Predictor Selection for CNN-based Statistical Downscaling of Monthly Precipitation

Convolutional neural networks (CNNs) have been widely studied and found to obtain favorable results in statistical downscaling to derive high-resolution climate variables from large-scale coarse general circulation models (GCMs). However, there is a lack of research exploring the predictor selection for CNN modeling. This paper presents an effective and efficient greedy elimination algorithm to address this problem. The algorithm has three main steps: predictor importance attribution, predictor removal, and CNN retraining, which are performed sequentially and iteratively. The importance of individual predictors is measured by a gradient-based importance metric computed by a CNN backpropagation technique, which was initially proposed for CNN interpretation. The algorithm is tested on the CNN-based statistical downscaling of monthly precipitation with 20 candidate predictors and compared with a correlation analysis-based approach. Linear models are implemented as benchmarks. The experiments illustrate that the predictor selection solution can reduce the number of input predictors by more than half, improve the accuracy of both linear and CNN models, and outperform the correlation analysis method. Although the RMSE (root-mean-square error) is reduced by only 0.8%, only 9 out of 20 predictors are used to build the CNN, and the FLOPs (Floating Point Operations) decrease by 20.4%. The results imply that the algorithm can find subset predictors that correlate more to the monthly precipitation of the target area and seasons in a nonlinear way. It is worth mentioning that the algorithm is compatible with other CNN models with stacked variables as input and has the potential for nonlinear correlation predictor selection.