Evaluation of IMERG Precipitation Product Downscaling Using Nine Machine Learning Algorithms in the Qinghai Lake Basin

High-quality precipitation data are vital for hydrological research. In regions with sparse observation stations, reliable gridded data cannot be obtained through interpolation, while the coarse resolution of satellite products fails to meet the demands of small watershed studies. Downscaling satellite-based precipitation products offers an effective solution for generating high-resolution data in such areas. Among these techniques, machine learning plays a pivotal role, with performance varying according to surface conditions and algorithmic mechanisms. Using the Qinghai Lake Basin as a case study and rain gauge observations as reference data, this research conducted a systematic comparative evaluation of nine machine learning algorithms (ANN, CLSTM, GAN, KNN, MSRLapN, RF, SVM, Transformer, and XGBoost) for downscaling IMERG precipitation products from 0.1° to 0.01° resolution. The primary objective was to identify the optimal downscaling method for the Qinghai Lake Basin by assessing spatial accuracy, seasonal performance, and residual sensitivity. Seven metrics were employed for assessment: correlation coefficient (CC), root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R2), standard deviation ratio (Sigma Ratio), Kling-Gupta Efficiency (KGE), and bias. On the annual scale, KNN delivered the best overall results (KGE = 0.70, RMSE = 17.09 mm, Bias = −3.31 mm), followed by Transformer (KGE = 0.69, RMSE = 17.20 mm, Bias = −3.24 mm). During the cold season, KNN and ANN both performed well (KGE = 0.63; RMSE = 5.97 mm and 6.09 mm; Bias = −1.76 mm and −1.75 mm), with SVM ranking next (KGE = 0.63, RMSE = 6.11 mm, Bias = −1.63 mm). In the warm season, Transformer yielded the best results (KGE = 0.74, RMSE = 23.35 mm, Bias = −1.03 mm), followed closely by ANN and KNN (KGE = 0.74; RMSE = 23.38 mm and 23.57 mm; Bias = −1.08 mm and −1.03 mm, respectively). GAN consistently underperformed across all temporal scales, with annual, cold-season, and warm-season KGE values of 0.61, 0.43, and 0.68, respectively—worse than the original 0.1° IMERG product. Considering the ability to represent spatial precipitation gradients, KNN emerged as the most suitable method for IMERG downscaling in the Qinghai Lake Basin. Residual analysis revealed error concentrations along the lakeshore, and model performance declined when residuals exceeded specific thresholds—highlighting the need to account for model-specific sensitivity during correction. SHAP analysis based on ANN, KNN, SVM, and Transformer identified NDVI (0.218), longitude (0.214), and latitude (0.208) as the three most influential predictors. While longitude and latitude affect vapor transport by representing land–sea positioning, NDVI is heavily influenced by anthropogenic activities and sandy surfaces in lakeshore regions, thus limiting prediction accuracy in these areas. This work delivers a high-resolution (0.01°) precipitation dataset for the Qinghai Lake Basin and provides a practical basis for selecting suitable downscaling methods in similar environments.