High-Resolution GeoAI framework for automated rutting depth prediction on logging trails using TransRAUnet and LiDAR

Understanding rutting depth resulting from machine traffic during thinning operations is a key indicator for evaluating operational quality and informing future planning in boreal forests. Existing rutting depth prediction methods often depend heavily on field inspections, which are labor-intensive, time-consuming, and limited in both accuracy and spatial scalability. This study presents a GeoAI-based approach that integrates a DINO Vision Transformer (ViT) with UNet-based architectures using airborne LiDAR-derived covariates to address these limitations. For this purpose, we first segmented logging trails (LTs) using a residual attention UNet (RAUnet) with LiDAR inputs. We then collected over 1,300 rutting depth samples from approximately 230 LT segments across 36 forest sites in southwest Finland. To densify the sparse annotation dataset, rutting depths at unsampled locations within each LT segment were interpolated using the inverse distance weighting (IDW) method. From LiDAR data, we derived micromorphologic and hydrologic covariates, which were then used to develop pipelines combining a hybrid ViT encoder and four UNet-based decoder architectures for rutting depth prediction. The results showed that RAUnet achieved high performance in segmenting LTs, with a Cohen’s kappa of 0.91. Although all 13 covariates contributed to the initial modeling process, the most accurate rutting depth predictions were obtained using the top six predictors, dominated by multiscale roughness, ruggedness, and depth-to-water (DTW). Among the tested architectures for rutting depth prediction, the hybrid ViT–RAUnet (TransRAUnet) achieved the highest accuracy, yielding an R2 of 0.84 and the lowest error metrics. In contrast, TransAUnet and TransSwinUnet showed reduced performance, with R2 values lower than that of the standard TransUnet. Our proposed approach provides a practical and scalable solution for automating rutting depth prediction, optimizing trafficability maps, and enabling data-driven post-harvest assessments in boreal forests.