Evaluation of quantitative precipitation forecasts over a monsoon-dominated catchment in Kerala

Numerical Weather Prediction (NWP) models have become an integral part of hydrological forecasting and research. Despite their advances and applications in hydrology, these models exhibit significant inherent errors due to imperfect representations of atmospheric physics, inaccuracies in initial conditions, and limitations in parameterisation schemes. Consequently, accurate quantitative precipitation forecasts (QPFs), which are critical for flood forecasting and early warning systems, remain challenging to obtain. Considering these limitations, it is essential to systematically evaluate available global QPFs derived from various NWP models before employing them as forcing inputs for hydrological models. This study evaluates the skill and reliability of three global quantitative precipitation forecast products archived in the TIGGE database - ECMWF, NCEP and NCMRWF, over the Chaliyar river basin, Kerala, against the gauge-based observation data for the Indian summer monsoon season from 2018 to 2023. The forecasts are evaluated at multiple lead times using a comprehensive set of deterministic and probabilistic metrics. The skill of the control forecasts is quantified by the correlation coefficient and root-mean-square error (RMSE), whereas perturbed forecasts were assessed using the mean Continuous Ranked Probability Score (CRPS). The analysis indicated that the NCMRWF model achieved the highest correlation skill, with values of 0.64 and 0.41 at 1-day and 2-day lead times, respectively, outperforming both ECMWF and NCEP. In terms of forecast errors, RMSE values indicated that ECMWF produced lower errors than NCMRWF and NCEP at both 1-day and 2-day lead times. In terms of probabilistic performance, NCMRWF achieved the lowest mean CRPS at a 1-day lead time, followed by ECMWF and NCEP. However, its probabilistic skill declined at the 2-day lead time, as indicated by higher CRPS values. Overall, both deterministic and probabilistic evaluations indicated that NCMRWF outperforms the other two models for the study area. As envisaged, forecasting skill significantly declined with increasing lead time across all models. These results highlight the need for further improvements, such as ensemble post-processing, to enhance the reliability of operational forecasting applications.