Remote Sensing–Based Terrain Analysis for Timber Harvesting Planning in Diverse Forest Environments

Sustainable mechanised timber harvesting depends on accurate terrain information to balance operational efficiency, cost-effectiveness, machine safety, and environmental protection. Conventional terrain classification systems (TCS) used in forestry are typically based on stand-level descriptions, manual field surveys, and expert judgement. While effective for broad planning purposes, these approaches provide limited spatial detail, are prone to observer bias, and are challenging to scale across large forest landscapes. This dissertation investigates the potential of remote sensing-based terrain analysis to quantify terrain attributes relevant for mechanised harvesting planning and to evaluate their analytical reliability and operational feasibility across different forest environments, sensor platforms, and spatial scales. The research is structured around three core terrain components underpinning established TCS frameworks: soil conditions, surface roughness, and slope. Using case studies based on UAV Structure-from-Motion (SfM) photogrammetry, airborne laser scanning (ALS), UAV LiDAR, and handheld and terrestrial laser scanning (TLS), terrain metrics were derived and analysed in relation to (a) machine-induced soil deformation, (b) near-ground obstacles, and (c) steepness and terrain complexity across spatial scales. For soil conditions, the work examined whether stump proximity and stump/root reinforcement proxies explain within-site variability in rut development under repeated machine traffic and evaluated models for rut-depth prediction. Models combining terrain-derived predictors with operational variables achieved R-squared values of 0.69-0.85 when predicting observed rut depth across machine passes and trail configurations. For surface roughness, the research operationalises the obstacle-based roughness concept used in forestry terrain classification by combining UAV photogrammetry with deep-learning-based obstacle segmentation. In the high-visibility study setting, this achieved a 95.6% obstacle detection rate (86/90) relative to field-marked obstacles. The work also evaluates continuous DTM-derived ruggedness metrics as complementary layers to characterise within-stand terrain variability and to assess their relationship to the obstacle-based, class-oriented framework. TLS is also used to support near-ground obstacle mapping via supervised point cloud semantic segmentation. For steepness, the analyses quantify the sensitivity of slope and terrain complexity outcomes to sensor type, grid resolution, and point-cloud processing choices. In this study, fine-scale terrain features and local steepness extremes were progressively smoothed as the grid size increased. Multi-sensor comparisons further show that slope-based category outcomes can vary markedly with resolution, and that apparent agreement at coarse grids can be misleading when spatial smoothing reduces the mapped extent and sometimes the occurrence of the steepest classes. Overall, the dissertation concludes that remote-sensing-derived terrain metrics should be generated and interpreted with respect to specific planning objectives and operational constraints, including survey efficiency, processing demand, cost, and data volume. The results show that remote sensing can support a shift from stand- or polygon-level terrain descriptors toward spatially explicit raster- and point-cloud-based terrain layers that are more directly usable for routing, exclusion mapping, and operational planning. However, because terrain attributes are scale- and context-dependent, stand-level TCS thresholds require explicit translation before being applied to continuous surfaces (e.g., by defining aggregation windows, rescaling, or threshold adjustment). No single remote sensing platform is optimal across all applications. Systematic comparison instead supports evidence-based sensor and product selection that aligns analytical requirements with operational feasibility.