Assessing the Performance of Deterministic Precipitation Nowcasting Algorithms with Weather Radar Data and Multimetric Verification

Abstract

This study assesses the performance of four deterministic radar-based nowcasting algorithms—LINDA, SPROG, ANVIL, and Extrapolation—using C-band weather radar data located at Tangerang. Forecasts were generated with the pysteps library and verified up to +96 minutes using spatial inspection, ROC analysis, and Taylor/Target diagrams. At short lead times (+8 to +24 minutes), all algorithms achieved high discrimination skill (AUC > 0.90), with SPROG and LINDA reaching peak AUC values of 0.96 and 0.95, respectively. Beyond +56 minutes, LINDA maintained the highest AUC (0.64), while SPROG and Extrapolation dropped below 0.60. Statistical verification showed that LINDA consistently preserved rainfall structure with correlation coefficients ≥ 0.80 at short range and ~0.65 at +80 minutes. Target diagrams indicated low bias (< ±0.1) and uRMSD stability for LINDA, while SPROG exhibited increasing overdispersion and structural error. Spatially, LINDA captured convective growth and peak intensities more realistically than other methods. These results demonstrate that LINDA offers the most balanced and skillful performance across metrics, especially in maintaining accuracy during medium-range forecasts. The findings support its operational suitability for nowcasting convective rainfall in tropical regions.