<p>The Loess Plateau in China is considered one of the most ecohydrologically sensitive regions globally, primarily due to its significant spatial and temporal variability in rainfall. Accurately obtaining the spatial distribution of precipitation is crucial for hydrological simulation, ecological restoration and disaster warning. Using the daily rainfall observation of 384 meteorological stations and SRTM elevation data in the Loess Plateau from 1980 to 2020, we systematically evaluated the performance of three typical interpolation techniques including Thin Plate Spline Interpolation (TPS), Inverse Distance Weighting (IDW), and Co-kriging (elevation as covariate) along with three machine learning methods including Random Forest (RF), Support Vector Machine (SVM) and Gaussian Process Regression (GPR). The training set and the validation set were divided using stratified sampling. We assessed the accuracy of different methods in interannual variation, seasonality and ecological zoning scale. The results show that TPS (RMSE = 2.76 mm/d, R2 = 0.71) and IDW (RMSE = 2.75 mm/d, R2 = 0.71)have the best overall performance. The accuracy of the Co-kriging method (R2 = 0.52) is notably compromised in areas of significant elevation change. Conversely, the machine learning method (with R2 ranging from 0.61 to 0.67) demonstrates an advantage in capturing the influence of elevation but tends to underestimate extreme rainfall values. The interpolation uncertainty exhibits seasonal and zonal differences; the largest errors occur in summer (mean RMSE = 5.98 mm/d) and in the gully-dominated regions of the Loess Plateau (Zone A1), while the highest accuracy observed in the sandy and irrigated agricultural areas (Zone C).</p>