Estimating Leaf Area Index and Leaf Nitrogen Content for Individual Populus Tree Using UAV Hyperspectral Data With Dual-Branch Deep Learning Architecture

Unmanned aerial vehicle (UAV) hyperspectral remote sensing provides a flexible approach for quantifying leaf area index (LAI) and leaf nitrogen content (LNC) of individualPopulustree. However, current methods neglect the influence of spatial scale, feature form and model structure, resulting in unclear scale determination and limited spectral–spatial feature utilization. This study proposed a dual-branch deep learning architecture to fully exploit spatial and spectral features across multiple scales and feature forms. Hyperspectral image cubes of individual trees were generated at four scales. Three feature forms were evaluated, including all bands, selected bands, and vegetation indices (VIs). The dual-branch comprises parallel spectral and spatial branches, where spectral branch processes the averaged canopy reflectance, and the spatial branch uses the canopy image as input. The spectral branch integrates one-dimensional convolutional neural network (1-D CNN), attention mechanism (AM), and long short-term memory modules, while the spatial branch employs 2-D CNN, residual learning, and AM modules. In each branch, the optimal feature forms and spatial scales were determined and then used for develop the dual-branch models. The results showed that the optimal spatial resolutions for LAI and LNC estimation were 0.10 m and 0.05 m, respectively. Both traits achieved optimal accuracy with selected bands used in the spectral branch and VIs used in the spatial branch. The proposed dual-branch models outperformed conventional DL models, with R2of 0.861±0.006 for LAI and 0.842±0.009 for LNC. This method provides an accurate high-throughput approach for monitoring ofPopulusgermplasm, facilitating precision breeding and phenotyping applications.