Characterization of Vegetation Dynamics on Linear Features Using Airborne Laser Scanning and Ensemble Learning

Linear feature networks are the roads, trails, pipelines, and seismic lines developed throughout many commercial boreal forests. These linear features, while providing access for industrial, recreational, silvicultural, fire management operations, also have environmental implications which involve both active non-active portions of network. Management existing across forests would lead to optimization maintenance construction costs as well minimization cumulative effects anthropogenic footprint. Remote sensing data predictive modelling valuable support tools multi-level this network by accurate detailed quantitative information aiming assess conditions (e.g., deterioration vegetation characteristic dynamics). However, potential remote datasets improve knowledge fine-scale dynamics within forest roads has not been fully explored. This study investigated use high-spatial resolution (1 m), airborne LiDAR, terrain, climatic, field survey data, provide on (i) developing a model characterization LiDAR-CHM cover dynamic (response metric) (ii) investigating causal factors driving using LiDAR (topography: slope, TWI, hillshade, orientation), Sentinel-2 optical imagery (NDVI), climate databases (sunlight wind speed), inventory (clearing width years post-clearing). For these purposes, we evaluated compared performance ordinary least squares (OLS) machine learning (ML) regression approaches commonly used in ecological modelling—multiple (mlr), multivariate adaptive splines (mars), generalized additive (gam), k-nearest neighbors (knn), gradient boosting machines (gbm), random (rf). We validated our models’ results an error metric—root mean square (RMSE)—and goodness-of-fit metric—coefficient determination (R2). The predictions were tested stratified cross-validation against independent dataset. Our findings revealed that rf showed most (cross-validation: R2 = 0.69, RMSE 18.69%, validation dataset: 0.62, 20.29%). informative clearing width, had strongest negative effect, suggesting underlying influence disturbance legacies, post-clearing, positive effect dynamic. long-term suggest timeframe no less than 20 is expected wide- narrow-width exhibit ~50% ~80% cover, respectively. improved understanding around qualitatively quantitatively. from useful short- Furthermore, demonstrates spatially explicit models reliable assessing roads. It provides avenues further research integrate approach with other studies. An patterns features can help sustainable management.