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

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.