Automated terrain roughness assessment using remotely sensed data

In forestry, assessing and navigating terrain is crucial for operational efficiency. Terrain trafficability, the capacity to support vehicle passage, affects machinery movement and soil health. While soil's weight-bearing ability is crucial, terrain classification is equally important for defining an area's characteristics. This study automates terrain roughness estimation using UAV imagery to generate high-resolution roughness maps. We used deep learning for object segmentation, followed by point cloud classification and ground roughness quantification. Manual field measurements validated the segmentation results while manipulating the obstacle dataset tested the roughness algorithm's sensitivity. Additionally, drone-derived Digital Surface Models (DSM) were used to calculate the Terrain Roughness Index (TRI), Vector Roughness Measure (VRM), and Area Ratio (AR) for comparison. Obstacle segmentation achieved 95.6% accuracy, while height estimation had an RMSE of 2.62 cm and an MRE of 11.4%. Manipulating the dataset demonstrated the method's responsiveness to changes in obstacle density and height. The trend in TRI and AR values (rho = 0.63, p < 0.05); (rho = 0.67, p < 0.05) showed that the method classifies areas similarly to TRI and AR. Contrarily, VRM (rho = 0.24, p = 0.13) did not align well with the roughness method used. This study highlights the potential to automate and improve roughness assessment, thereby improving operational efficiency and enabling better adjustments in performance expectations and cost estimation in forest operations.