A Coordinate Attention-Based Fusion Network for HSI-LiDAR Classification With Sharpness-Aware Optimization and Label Propagation
Jie Zhao, Xiaoting Jia, Junhua Ku · IEEE Access · 2026
Current HSI-LiDAR classifiers encounter three principal challenges: aligning heterogeneous spectral-structural features, mitigating instability and overfitting caused by limited labels and class imbalance, and addressing fragmented boundaries in complex scenes due to localized prediction pipelines. To overcome these issues, this work introduces an attention-enhanced fusion network (AEFN) that employs a dual-branch convolutional neural network (CNN) backbone. The architecture integrates coordinate attention (CA) for direction-aware, position-sensitive feature recalibration and adaptive gated fusion (AGF) to explicitly balance HSI and LiDAR contributions during feature interaction. To improve generalization with scarce labeled data, the training pipeline incorporates sharpness-aware minimization (SAM) with AdamW, exponential moving average (EMA) weight averaging, and MixUp regularization in the dual-input space. For comprehensive scene mapping, dense inference is performed, and posterior logits can be refined using graph-based label propagation with confidence-preserving clamping. To ensure fairness and interpretability in cross-method comparisons, AEFN-core (excluding test-time augmentation and label propagation) serves as the primary comparison setting, while optional refinements (AEFN+) are presented separately. Evaluation is conducted on two HSI-LiDAR benchmarks (Houston 2013 and Trento) and one HSI-only benchmark (Indian Pines) to distinguish multimodal fusion gains from unimodal performance.