Precipitation Nowcasting using Data-driven Reduced-order Model

Radar-based precipitation nowcasting refers to predicting rain for a short period of time using radar reflectivity images. For dynamic nowcasting, motion fields can be extrapolated using an approximate and localized reduced-order model. Motion field estimation based on traditional Horn-Schunck (HS) algorithm suffers from over-smoothing at discontinuities in non-rigid and dissolving texture present in precipitation nowcasting products. An attempt to preserve the discontinuities using an L1 norm formulation in HS led to the use of Total Variation L1 norm (TVL1). In this paper, we propose a radar-based precipitation nowcasting model with TVL1-based estimation of motion field and Sparse Identification of Non-linear Dynamics (SINDy)-based estimation of non-linear dynamics. TVL1 is effective in preserving the edges especially in the case of the eye of typhoons and squall lines while estimating motion vectors. SINDy captures the non-linear dynamics and generates the subsequent update values for the motion field based on a reduced-order representation. Finally, the SINDy-generated ensemble of motion field is used along with the radar reflectivity image for generating precipitation nowcasts. We evaluated the effectiveness of TVL1 in preserving edges while capturing the motion field from non-rigid surfaces. The performance of the proposed TVL1-SINDy model in nowcasting weather events such as Typhoons and Squall lines are evaluated using performance metrics such as Mean Absolute Error (MAE), and Critical Success Index (CSI). Experimental results show that the proposed nowcasting system demonstrates better performance compared to the benchmark nowcasting models with lower MAE, higher CSI at higher lead times.