Estimating annual runoff in response to forest change: A statistical method based on random forest

Abstract Population growth and climate change have put pressure on policy makers in southwest Western Australia to increase water supply to urban areas. A potential contribution to solving this problem is thinning of forested catchments to increase runoff. This study uses a machine learning approach, random forest, to relate catchment annual runoff to a range of predictors including climate variables and catchment attributes, and to estimate runoff increases from forest thinning. This approach identifies important predictors and enables prediction. The most important predictor is ‘ForestIndex’ calculated from calibrated satellite imagery and providing a consistent surrogate measure of forest density. This approach estimates annual runoff and carefully assesses potential model predictability by three modes of cross-validation. Our approach leads to more accurate annual runoff predictions than linear regression and the Fu’s model (e.g. reducing RMSE by 41% and 63% respectively). We provide an example to predict the change in annual runoff in response to forest reduction under certain rainfall scenarios. The predicted runoff increase varies greatly amongst catchments from zero to 60 mm per 5 unit ForestIndex reduction.