Assessing the effectiveness of ANN model in spatial downscaling of d4PDF hourly precipitation data: a case study in Japan

Abstract Obtaining high-resolution data is essential in the hydrology field. This study assesses the effectiveness of the ANN technique in spatially downscaling d4PDF hourly data from 20 km coarse resolution to finer 5 km resolution, considering precipitation as the only atmospheric variable. Two sets of d4PDF data have been utilized for downscaling: 20 km and 5 km resolution. This study was conducted through two main experiments: synthetic and realistic. Firstly, the performance of the ANN is assessed for the downscaling of hourly d4PDF data from a synthetic resolution of 20 km to 5 km, which is applied in two regions. Secondly, the ANN algorithm is applied for downscaling hourly d4PDF from 20 km to 5 km resolution under a real-world scenario considering three different regions in Japan. The findings indicate the applicability of the ANN for spatial downscaling of hourly data in the first experiment. In the second experiment, the performance of the ANN varies depending on the region and season. While it captures the long-term mean during winter, it exhibits high biases in the mean, standard deviation (SD), and maximum values, particularly during the summer seasons. The high values of RMSE, MAE, and CC indicate the poor performance of the ANN during this season. The RMSE and MAE range from 0.465 to 3.616 mm/hour and 0.165 to 1.46 mm/hour, respectively. The ANN exhibits potential for downscaling hourly d4PDF from 20 km to 5 km spatial resolution during the winter seasons. The ANN’s limitations in preserving statistical properties diminish its reliability for the downscaling of hourly data particularly in summer. This is pertinent given that precipitation is the only atmospheric variable considered input for the ANN.