Individual Tree Species Classification in Mining Areas Fusing UAV LiDAR and Hyperspectral Data via a CNN-RepMLP Framework

Abstract

Intensive coal mining in the Yellow River Basin has severely degraded native vegetation, making accurate individual tree species identification a prerequisite for ecological restoration monitoring. In this study, we proposed a deep learning framework based on a 1D-Convolutional Neural Network (CNN)-reparameterization multilayer perceptron (RepMLP) ensemble for tree species classification in complex mining landscapes by fusing uncrewed aerial vehicle-borne Light Detection and Ranging and hyperspectral data. We extracted structural metrics (height statistics, intensity percentiles, and Canopy Height Model) and spectral features from tree samples. A hierarchical feature ablation experiment was conducted via sixfold cross-validation. The results indicated that the optimal 70-feature set achieved 94.53% with a stable cross-validation mean of 96.18%, effectively mitigating the overfitting risks typically associated with small-sample deep learning. Utilizing this optimal 1-D feature set, the proposed CNN-RepMLP was evaluated against seven machine learning classifiers. It outperformed all baseline models, achieving an overall accuracy of 94.53%, a Kappa coefficient of 0.90, and a weighted F1-score of 0.95, followed by LightGBM (93.31%), while maintaining a structurally efficient inference process. Statistical significance tests further corroborated the superiority of the CNN-RepMLP architecture. Our findings demonstrate that: first, the integration of structural and spectral features substantially enhances classification performance; and second, the CNN-RepMLP model provides a robust, accurate, and generalizable solution for individual tree species identification. This framework offers a effective tool for high-precision vegetation monitoring and ecological recovery assessment in degraded regions.