model maintained the best performance (average test set R²=0.859, RMSE=0.282), thus outperforming the DT model (R²=0.821, RMSE=0.325) and the SVR model (R²=0.853, RMSE=0.283). Model adaptability varied with canopy density: The RF model performed better in stages with complex canopies (fruit expansion and maturity), while the SVR model was more suitable for sparse canopy stages (final flowering and young fruit). The fruit expansion stage shared the highest inversion accuracy, due to its stable canopy structure, while the final flowering stage shared the lowest accuracy (e.g., RF R²=0.823). UAV-based VIs and TFs were integrated with machine learning, especially the RF model. Accurate and efficient LAI inversion was realized for kiwifruit at all growth stages. Feature fusion was used to reduce single-source data using physiological and structural information, particularly for complex canopies. The RF was recommended for robust applications. While SVR can offer an alternative for sparse canopy periods. This approach can also provide a reliable tool for precision agriculture, thereby monitoring kiwifruit growth in a sustainable orchard.