A UAV-Based Highway Slope Inspection Method Based on Multimodal Sensors and Decision-Level Fusion Strategy

Abstract

Highway slope maintenance is an important task in transportation infrastructure. In recent years, unmanned aerial vehicle (UAV)-based remote sensing has emerged as a flexible, low-cost approach to slope inspection. However, heterogeneous multimodal data and limited image labels pose major challenges to high-performance UAV image analysis. In this study, we propose a multimodal UAV-based slope inspection framework with Vision Foundation Models and decision-level fusion. First, a UAV inspection is conducted to gather RGB, TIR, and LiDAR data from the investigated area. The Cloth Simulation Filter and grid-based elevation encoding generate a color-encoded digital elevation model (DEM). Second, a frozen DINOv2-ViT-S/14 encoder extracts modality-specific features from the aligned patches. Three independent Extreme Gradient Boosting (XGBoost) classifiers generate five-class probability distributions. Third, decision-level fusion preserves the modality-specific predictions and integrates their class support, prediction confidence, and cross-modal differences through L2-regularized logistic regression. The patch label corresponds to the class with the highest fused probability. The proposed framework was evaluated using 2007 patch triplets from five highway slopes in Shandong, China. Under mixed-site stratified five-fold validation, it achieved 90.48% accuracy, 85.13% macro-averaged recall, and an F1-score of 82.90%. Spatial aggregation converted the classification results into high-priority inspection regions for engineering review and field verification.