JoH paper: 一种基于双机器学习的降水融合方法

我们提出了一种基于双机器学习的降水融合方法,取得了较好的效果。

算法地址:https://github.com/zhanglingky/MLPrecMerg

文章地址: https://doi.org/10.1016/j.jhydrol.2021.125969

Abstract: This study proposed a novel double machine learning (DML) approach to merge multiple satellite-based precipitation products (SPPs) and gauge observations, and tested its reliability and validity over the Chinese mainland. The DML approach was mainly developed based on the classification model of random forest (RF) in combination with the regression models of the machine learning (ML) algorithms including RF, artificial neural network (ANN), support vector machine (SVM) and extreme learning machine (ELM). This led to four DML algorithms, i.e., RF-RF, RF-ANN, RF-SVM, and RF-LM. The performance of the DML algorithms were compared to the single machine learning (SML) algorithms developed based solely on the regression models of RF, ANN, SVM, and ELM, and to the liner merging methods including the inverse error variance weighting, the one-outlier-removed average, and the optimized weight average. In total, we produced twelve precipitation products including four of the DML algorithms, four of the SML algorithms, three of the liner merging methods, and another one generated via the gauge-only interpolation. The precipitation observations at 697 gauges were spatially and randomly divided into two parts (i.e., 70% and 30%), one was used for the training of the ML algorithms or for the interpolation, while the other for the performance evaluations. Results indicate that the DML algorithms outperform the other merging methods, the gauge-only interpolation, and the original SPPs over the Chinese mainland. The median Kling-Gupta efficiency (KGE) ranges 0.67–0.71 for the merged products of DML, which are obviously higher than the original SPPs (0.31–0.54), the linear merged product (0.54–0.55), gauge-only interpolated product (0.62), and the SML-based products (0.47–0.65). The DML-based products also exhibit better performances than the other products in detecting precipitation events with the threshold of 1 mm/day, and outperform the original SPPs regardless of the precipitation thresholds. Further analyses imply that: (i) the DML-based products could outperform the original SPPs even with a small training dataset size; (ii) the superiority of the DML approach to SML is mainly due to that the former can better capture the temporal dynamics of precipitation; (iii) the added values of the merged products of DML relative to the original SPPs and the gauge-only product vary with the sizes of the training dataset; and (iv) the ensemble of the DML algorithms could not further improve the accuracy of the precipitation estimates. This study not only provided an effective and robust tool for the fusion of multiple SPPs and gauge observations, but also, for the first time, compared the performance of various ML algorithms in merging satellite and gauge-based precipitation

Leave a Reply

  Subscribe  
Notify of