Abstract:Near-surface air temperature (Ta) at high spatiotemporal resolution is of great significance for meteorology, hydrology, agroecology, production and life, and disaster prevention and reduction. The heterogeneity of Ta in mountainous areas is strong, and there are few stations. How to obtain high spatial resolution forecast Ta under complex mountainous terrain remains a challenge in operational applications and scientific fields. The Global Forecast System (GFS) provides forecast Ta products with global coverage, but its low spatial resolution is not applicable to small-scale areas, especially mountainous areas with complex terrain. Currently, downscaling methods are widely used, which transform large-scale Ta into small-scale Ta. Taking Chongqing with complex terrain as a test area, this study uses elevation data with two spatial resolutions (0.25° and 200 m), latitude and longitude, temporal information, and GFS forecast Ta (0.25°) to construct a downscaling model of GFS forecast Ta for Chongqing based on the statistical downscaling method of neural networks. Validated using Chongqing station data in 2020, the root-mean-square error (RMSE) between the downscaled Ta and station Ta reduces from 2.55 ℃ to 1.65 ℃, the correlation coefficient (R) improves from 0.959 to 0.98, and the absolute value of the mean bias (Bias) reduces from 1 ℃ to 0.02 ℃. To assess the spatiotemporal stability of the model, this paper analyses the model errors by four aspects: month, day, different moments, and different forecast timescales. The results show that the downscaled Ta performance is more stable, with the 12-month RMSE decreasing from 2.22-3.15 ℃ to 1.37-1.86 ℃, and the absolute value of Bias decreasing from 0.10-2.24 ℃ to 0.01-0.29 ℃. The improvement is more obvious in the hot season (July to October). Comparing the time series of modelled Ta, GFS Ta, and station Ta at the three sites, the downscaled Ta aligns better with the station Ta, and the RMSE reduces from 1.63-2.54 ℃ to 1.52-1.87 ℃. Comparing the model performance at different moments, the downscaled Ta has smaller RMSE and Bias, and the model significantly improves accuracy even at moments when the GFS Ta errors are larger (03:00 and 09:00). Comparing the two forecast time scales, there is no significant difference in model performance after downscaling. In addition, the error spatial analysis results show that the station locations throughout Chongqing perform better after downscaling, and the effect is obvious at some stations, with the RMSE decreasing from 3.61 ℃ to 1.70 ℃, the R improving from 0.965 to 0.978, and the Bias reducing from -2.95 ℃ to -0.25 ℃. Finally, the model applies to GFS 3 h forecast Ta (0.25°×0.25°) to obtain small-scale Ta (200 m×200 m). The effect of Ta spatial distribution before and after downscaling is demonstrated for four moments (January 1, 03:00; April 1, 03:00; July 1, 03:00; and October 1, 03:00, 2022) in Chongqing. In contrast, the spatial resolution of GFS Ta before downscaling is low and shows an obvious sense of fuzzy patches, while the spatial resolution of Ta after downscaling is significantly improved and shows richer Ta spatial details, especially in high-altitude mountainous areas.
