Regionalization and Projection of CLIGEN input parameters based on machine learning technique

Long-term weather sequences generated by the CLImate GENerator (CLIGEN) are widely used as climate input for hydrological and erosion modeling. This study developed a machine learning (ML)-based framework to regionalize and project CLIGEN input parameters by linking gauge-observed parameters with those derived from daily outputs of Global Climate Models (GCMs). The framework was trained and evaluated using observed daily temperature and precipitation, hourly precipitation at 2,405 stations, and daily solar radiation at 130 stations across mainland China over a 44-year period (1971–2014). The ML-estimated CLIGEN parameters exhibited high predictive accuracy, with daily temperature, solar radiation, and precipitation-related parameters generally achieving Kling-Gupta Efficiency (KGE) ≥0.80. For sub-daily precipitation parameters, KGE reached 0.93 for the maximum 30-min intensity (MX.5P) and 0.75 for the time to peak intensity (TimePk). For CLIGEN outputs, the mean absolute relative errors (MAREs) for the average of four precipitation-related variables were all below 9%. The KGE for two indirect erosivity indicators, the R-factor and 10-year storm EI, were 0.96 and 0.88, respectively. For climate change scenarios SSP1-2.6 and SSP5-8.5, main precipitation statistics were projected to increase by at least 3.1% by 2060 and 4.6% by 2100. Using this framework, grid-based datasets of CLIGEN parameter fields at 0.5 degree resolution were produced for mainland China, covering both historical and future periods under two climate scenarios. These datasets enable generation of daily weather sequences for assessment of climate change impacts on runoff and soil loss with CLIGEN and the Water Erosion Prediction Project (WEPP) model.