Determining the extent and drivers of attrition losses from wind using long-term datasets and machine learning techniques

Abstract

Abstract In addition to causing large-scale catastrophic damage to forests, wind can also cause damage to individual trees or small groups of trees. Over time, the cumulative effect of this wind-induced attrition can result in a significant reduction in yield in managed forests. Better understanding of the extent of these losses and the factors associated with them can aid better forest management. Information on wind damage attrition is often captured in long-term growth monitoring plots but analysing these large datasets to identify factors associated with the damage can be problematic. Machine learning techniques offer the potential to overcome some of the challenges with analysing these datasets. In this study, we applied two commonly-available machine learning algorithms (Random Forests and Gradient Boosting Trees) to a large, long-term dataset of tree growth for radiata pine (Pinus radiata D. Don) in New Zealand containing more than 157 000 observations. Both algorithms identified stand density and height-to-diameter ratio as being the two most important variables associated with the proportion of basal area lost to wind. The algorithms differed in their ease of parameterization and processing time as well as their overall ability to predict wind damage loss. The Random Forest model was able to predict ~43 per cent of the variation in the proportion of basal area lost to wind damage in the training dataset (a random sample of 80 per cent of the original data) and 45 per cent of the validation dataset (the remaining 20 per cent of the data). Conversely, the Gradient Boosting Tree model was able to predict more than 99 per cent of the variation in wind damage loss in the training dataset, but only ~49 per cent of the variation in the validation dataset, which highlights the potential for overfitting models to specific datasets. When applying these techniques to long-term datasets, it is also important to be aware of potential issues with the underlying data such as missing observations resulting from plots being abandoned without measurement when damage levels have been very high.