Forest stand susceptibility mapping during harvesting using logistic regression and boosted regression tree machine learning models

Abstract

Listen

Translate article

Economic and environmental consequences of harvesting operations are important in forestry because of the damages caused by tree felling and timber transportation. In this research, we analyzed forest stand damage susceptibility using GIS-based logistic regression (LR) and boosted regression tree (BRT) machine learning models for the Shourab Forests. These forests contain valuable stands of Hyrcanian forests in northern Iran that represent many of the plant covers, geographical, and forest management features of this ancient ecosystem. To this end, we applied traversed transect to cover the forest unit. This allowed us to sample any damaged residual tree intersected by the transect resulting in a dataset of 152 damaged trees in the study area. Of this dataset, we utilized 106 damaged trees for training purposes, while the remaining instances (46 damage locations) served as the verification dataset. We utilized eight input factors as stand damage conditioning variables to model damage susceptibility, including distance from the skid trails, distance from the road, forest density, type of forest, topographic position index (TPI), length of slope (LS), altitude, slope aspect, and slope degree. Using these predictive variables, we calculated damage susceptibility using LR and BRT machine learning models, and, subsequently, devised the outputs in ArcGIS to produce the final stand damage susceptibility maps. As the statistical evaluative criteria, we calculated the area under the curve (AUC) and Akaike Information Criterion (AIC) for the validation and training data sets. According to the validation results, the BRT model (AIC = −109.04; AUC = 0.91) strongly outperformed the LR model (AIC = 96.14; AUC = 0.77). According to the results, among the explanatory variables, the slope degree and slope length had more significant impact on the stand damage. Moreover, the susceptibility maps depicted that more than one-third of the forest area were surveyed in high and very high damage susceptibility classes. According to the results mentioned above, using machine learning technique is an inevitable approach in controlling and mitigating the damages caused by human interferences in natural ecosystems especially in mountain forests. Our findings revealed that the susceptibility maps produced by BRT are highly relevant and beneficial for planning forest harvest operations.