Portability of stand-level empirical windthrow risk models

Abstract Wind damage to stand edges recently exposed by harvesting is a significant management problem. Large datasets were assembled for clearcut edges in three coastal and three continental locations in British Columbia. The datasets were produced by dividing cutblock boundaries into 25 m deep by 25 m long edge segments. The sample was restricted to cutblocks that had been harvested between 1 and 10 years prior to the most recent aerial photography and further restricted to produce equal sample sizes of 6700 segments per location. Each dataset was then randomly partitioned into 80% for model building and 20% for model testing. Forest cover, ecosystem, wind speed and elevation data were compiled within a geographic information system. Additional topographic variables were derived from digital elevation models. The orientation and exposure of each edge segment was derived with customized scripts. Windthrow polygons were mapped using stereo-photographs. Logistic regression models were fit for each location and were then tested in each other location. Datasets were pooled to enable fitting and testing of generic models. Models correctly predicted outcomes 67–83% of the time for the model building locations. Portability of local models to other locations varied from excellent to poor. The models built for continental locations were the least portable. Calibration by multiplying predictions by the ratio of local mean observed damage to the mean predicted damage substantially improved local model portability. Well-fitting models were produced with pooled coastal, interior and provincial datasets. The similarity between models in the contribution of geographic wind exposure, boundary wind exposure and stand stability factors indicate an underlying consistency in the factors leading to windthrow across the province. The methods are applicable to other forest regions where synoptic weather systems produce damaging winds.