GoKawiilRadar reflectivity data
To characterize the three-dimensional structure of storms, weather radar provides the most suitable observational data. Unlike conventional satellite or rain-gauge observations, which offer precipitation intensity on a two-dimensional spatial scale, radar measures the strength of signals reflected to the receiver at different angles, thereby preserving the vertical structure of storms. These reflected signals, known as reflectivity, represent a combination of raindrop size and number concentration and are therefore related to precipitation intensity. Reflectivity is measured as Z (mm6 m−3) and is typically expressed on a logarithmic scale as decibels of reflectivity (dBZ). This study analyses reflectivity directly rather than converting it to rain rate (R), thereby avoiding uncertainties in the Z–R relationship, which is a main source of error in radar-based precipitation retrievals.
This research uses GridRad data version 3.1, which is specifically designed to investigate deep convective systems and evaluate the vertical structure of storms40,41. This dataset provides hourly, three-dimensional reflectivity (0.02° latitude × 0.02° longitude × 1 km altitude) across the contiguous United States from 1995 to 2017, determining the time frame of this study25. GridRad data have been widely used to analyse mesoscale convective systems42,43, hail44 and tropopause-overshooting convection45. It merges reflectivity observation from 125 National Weather Service NEXRAD WSR-88D Level II weather radars46 onto a common three-dimensional grid through a four-step algorithmic procedure: (1) reading raw data; (2) identifying grid volumes; (3) computing space–time weights; and (4) applying weighted binning40. As well as the space–time weighting scheme that reduces noise from low-quality or distant data, GridRad applies several quality control procedures, such as filtering low-confidence echoes and removing ground clutter, to minimize non-meteorological artefacts. Details of these quality control procedures and validation against other radar datasets are described in the algorithm description documentation40. To further reduce residual noise caused by remaining artefacts, this research applies two reflectivity thresholds to identify storms: 20 dBZ to differentiate rain from drizzle, very light rain or non-precipitating echoes, and 40 dBZ (approximately 10 mm h−1 rain rate) to identify heavy rainfall47,48. The identified storms with areas smaller than 100 km2 are excluded to further reduce uncertainties associated with small-scale artefacts.
Research domains
This research focuses on four main cities in Texas: Dallas–Fort Worth, Austin, San Antonio and Houston. We first define urban areas based on the 2019 National Land Cover Database (NLCD)49. To provide a basis for comparison, we set up four rural domains to the north, south, west and east of each city as rural comparisons. In total, we define 20 research domains: one urban and four rural domains for each of the four cities (Extended Data Fig. 1). To determine the locations of the four rural comparison domains, we calculate the spatial extent of city development along the longitude and latitude directions, the average of which is defined as the urban diameter. Because urban influences on precipitation extend beyond city boundaries, we test several distances of 0.5, 1.0 and 1.5 times the urban diameter from the urban domain as rural domains. Distances that are too small may result in overlapping urban influences and reduce contrast between urban and rural storms, whereas larger distances may lead to proximity to other cities owing to the dense distribution of Texas urban areas. We finally select the rural comparison domains as areas one urban diameter beyond each urban domain. For Houston, which is located along the Gulf Coast and has a large urban extent, further adjustments are made. Specially, we move the eastern rural domain northward by 0.5 urban diameters to avoid ocean coverage and move the western rural domain eastward by 0.5 urban diameters to avoid overlap with the Austin region. The final domain configuration is shown in Extended Data Fig. 1. To capture the dynamics of storm evolution approaching and moving away, we define an extended observation window for each domain by adding a further urban diameter in all directions.
Storm identification
We extract individual storm events from the hourly three-dimensional gridded reflectivity data based on two different thresholds (20 dBZ and 40 dBZ). Our method began by locating high-reflectivity grid cells above 40 dBZ in the 20 research domains during the warm season (May–September) across 23 years. These high-reflectivity grid cells serve as starting points, from which we expand to all adjacent rainy grid cells above 20 dBZ in a three-dimensional volume to define the full rainy area. Finally, to capture the entire storm event from the time series, we track overlapping rainy areas across consecutive hourly time intervals. Those overlapping rainy areas are combined into a single storm event. Therefore, distinct and unconnected convective cells are identified as individual storm events. As a result, it is possible to observe several storm events (typically fewer than three) occurring simultaneously within a given domain.
In this way, we preliminarily identify approximately 1,000 to 12,000 storm events across each of the 20 research domains. Through the visual inspection of radar animations, we identify several issues, prompting us to enhance the algorithm with further steps for correction. First, we find that isolated spikes in radar reflectivity data are occasionally misidentified as small storm events. To reduce such artefacts, we exclude storms with a peak rainy area smaller than 100 km2. Second, some distinct storms may be erroneously merged into a single event owing to overlapping rainfall areas. For example, a warm front followed by a cold front may be mistakenly classified as a single storm event. To resolve this, we separate time series containing several periods of high reflectivity into distinct storm events. For example, a storm event with a time sequence of maximum reflectivity as 21, 30, 42, 42, 25, 21, 45, 30 and 26 dBZ will be divided into two separate events, with the split occurring at the time step corresponding to the minimum number of rain grid cells (≥20 dBZ) between peaks. Finally, some storms intersect marginally with research domains, with most of the areas outside the area of interest. To address this, we track the storm motion by calculating the reflectivity-weighted centroid at each time step (weights are defined as reflectivity values above the 40-dBZ threshold). If the storm centroid trajectory does not intersect with the study scope (including a 0.05° buffer zone) and the high-reflectivity area covers less than 20% of the domain, we presume limited urban influences and excluded those storms from analyses. After these refinements, the final dataset includes approximately 1,000–5,000 storm events per domain. A flowchart illustrating storm identification and post-processing procedures is shown in Supplementary Fig. 6.
Storm classification
We classify the identified warm-season storm events into five types: single-cell storms, isolated storms, tropical systems and warm-frontal and cold-frontal storms. Similar to the storm-identification process, we first examine radar animations of numerous storms to subjectively characterize the features of each storm type in Texas. For instance, we observe that single-cell storms and isolated storms are usually localized and short-lived compared with other storms; frontal storms tend to persist longer with warm and cold fronts moving in different directions; tropical systems generally have longer duration and larger rainfall areas, and their large spatial scale causes storm centroids to move more slowly than frontal storms. We then translate these observed storm-performance characteristics into objective criteria based on storm properties, including rainy area, heavy rain area, duration, movement speed and direction: the rainy area and heavy rain area are quantified as the largest two-dimensional contiguous regions with the column-maximum reflectivity exceeding 20 dBZ and 40 dBZ, respectively, throughout the storm passing through the observation window; storm speed and direction are calculated from the movement of storm centroid, in which the hourly centroid displacements are computed, and their median values are used.
... continue reading