Attention-Guided Hybrid Learning for Tree Species Classification in Mining Restoration Areas Using UAV Hyperspectral and LiDAR Data

Abstract

Accurate tree species classification in open-pit mining regions is crucial for assessing ecological restoration effectiveness. However, it remains challenging due to extreme surface reflectance, dust contamination, and spectral whitening effects prevalent. To address these challenges, this letter proposes an attention-guided misclassification-aware hybrid learning (AG-MAHL) framework for tree species classification. Unlike existing fusion methods that simply concatenate features or ignore hard samples, our framework introduces: 1) an attention-guided fusion strategy to align spectral–structural features under high-reflectance conditions and 2) a misclassification-aware learning mechanism that dynamically reweights difficult samples. Experiments conducted at the Haikou phosphogypsum tailings reservoir in Kunming, China, demonstrate that the proposed AG-MAHL framework achieves an overall accuracy (OA) of 0.8571, a weighted $F1$ -score of 0.8423, and a balanced accuracy of 0.8402. This performance surpasses that of conventional machine learning methods and representative deep learning baselines. The results indicate that the proposed attention-guided hyperspectral–LiDAR fusion, coupled with misclassification-aware learning, offers an effective solution for tree species classification in spectrally challenging open-pit mining restoration environments.