Stacking machine learning models versus a locally weighted linear model to generate high-resolution monthly precipitation over a topographically complex area

This study applied a stacked generalization ensemble approach to generate high-resolution precipitation estimates and compared its performance with an optimized local weighted linear regression (LWLR) algorithm, a well-known local precipitation-elevation interpolation method incorporating physiographical factors. The stacked generalization ensemble consists of multilayer perceptron neural network (MLP), support vector machine (SVM), and random forest (RF) combined through a meta-learning algorithm with/without rescanning input covariates. Sixteen input covariates, including 2 topographic features, 5 cloud properties, 5 environmental variables, 3 precipitation products (PPs), and inverse distance weighted (IDW) estimates as the precipitation background field were fed into the machine learning models. Hold out approach was adopted for performance evaluation in which 50% of the 174 gauges were used for training, and the rest was used for validation. The results indicated that the overall monthly MAE, RMSE, and rBias of the proposed stacking model for the validation dataset were 3.3%, 6.8%, and 50%, respectively, less than that of the RF model, which is the best individual model. Also, the stacking model outperformed LWLR by decreasing monthly MAE, RMSE, and rBias 10.7%, 19.1%, and 28.6%, respectively. Further analysis implied that (1) the stacked model is more robust than LWLR and less dependent on the density of gauges, thus suitable for areas with scarce gauge coverage; (2) comparing the spatial distribution of mean monthly precipitation maps, generated by stacking and LWLR models, stacking algorithm can successfully screen out the bulls' eyes of background IDW precipitation field, and both patterns are consistent with the topography of the area; and (3) the stacking scheme is found to have a better extrapolation ability than LWLR over high elevations.