The identification of some radar reflectivity signatures plays a vital role in severe thunderstorm nowcasting. A weak echo region is one of the signatures that could indicate updraft, which is a fundamental condition for hail production. However, this signature is underutilized in automatic forecasting systems due to the lack of a reliable detection method and the uncertain relationships between different weak echo regions and hail-producing thunderstorms. In this paper, three algorithms related to weak echo regions are proposed. The first is a quasi-real-time weak echo region morphology identification algorithm using the radar echo bottom height image. The second is an automatic vertical cross-section-making algorithm. It provides a convenient tool for automatically determining the location of a vertical cross-section that exhibits a visible weak echo region to help forecasters assess the vertical structures of thunderstorms with less time consumption. The last is a weak echo region quantification algorithm mainly used for hail nowcasting. It could generate a parameter describing the scale of a weak echo region to distinguish hail and no-hail thunderstorms. Evaluation with real data of the Tianjin radar indicates that the critical success index of the weak echo region identification algorithm is 0.61. Statistics on these data also show that when the weak echo region parameters generated by the quantification algorithm are in a particular range, more than 85% of the convective cells produced hail.

The velocity dealiasing is an essential work of automatic weather phenomenon identification, nowcasting, and disaster monitoring based on radial velocity data. The noise data, strong wind shear, and isolated echo region in the Doppler radar radial velocity data severely interfere with the velocity dealiasing algorithm. This paper proposes a two-step velocity dealiasing algorithm based on the minimization of velocity differences between regions to solve this problem. The first step is to correct aliased velocities by minimizing the sum of gradients in every region to eliminate abnormal velocity gradients between points. The interference of noise data and strong wind shear can be reduced by minimizing the whole gradients in a region. The second step is to dealiase velocities by the velocity differences between different isolated regions. The velocity of an unknown isolated region is determined by the velocities of all known regions. This step improves the dealiasing results of isolated regions. In this paper, 604 volume scan samples, including typhoons, squall lines, and heavy precipitation, were used to test the algorithm. The statistical results and analysis show that the proposed algorithm can dealiase the velocity field with a high probability of detection and a low false alarm rate.

Precipitation forecasting is a fundamental task in atmospheric science. Recent advances in deep learning have shown significant promise in spatiotemporal modeling. Nevertheless, existing models still struggle to capture intricate spatial structures and achieve effective spatiotemporal coordination. To tackle these challenges, we introduce the Spatio-Temporal Structural Correlation Modeling Network (CorrCast), a framework for reliable precipitation forecasting. CorrCast incorporates three core modules: the Spatial Correlation Scale-Aware (SCSA) module, the Spatio-Temporal Pooling (STP) module, and the Spatio-Temporal Feed-forward (STF) module. Specifically, the SCSA module leverages multi-scale dilated convolutions and spatial attention mechanisms to enhance the extraction and selection of spatial correlations across varying scales. The STP module employs various pooling operations to enhance the model's ability to forecast precipitation events. The STF module strengthens global spatiotemporal feature modeling through a mutual constraint mechanism between temporal and spatial pathways. Comprehensive experiments show that CorrCast surpasses state-of-the-art methods across multiple metrics, including prediction accuracy (MAE, RMSE), image quality (PSNR), and event-level detection (CSI, POD, FAR). Ablation studies further validate the effectiveness and complementarity of each proposed module.