Machine Learning-based Precipitation Merging: Selection of Input Features and Evaluation Benchmarks

Multi-source merging is essential for creating high-quality gridded precipitation datasets. Machine learning (ML) and/or deep learning (DL) algorithms have achieved inspiring success in this area. This study aims to explore two critical yet underexplored issues in ML-based precipitation merging, i.e., the selection of input features and evaluation benchmarks.

The first issue is about input features of ML models, which often include precipitation products such as satellite and reanalysis datasets, along with auxiliary features like topographical and meteorological variables. One major concern is data independence. Many precipitation products, particularly satellite datasets, are calibrated using similar gauge data, yet the impact of this interdependence on ML-based merging performance is largely unknown. Another challenge is the interaction between input features and regional characteristics, such as climatic regimes, topographical features and gauge density, which affects model generalization across regions or scales but receives little attention in current research.

The second issue relates to benchmark selection. Processes like merging, bias correction, downscaling, and interpolation often employ similar supervised learning frameworks: utilizing high-accuracy reference data (e.g., gauge observations) as training labels with various static or dynamic variables (e.g., latitude and longitude, low-accuracy precipitation products) as features. The ambiguous boundaries between these techniques leads to diverse benchmark choices, ranging from original datasets to sophisticated methods such as geographically weighted regression (GWR). This inconsistency fosters subjective and potentially misleading evaluations, impeding progress in merging precipitation datasets with ML methods.

We investigate these issues through a series of experiments merging five precipitation datasets and high-density gauge data in mainland China, using multiple ML methods including random forest, convolutional neural network and artificial neural network with self-attention modules. The experiments involve varying degrees of data dependence, across eight sub-regions with diverse geographical conditions and gauge densities, and are compared against several benchmark datasets and methods.

By controlling the data dependence, our findings highlight its impact on spatial estimation. Additionally, we identify optimal feature selections across different regions and gauge densities. Interestingly, in areas with low gauge density, simple feature sets without auxiliary environmental variables often outperform those with complex predictors. Moreover, our results show that the ML models function more as interpolation rather than merging, suggesting that complex interpolation algorithms such as GWR might serve as more fitting benchmarks. Our work offers critical insights not only for precipitation datasets but also applicable to a wide range of geoscience data, emphasizing the importance of comprehensive evaluations beyond simplistic comparisons and hasty conclusions.