Study on two downscaling models for precipitation and temperature based on the CMIP6 climate model — a case study above Lanzhou on the Yellow River

The acquisition of high-resolution temperature and precipitation data is crucial for climate change and flood and drought risk assessments in watersheds. Using measured temperature and precipitation data from 16 meteorological stations above Lanzhou in the Yellow River Basin and NCEP reanalysis data, the optimal grid and atmospheric circulation factors were selected to establish the SDSM statistical downscaling model and the Random Forest (RF) downscaling model. Accuracy evaluations were conducted using station-measured data. Based on this, the preferred atmospheric circulation factors from the MRI-ESM2-0 climate model in CMIP6 were used to further verify the reliability of the two downscaling model results. The results show that: ① During the calibration period (1971–2004), the downscaling model was constructed based on the screened NCEP reanalysis data. The Random Forest model outperforms the SDSM model, with the R^2 and Nash-Sutcliffe efficiency coefficient for daily temperature simulation reaching 0.97, and the R^2 and Nash-Sutcliffe efficiency coefficient for precipitation simulation exceeding 0.64 at the monthly scale. ② During the validation period (2005–2014), using the preferred atmospheric circulation factors of the MRI-ESM2-0 model, compared to the SDSM model, the proportion of absolute error ( E_A ) in the Random Forest model's temperature downscaling where E_A\leqslant 0.5 ℃ increased by 4.3%, while the proportion where E_A\geqslant 1.0 ℃ decreased by 9.1%. The proportion of monthly precipitation simulations where E_A\leqslant 2.0 mm increased by 6.7%, while the proportion where E_A\geqslant 10.0 mm decreased by 15.8%. The Random Forest model also performed better than the SDSM model in both wet and dry months. ③ Overall, both downscaling models are applicable above Lanzhou in the Yellow River Basin, with daily temperature simulation being more effective than precipitation simulation, and the Random Forest model demonstrating relatively higher simulation accuracy. The study results provide a reference for selecting statistical downscaling methods in the region and future climate change research.