A spatial downscaling method for multielement meteorological data: case study from a water conservation area of the upper Yellow River basin

Meteorological datasets with high-precision and high spatiotemporal resolution are an important base in many applications, such as climatology, ecology, and hydrology. To improve the spatial resolution and accuracy of meteorological data with different elements, this study proposes a method, whereby a machine learning (ML) algorithm is jointly applied to spatial downscaling and post-downscaling error correction (ML-ML). Taking a water conservation area of the upper Yellow River basin (UYRB) as a case study, and using the China Meteorological Forcing Dataset (CMFD), four ML algorithms (Gaussian process regression (GPR), neural network (NN), random forest (RF), and support vector machine (SVM)) were selected to verify the effectiveness of ML-ML and explore the optimal downscaling model suitable for different meteorological elements. The experimental results show the following: (1) the CMFD has good applicability in the UYRB; (2) in addition to the RF, the GRP, NN, and SVM models can successfully retain the original spatial distribution patterns of the CMFD dataset and reflect increased spatial detail; and (3) by comparing the performance of the four models in spatial downscaling and error correction of different meteorological elements, we find that the GPR model is best for precipitation, and the SVM model is best for relative humidity, 2-m air temperature, and 10-m wind speed. On the basis of the thinking behind the ML-ML method, the downscaling models applicable to different meteorological elements screened in this study can provide a reference for generating high-precision and high-resolution meteorological datasets.