To overcome the accuracy bottleneck in remote sensing monitoring of tree species diversity in complex terrain areas, and to meet the requirements for rapid dynamic monitoring at regional and landscape scales, precise monitoring provides robust support for regional biodiversity assessment, forest carbon stock accounting, and scientific management decision-making.

A multispectral remote sensing and light detection and ranging (LiDAR) data fusion approach was adopted for extracting tree species diversity information. This study integrated Planet Labs satellite constellation (PlanetScope) with LiDAR point cloud data, selecting a topographically complex region with rich vegetation types as the study area, and complemented by ground quadrat surveys. Initially, a linear spectral mixture model was applied to the imagery for spectral unmixing, enabling preliminary extraction of tree species spectral endmembers. Subsequently, a canopy height model (CHM) was generated from LiDAR data to quantitatively characterize the three-dimensional spatial structure of vegetation. Finally, a convolutional neural network (CNN) model was constructed to perform deep feature fusion of the aforementioned spectral and structural features, achieving refined tree species classification and regional diversity mapping.

The results demonstrated that fine-scale identification of nine dominant tree species was achieved, with the synergistic integration of spectral and structural information significantly improving the overall classification accuracy from 0.78 (using spectral decomposition alone) to 0.86. Confusion matrix analysis revealed substantially enhanced discriminative capability for spectrally similar species: classification accuracy for Schima increased from 0.67 to 1.00, eliminating misclassification with pine; maple improved from 0.67 to 1.00; bamboo rose from 0.50 to 0.75, reducing confusion with other forest types and pine; and pine increased from 0.85 to 0.92. Spatial distribution maps revealed that Schima and pine exhibited locally aggregated distributions within specific microhabitats, whereas fir and bamboo demonstrated fragmented, scattered distribution patterns. The structural information derived from the CHM proved valuable for distinguishing species with distinct vertical morphologies, effectively mitigating confusion among spectrally similar classes. However, discrimination between fir and congeneric conifers such as pine remained challenging due to their highly similar morphological characteristics. Additionally, bamboo continued to present classification difficulties attributable to its patchy, fragmented distribution and spectral mixing with surrounding vegetation. The achieved accuracy of 0.86 across nine tree species categories was comparable to that of previous CNN-based studies with fewer classes and significantly outperformed the accuracy reported for random forest algorithms (0.80) in comparable studies.

This method successfully integrates PlanetScope multispectral imagery with LiDAR point cloud data, achieving effective complementarity between spectral and structural information. It successfully overcomes the accuracy bottleneck in remote sensing monitoring of tree species diversity within complex terrain areas, providing an efficient and accurate technical solution for regional-scale forest resource monitoring and biodiversity assessment, while offering robust technical support for carbon sink assessment and ecological management.