A Radar Object-Based Examination of Rain System Climatology and Including Climate Variability

We know the climate is warming and this is changing some aspects of storms, but we have little knowledge of storm characteristics beyond intensity, which limits our understanding of storms overall. In this study, we apply a cell-tracking algorithm to 20 years of radar data at a mid-latitude coastal-site (Sydney, Australia), to establish a regional storm climatology. The results show that extreme storms in terms of translation-speed, size and rainfall intensity usually occur in the warm season, and are slower and more intense over land between ~10am and ~8pm (AEST), peaking in the afternoon. Storms are more frequent in the cold season and often initiate over the ocean and move northward, leading to precipitation mostly over the ocean. Using clustering algorithms, we have found five storm types with distinct properties, occurring throughout the year but peaking in different seasons. While overall rainfall statistics don't show any link to climate modes, links do appear for some storm types using a multivariate approach. This climatology for a variety of storm characteristics will allow future study of any changes in these characteristics due to climate change.


Abstract
We know the climate is warming and this is changing some aspects of storms, but we have little knowledge of storm characteristics beyond intensity, which limits our understanding of storms overall. In this study, we apply a cell-tracking algorithm to 20 years of radar data at a mid-latitude coastal-site (Sydney, Australia), to establish a regional storm climatology. The results show that extreme storms in terms of translation-speed, size and rainfall intensity usually occur in the warm season, and are slower and more intense over land between ~10am and ~8pm (AEST), peaking in the afternoon. Storms are more frequent in the cold season and often initiate over the ocean and move northward, leading to precipitation mostly over the ocean. Using clustering algorithms, we have found five storm types with distinct properties, occurring throughout the year but peaking in different seasons. While overall rainfall statistics don't show any link to climate modes, links do appear for some storm types using a multivariate approach. This climatology for a variety of storm characteristics will allow future study of any changes in these characteristics due to climate change.

Introduction
1 Heavy rainfall is a significant threat to life and property in many parts of the world, 2 especially when it is accompanied by flash floods (Johnson et al. 2016; Allen and Allen 2016). 3 Many studies have shown the potential for climate change to impact rainfall intensity, but how it 4 will affect other storm characteristics (size, translation speed, orientation, etc.) remains largely 5 unexplored. The first step towards exploring the potential future changes is to establish an 6 observed baseline for a wide variety of storm characteristics. 7 Climatological studies can help us to better understand storm characteristics and their 8 (local and remote) drivers in different seasons. Many studies have used gridded datasets (e.g., 9 global climate models, reanalysis data) to perform climatological studies globally and over 10 specific regions; however, the coarse-resolution of these datasets are often unable to properly 11 capture small scale storms like thunderstorms. Therefore, in order to capture these small scale 12 storms using these datasets, researchers have tried to establish thunderstorm climatologies based 13 on the concept of favourable conditions for thunderstorms, which usually includes a combination 14 of convective available potential energy (CAPE) and vertical wind shear in a region ( reported observation data. They showed that these types of storms are more frequent in 19 December during the afternoon, consistent with the seasonal and diurnal cycle of surface 20 temperature and the maximum availability of heating. Although this approach provides valuable 21 information, environmentally favorable conditions do not necessarily lead to a thunderstorm, 22 causing a misestimation of the thunderstorm frequency. In addition, this approach only provides 23 us with storm frequency and it is unable to provide information on other storm characteristics 24 (Allen and Karoly 2014). 25 Using coarse-resolution datasets, previous authors have tried to investigate the effect of 26 natural climate variability (e.g., El Niño-Southern Oscillation (ENSO), the Indian Ocean Dipole 27 (IOD)) on the rainfall over Australia. Ashok et al. (2003) showed that IOD has significant 28 negative partial correlations with rainfall over the western and southern regions of Australia 29 using an atmospheric general circulation model. Allen and Karoly (2014) employed the ECMWF 30 Interim Re-Analysis (ERA-Interim) data and have found that ENSO has a substantial impact on 31 the spatial distribution of severe thunderstorm environments over the continent. In another 32 reanalysis-based study, Hauser et al. (2020) investigated the winter-spring rainfall variability in 33 southeastern Australia (SEA) during El Niño events by quantifying the contribution of clustered 34 mid-latitude weather systems to monthly precipitation anomalies. The authors found that the 35 cluster with below-average rainfall is more frequent compared to the other clusters during El 36 Niño, which confirmed the general suppression of SEA rainfall during these events. Since 37 precipitation in some regions is correlated with more than one large scale driver, and indices are 38 often correlated with each other, the interconnected nature of precipitation dependence suggests 39 the need for a multivariate rather than bivariate approach to this problem. Maher and Sherwood 40 (2014) applied a multivariate approach to Australian precipitation to disentangle the multiple 41 sources of large-scale variability using the ERA-Interim and Australian Water Availability 42 Project datasets, and they showed that ENSO, blocking, and the intensity and position of the 43 ridge are driving wintertime precipitation in Australia, with a minor role played by the jet 44 intensity and the IOD. All of these studies investigated the effects of natural climate variability 45 on rainfall intensity and frequency, and the relationships between other characteristics of storms 46 (i.e., size, shape, translation speed, etc.) and climate modes are not understood. 47 One way of studying the thunderstorm climatology is by measuring the occurrence of 48 lightning using satellite instruments such as the Optical Transient Detector (OTD) and/or the 49 Lightning Imaging Sensor (LIS Although object-based techniques can provide us with more information on storm 123 characteristics, the investigated storm properties in most of these studies were limited to storm 124 number, area, and rainfall intensity, whereas other storm characteristics like storm translation 125 speed, shape and aspect ratio, orientation, direction and volume could also be of interest. In 126 addition, the object-based techniques employed in these studies are limited by the object 127 split/merge issue, which is a common problem in object tracking methods and can lead to 128 calculating misleading storm properties (Muñoz et al. 2018). 129 In this research, we employ the Method of Object-based Diagnostic Evaluation (MODE) 130 Time Domain (MTD), which is modified by the authors so as to handle splitting and merging of 131 objects. We apply this to the Wollongong (near Sydney) radar, which has around 20 years of 132 records, to establish an object-based climatology of precipitation in different seasons over the 133 radar footprint areas (i.e., Greater Sydney, Illawarra and other land/ocean regions within 150 km 134 of the radar). An effort has been made to group the main contributing storms with similar object-135 based characteristics over this region using clustering algorithms followed by investigating their 136 relationships with different climate modes. 137 This study is presented in seven sections. Section 2 describes the Wollongong radar data 138 and its characteristics. Section 3 introduces the object-based and clustering methods along with 139 the statistics employed in this study. Section 4 describes the study area and section 5 presents the 140 results of the object-based climatology over the study area. Section 6 discusses the results shown 141 in the previous section, and finally, the summary of findings is presented in section 7. 142 The site experiences partial blocking up to 3 dB in the lowest scan (0.5° elevation) from the 147 northwest to the southeast due to the significant terrain associated with the Great Dividing 148

Dataset
Range. The archive for this radar started in November 1996 and continues to operate as of 2021. 149 However, the study period is limited to June 2018 in this study. Several hardware and 150 configuration changes have taken place over the last 24 years. Initially, the radar operated on a 151 10-minute volume cycle with 16-level reflectivity data. In December 1999 the number of 152 reflectivity levels was increased to 64. Between October 2010 and January 2021, a major 153 hardware upgrade delivered 160-level reflectivity data and a 6-minute volume cycle. One 154 significant gap is present in the archive from 1/1/1998 to 15/12/1998. 155 To ensure the accuracy of reflectivity values across the entire dataset, an absolute 156 calibration technique is applied using precipitation radar measurements from the Tropical 157

Rainfall Measuring Mission (TRMM) and the Global Precipitation Measurement Mission 158
(GPM). Satellite overpasses with precipitation are compared with ground radar measurement 159 using the volume matching technique described by Louf et al. (2019), providing a mean 160 calibration value for every pass. Periods of stable calibration are identified, and the mean 161 absolute calibration value for these periods is applied as an offset to the ground radar data. 162 Removal of non-meteorological echoes from reflectivity datasets is challenging. In addition to 163 the ground clutter filtering performed by the signal processor, the technique described by Gabella 164 et al. (2002) is applied using filters for echo continuity and minimum echo area. Unfortunately, 165 this technique is not suitable for removing anomalous propagation which is commonly observed 166 over the adjacent South Pacific Ocean. 167 Reflectivity data is transformed to rain rates using a fitted Z-R relationship derived from 168 9 years of hourly rain gauge data using the Camden Airport AWS (35.04° S, 150.69° E). The A  169 and B coefficients for this relationship were 81 and 1.8 respectively. The maximum rainrate is 170 limited to 100 mm/hr to limit contamination from hail. Volumetric rain rates are transformed into 171 a Cartesian grid at a 0.5 km altitude using the Barnes weighting function and a 2.5km constant 172 radius of influence. The final grid has a horizontal resolution of 1 km and a domain size of 300 173 km by 300 km. 174

Challenges with anomalous propagation 175
Despite all the efforts made in removing non-meteorological echoes, the Wollongong 176 radar site experiences significant anomalous propagation over the ocean within the eastern 177 portion of its coverage. These echoes primarily occur during heat-wave conditions, where strong 178 low-level vertical gradients of humidity and temperature create regions of super refraction for 179 lower elevation scans. The resultant echoes have similar reflectivity gradients, size and shape as 180 precipitation echoes, while also being non-stationary, limiting the effectiveness of any 181 algorithms to remove non-meteorological echoes from reflectivity data (fig S5a; Online 182 Resource 1). In order to reduce the effect of these clutter sources in precipitation estimates, we 183 have opted for a two-step clutter removal process over the whole dataset: 1) Since these clutters 184 often include pixels with low intensity, applying the 3 mm/hr threshold over the convolved data 185 ( The days with extreme clutter were removed manually from the dataset using maximum daily 188 reflectivity maps (see fig S4; Online Resource 1) such that the days with high values of 189 maximum reflectivity over the ocean and low values over land were considered as a day with 190 extreme clutter over the ocean. Several factors may impact the precipitation over these regions and its seasonality. 201 Generally, precipitation peaks in the first half of the year and decreases in the second half 202 (Bureau of Meteorology 2013). In the summer, the easterly (or inland) trough is a major 203 contributor to rainfall over these regions with a peak in the evening. Its impact can be enhanced 204 by interacting with any upper-level troughs or cold fronts crossing over these regions. Frontal 205 systems also bring rainfall to these regions throughout the year, but mostly in winter when the 206 subtropical ridge moves northward over inland Australia (Bureau of Meteorology 2010). Another 207 source of precipitation over these regions is cut-off lows, which can occur at any time of year but 208 are most common during autumn and winter. These can be intense and last up to a week when 209 formed as part of a blocking pair or east coast lows, in which case they are accompanied by long-210 lasting heavy rainfall and gusty winds (Bureau of Meteorology 2007). Northwest cloud bands 211 (stretching from northwest to southeast Australia) can also bring precipitation over these regions. 212 They may interact with cold fronts and cut-off lows over southeastern Australia to produce very 213 heavy rainfall over these regions (Reid et  In this method, the "storm objects" at each time step are the connected pixels higher than a 222 specified threshold in the convolved precipitation map (smoothed by an "n⨯n-pixel" moving 223 window across the map). Every "storm object" at each time-step has a unique label number 224 unless it has overlap with another storm object at the previous time-step (each blob in figure 1b). 225 In this case, it takes the label of the storm object at the previous time-step. If a storm-object has 226 overlap with two (or more) storm objects in the previous time step (merging) or two (or more) 227 objects have overlaps with a storm-object at the previous time step (splitting), the storm-228 object/storm-objects at the current timestep takes/take a new label. Based on these definitions, A 229 "sequence of objects (or 3D objects)" is the connected storm objects in time that has a unique 230 label number and doesn't have split/merge events during its lifetime (connected blobs with the 231 same colours and numbers in figure 1b). Finally, a "storm" is a group of the 3D objects that are 232 connected via split/merging events (The whole diagram in figure 1b). 233 Figure 1a represents an example of running the modified MTD on Wollongong radar data 234 during an event that occurred on 2018-10-4. The lines with different colors are showing the 3D-235 object tracks. In this event, land-originating storms' parts had a southeastward direction and later 236 were merged with the storms' parts that had formed over the ocean. Considering split/merge 237 events in this event has shown how successfully this approach could separate storms' parts over 238 land and ocean before merging which is not possible using the original version of MTD. In 239 addition, with the help of the new approach in tracking the storms, it's possible to extract storm 240 characteristics with more details from different parts of the storms and better calculate 241 characteristics like translation speed, direction and track length during the lifetime of the storms 242 with high rate of split/merge events. 243 In this research, we are studying the extracted characteristics related to "storms" and 244 "storm objects" using the modified MTD method. The selected threshold to filter the objects is 3 245 mm/hr in the convolved data smoothed by a "3×3-pixel" moving window across the map. The 246 storm-object characteristics of interest in this study include: 1) area: the number of pixels in the 247 storm object; 2) translation speed: the ratio of the distance between the volumetric centroid of 248 two connected objects in time to the temporal resolution of the dataset; 3) maximum intensity: 249 the maximum precipitation rate within a storm object; 4) object average intensity: the average 250 precipitation rate of all cells within a storm object; 5) object volume discharge: the volumetric 251 rain rate that passes through the storm object area during a specified period; 6) aspect ratio: the 252 ratio of the minor and major axis of the fitted ellipse over the storm object; 7) object direction: 253 the compass direction of the line connecting the centroids of two consecutive objects in a 254 sequence, and finally, 8) orientation that is the compass direction of the major axis of the fitted 255 ellipse. 256 The studied storm characteristics in this research include: 1) storm area: the average 257 storm snapshot areas in the storm lifetime; 2) storm volume discharge: the average storm 258 snapshot volume discharge in the storm lifetime; 3) storm average intensity: the average of 259 precipitation rate in the storm lifetime; 4) storm max average intensity: the maximum of storm 260 averaged intensity (calculated at each snapshot) in the storm lifetime; 5) storm translation speed: 261 the area-weighted average translation speed of the storm snapshots in the storm lifetime; 6) storm 262 direction: the area-weighted average direction of the storm snapshots in the storm lifetime; 7) 263 storm contributing objects: The number of root storm objects in the storm graph diagram (e.g. in 264 figure 1 storm object numbers 1, 2 and 3 are the root objects in the storm diagram) and 8) storm 265 split/merge event number: the number of split/merge events in the storm lifetime. Note that no 266 thresholds have been applied over the defined storm/storm-object properties in this study. t-SNE is a statistical technique to visualize high-dimensional data by projecting it on a 280 two or three-dimensional map. Here is a brief overview of the main stages in this method: 1) It 281 starts with constructing a probability distribution of similarities over pairs of events in high-282 dimensional data such that a similar pair of events have a higher value compared to the one that 283 is less similar; then, 2) another probability distribution of similarities is defined over the points in 284 the low-dimensional map, and finally, 3) the algorithm minimizes the divergence between two 285 distributions using Kullback-Leibler divergence parameter (KL divergence) between the two 286 distributions with respect to the locations of the points in the map. Note that the KL divergence 287 parameter is a measure of how one probability distribution diverges from another using a 288 gradient descent method. For more details, please refer to the original research paper by Maaten 289 and Hinton (2008). The agglomerative technique is one of the common types of hierarchical clustering in 293 grouping data based on their similarity. This technique works in a "bottom-up" manner by 294 treating each object as a separate group in the beginning. Next, at each step, the two clusters with 295 the most similarity are merged into a bigger cluster and this process continues until all objects 296 are merged into one single big cluster (Subasi 2020). Here we have used the t-SNE algorithm to 297 project our n-dimensional data on a two-dimension map (see fig 9a) and increase the divergence 298 of potential clusters. Then, the agglomerative technique has been employed over the projected 299 data to find the clusters. Similar process has been repeated by applying KMeans clustering 300 algorithm (see fig S3; Online Resource 1). However, based on the density map, the cluster 301 borders have been better recognized by the agglomerative technique. 302 In order to find independent properties as the input for the clustering algorithm, the 303 correlations between all pairs of the storm-object properties have been calculated and the pairs 304 with correlation higher than 0.5 are considered as dependent variables and haven't been used 305 together as the input in the clustering algorithm. Note that all input data are normalized (ranging 306 from 0-1) by dividing each input storm property by its maximum and standard deviation. 307 One of the problems with hierarchical clustering is that it doesn't give information 308 regarding the number of clusters, or where to stop the merging process in the algorithm. In order 309 to overcome this limitation, we have employed the Calinski Harabasz index (CHI) to define the 310 number of clusters which is the optimum value of CHI by increasing the number of clusters. 311 CHI is the ratio of between-cluster variance (VAR B ) to within-cluster variance (VAR W ): Here, N is the population of the data, K denotes the number of clusters, and Z k and z refer 318 to the centroid of cluster k and the entire data, respectively. In the second and third equations, n k 319 is the population of cluster k and x i denotes each member of that cluster (Li et al. 2018 Where n is the number of observations, p is the number of independent variables, y is the 354 n×1 matrix of the dependent parameter, X is the n×p matrix of independent variables. 355 Suppose b is a "candidate" value for the parameter β which is the n×1 matrix of 356 coefficients for independent variables. 357 An effort has been made in this study to investigate the relationships between climate 358 indices (i.e., El Niño-Southern Oscillation (ENSO), Indian Ocean Dipole (IOD), and Southern 359 Annular Mode (SAM)) and storm properties in each cluster (derived from the previous section). 360 Since precipitation can be correlated with more than one index, and the indices are often 361 correlated with each other (

378
The  (fig 2c, 2d, 2m, 2o) shows that typical storms in autumn are more intense 402 compared to the other seasons. However, extreme rain intensity is higher the warmer the season, 403 since storms with maximum intensity above 40 mm/hr (top 10% of storms in maximum  404  intensity) during spring, summer, autumn and winter have occurred 1194, 1662, 1182 and 441  405 times, respectively, during the study timeframe. 406 In autumn, storms tend to move slower (fig 2f, 2n) and look more symmetric (fig 2i) than 407 in other seasons. In summer, storm objects are mostly oriented near 315 ∘ from the positive x-axis 408 while in autumn and spring the object orientation angles mostly change to 330 ∘ and in winter 409 storm objects are mostly oriented west-east (fig 2h). Along with having larger and more severe 410 storms in summer than in winter, summertime storms often include more contributing objects 411 and split/merge events during their lifetimes (fig 2j, 2k). Although all of the mentioned 412 differences are statistically significant based on Kolmogorov-Smirnov test, some storm 413 properties clearly vary with seasons such as rainfall intensity and storm direction. However  (fig 3e-f), which is also true for extreme land-originating storms during autumn 432 compared to their counterparts originating on the ocean. However, typical ocean storms in 433 autumn and also winter (in terms of precipitation intensity) have slightly higher rain rates 434 compared with storms originating on land. Therefore, typical ocean-originating storms are more 435 spatially concentrated with higher rainrate and smaller areas compared to land-originating 436 storms. In summer and spring, both types of storms (land and ocean) mostly move towards the 437 east-southeast. However, in autumn and winter, they have different directions. Land storms still 438 tend to move towards east-southeast, but ocean-storms usually move northward. 439 440 441 442 Fig. 4 PDFs of storm snapshot properties originating on land and transitioning to the ocean 443 (panels a-d) and originating on the ocean and hitting the land (panels e-l). Solid and dashed lines 444 are related to the part of the storms that are over land and over the ocean, respectively. St_no 445 refers to the number of storms that start from land and reach the ocean or vice-versa, and 446 St_snp_L and St_snp_O refer to the number of storm snapshots over the land and the ocean, 447 respectively. All differences between land and ocean distributions are statistically significant 448 based on the Kolmogorov-Smirnov test. 449 450 Next, we examine how storm properties change during the transition of storms between 451 land and ocean. Land-originating storms during summer/spring have higher max intensity over 452 land than later over the ocean (fig 4a, b). However, during the autumn and winter, these 453 differences are smaller (fig 4c, d). Summer and spring storms originating over the ocean and 454 reaching land tend to be smaller (fig 4e, f) without much change in rainfall intensity (fig 4i, j). In 455 addition, autumn and winter ocean-originating storms (that reach the land) are also more 456 spatially concentrated (with higher max intensity (fig 4k, l) and smaller sizes) when they are 457 raining over the ocean compared to when they are over land (fig 4g, h) 6e) and intensity (fig 6f, g). However, winter 488 storms don't have such a peak during the day (fig 6e-f). Note that the afternoon peak intensity 489 (which exists for all seasons except winter) is mostly related to intense storm-objects over land. 490 In all seasons except winter, high-intensity storm-objects over the ocean are more intense around 491 10:00-23:00 UTC (20:00-09:00 AEST; fig 6l, 6p, 6t, 6m, 6n and 6q) compared to their land 492 counterparts, but in the afternoon and evening, the opposite is true. During spring, summer and 493 autumn, fast-moving ocean storm-objects move faster than land storm-objects during 0:00-10:00 494 UTC (10:00-20:00 AEST) with a peak around 5:00 UTC (15:00 AEST; fig 6n and 6r). However, 495 during the other times of the day, extreme land storm-objects have faster translation speeds. Note 496 that in winter, these peaks are less clear and land storm-objects are mostly faster than their ocean 497 counterparts (fig 6z). 498

Clustering analysis 499
Using clustering algorithms, we have grouped the storms with similar object-based 500 properties over the study area. Here, we employed the Agglomerative clustering technique over 501 the projected multi-dimensional storm data on a 2D map using t-SNE algorithm (more details in 502 section 4.2). The selected input storm properties in the clustering algorithm should not be highly 503 dependent on each other. Therefore, we have calculated the correlation between all pairs of the 504 studied storm-object properties to identify the dependent properties. Figure 7 shows the 505 correlated properties at the significance level of 0.01. By considering a threshold of 0.5 for 506 correlation coefficient, we have found that area vs. volume discharge (fig 7a) and maximum 507 intensity vs. average intensity (fig 7h) are highly dependent on each other and should not be used 508 together as the input in the clustering algorithm. Based on this analysis, the selected input 509 properties for the clustering algorithm are: 1) storm area, 2) storm translation speed, 3) storm 510 max intensity and 4) storm direction which is decomposed into x and y components and have 511 been considered as two independent input variables in the clustering algorithm. 512 The bi-variate histograms in figure 7 also show that storm-objects with higher intensities 513 are generally larger in size and volume (fig 7b, 7c, 7f and 7g). In addition, when the size of the 514 storm-objects increases, the shape of the storm-objects (on average) tends to be more linear (fig  515  7d, 7e). The results show that there are five storm clusters (types) with similar object-based 530 properties occurring over the study area. This number is based on the optimum value of CHI 531 against the number of clusters (fig 9b; for more details see section 4.2.2). Figure 8d-h shows the 532 PDFs of storm properties for different storm types with similar quantitative characteristics that 533 have been identified using the Agglomerative and t-SNE algorithms over the study area. Based 534 on these results, the detected storm types have the following characteristics: 535  Type 1 (T1) storms have a peak frequency in autumn and include mostly average size 536 storms with the lowest translation speeds but very high rainfall intensities compared to the other 537 groups. They often move towards the north (over the ocean) to the northwest (hit the land) 538


Type 2 (T2) storms often move south-eastward and include the fastest and largest storms 539 but with low rainfall intensity. They are frequent during the whole year with a frequency peak in 540 spring 541  Type 3 (T3) storms mostly occur during winter with a frequency peak in June, mostly 542 moving northward over the ocean, and include the very slow storms with the smallest size and 543 low intensity compared to the other types. 544  Type 4 (T4) includes the most extreme storm in terms of rainfall intensity and often 545 appears in large sizes moving eastward with low translation speed. They mostly occur during the 546 summer. 547  Type 5 (T5) storms mostly include very fast storms with small sizes and low rainfall 548 intensities that often occur during the winter, mostly moving northward (over the ocean). 549 To further demonstrate the characteristics of each cluster, a video representing the typical 550 storm for each cluster is provided in the supplemental material (Online Resources 2). 551 552 553 554 Fig. 9 Selected regression coefficients for the relationship between climate mode indices 555 and the object-based storm properties in different months. The detail of selection criteria is 556 explained in the manuscript. The coloured bars show the normalized coefficients that are 557 significantly different from zero. Note that the colours represent the storm groups and are 558 matched with the groups' colours in figure 8. 559 560 We also investigated whether significant relationships exist between storm properties and 561 climate mode indices. Although no statistically significant relationships were found when 562 investigating all storms, we have found some significant relationships between climate indices 563 and storm properties in each cluster (identified in the previous section) using a multiple 564 regression model described in section 4.3. Based on this approach, 75 regression models have 565 been produced (3 indices × 5 storm properties × 5 clusters). In order to identify robust 566 associations, we have identified instances when at least five coefficients in a row have the same 567 sign (positive or negative), and among them at least three are significantly different from zero. 568 Since these climate indices often have impacts on weather for a period from 3-9 months, this 569 restriction helps us to better exclude those short periods in which precipitation has a statistically 570 acceptable link with climate indices but probably not in reality. Thus, from all of the results, only 571 five regressions passed this criterion and are shown in figure 9. The results show that during El-572 Niño events in cold seasons, T3 and T5 storms have negative correlation with ENSO in cold 573 season with lower rainfall intensity in El Niño and higher rainfall intensity in La-Niña (fig 9a,  574 9b). ENSO also has an impact on T1 and T3 storms' translation speed during the warm season 575 with a positive correlation (faster in El Niño and slower in La- Niña; fig 9c, 9d). Finally, IOD has 576 also shown to have a positive correlation with T1 storms' translation speed from mid-summer to 577 early winter (Feb, Apr and Jun; fig 9e). Note that in figure 9, all the coefficients have been 578 normalized to derive the partial correlation between every index and storm property as below: Where σ is the standard deviation in this equation, "index coefficient" refers to any 583 calculated coefficient from equation 11 and "index normalized coefficient" is the partial 584 correlation between every index and storm property. 585

587
The results presented in Section 4.1 are broadly consistent with previous studies, but with 588 some notable exceptions. For example, during summer, storms are mostly larger, move faster 589 and are accompanied by higher rainfall intensities compared to the storms in winter (fig 2). This are more intense when they are raining over land between 10:00 to 20:00 (AEST; peaking in the 594 afternoon) compared to when they are raining over the ocean, consistent with the diurnal cycle of 595 surface temperature and the maximum availability of heating over land. The diurnal peak of 596 severe thunderstorm over land during the warm season was also reported by Griffiths et al. 597 (1993)  seasons. In addition, most of the previous studies were focused on the severe thunderstorms or 608 storms with deep convective clouds and high storm tops that are often accompanied by 609 electrification, which are less frequent during the cold seasons. Using lightning records, Dowdy 610 and Kuleshov (2014) also showed that a maximum in lightning activity during the cooler months 611 occurs over the ocean to the east of Australia, which is consistent with our results. However, they 612 reported a higher frequency of thunderstorms during the warm season. Since storms in cold 613 season are small-scale with low rainfall intensity, probably many of them are not accompanied 614 by lightning to be captured by the sensors. Therefore, a large number of storms over the ocean 615 during this season are probably missed in the mentioned study. 616 We have calculated the average wind direction during the rainy days (based on radar 617 data) at 700 hPa and 850 hPa in ERA5 Reanalysis data (see fig S6 and  States. Additionally, we have found that this peak is more prominent during the summer for 632 severe storms raining over land. Storm characteristics also change during the transition of storms 633 between land and ocean like the decrease in max intensity for summertime land-originating 634 storms when moving over the ocean and wintertime ocean-originating storms when moving over 635 land. These variations are probably related to a change in boundary layer instability in this 636 process, and shows the immediate effect of change in air mass characteristics (land/ocean) on the 637 storms. These characteristics can include surface temperature and humidity, sea/land breezes and 638 topographical interactions, the effects of elevated mixed layers advected over the coast, low-level 639 wind shear and convergence. 640 Using clustering analysis, we have found that there are five storm types with similar 641 object-based properties over this region and described in detail in section 5.3. Among these 642 clusters, three storm types might be accompanied by natural disasters, due to their special 643 characteristics. The first storm type (T1) mostly includes the slowest storms that have high 644 rainfall intensities with small areas and mostly move towards the north-northwest with a peak 645 frequency in autumn. we have found that the rainfall intensity in T3 and T5 storm types (that are more frequent in 664 winter) decreases during these events (fig 9a, b). 665 We have found that there are relationships between some object-based storm properties 666 over the study area that have also been reported in previous studies over other regions. For 667 example, we find that storm objects with large volume and size tend to be more linear and are 668 accompanied by higher rainfall intensities (fig 7); this is consistent with the findings of Ayat et. convective systems with a convection-permitting climate model. Large-scale heavy storms over 672 the study area mostly fall in T2 and T4 storm categories, and typical T2 and T4 storms showed 673 that they are mostly frontal systems that elongated/oriented parallel to the front borders. 674 In summary, the results are showing that the storm intensity variations are consistent with 675 diurnal/seasonal cycles and are related to climate mode oscillations. However, other 676 characteristics of the storms like storm size and translation speed do not seem to always follow 677 the same relationship. This suggests that further investigations are required to find a more 678 definitive answer to the effect of atmospheric parameters variations (e.g. temperature, humidity, 679 etc.) on storm properties other than intensity. 680

681
In this study, we establish an object-based storm climatology using an S-band weather 682 radar located near Wollongong, NSW (34.26° S, 150.87° E, 471 m altitude) with more than 20 683 years of records . The study area is the radar coverage region (including land and 684 ocean), within 150 km of the radar. Here, we employed the Method of Object-based Diagnostic 685 Evaluation (MODE) Time Domain (MTD) to detect and track the storms. Using this object-686 based approach helps us to better understand the climatology of storm properties (other than 687 rainfall intensity and frequency) that haven't been explored in the previous studies over the study 688 area. 689 The extreme storms in terms of size, intensity and translation speed are more frequent 690 during summer and spring. Storms in these seasons mostly originate on land, move towards the 691 east-southeast, are larger, faster and more intense compared to the storms originating on the 692 ocean. In these seasons, between ~10am and ~8pm (AEST), the extreme storms raining over land 693 (wherever they originate) are larger and have higher rain rate but slower compared to when they 694 are raining over the ocean (with a peak in the afternoon that is consistent with the diurnal 695 maximum of boundary layer instability). However, the opposite is true for storms later at night 696 into early morning. 697 Although severe storms are more frequent during summer and spring, typical storms in 698 autumn have higher rainfall intensity compared to the other seasons. In addition, the storms 699 (including non-severe ones) are more frequent in autumn and winter compared to summer and 700 spring. During the cold season, storms mostly initiate from the ocean and tend to move 701 northward, which causes more precipitation over the ocean than land. Ocean-originating storms 702 during these seasons like summer and spring are typically smaller and move slower but have 703 higher rainfall intensity than the land-originating storms which still tend to move east-704 southeastward. 705 The results show that the change in the air mass characteristics (land/ocean) can 706 immediately affect the storm properties. For instance, the land-originating storms that cross to 707 the ocean in summer are more intense over land than ocean. However, in winter ocean 708 originating storms that can reach land are more intense over ocean than land. These changes in 709 storm properties during the storm lifetimes might be related to the differences in surface 710 temperature and humidity, sea/land breezes and topographical interactions, the effects of 711 elevated mixed layers advected over the coast, low-level wind shear and convergence. However, 712 further research is needed to find a definitive answer. 713 We have found five types of storms with distinct object-based properties using clustering 714 techniques. The first storm type (T1) peaks in autumn and mostly includes small-scale and slow-715 moving storms but with high rainfall intensities, often moving northward over the ocean. This 716 type of storm has the potential to create flash floods when they move offshore. T2 storms are the 717 largest and fastest storms with low rainfall intensities, often moving southeastward with a peak 718 in spring. This storm type can be accompanied by severe gust winds. T3 storms, include the 719 smallest size storms moving northward with low intensities and translation speed and peak in 720 winter. T4 storms include the most extreme storms in terms of rainfall intensity with large areas 721 often moving slowly towards the east with a peak in summer, and can be a source of devastating 722 flash and riverine floods over the study area in this season. Finally, T5 storms are wintertime 723 small scale storms over the ocean, moving northward with high translation speeds. We also 724 studied the connection between climate modes and storm properties for different clusters using a 725 linear multivariate approach and the results show that during El-Niño events in cold seasons, T3 726 and T5 storms have negative correlation with ENSO in cold season with lower rainfall intensity 727 in El Niño and higher rainfall intensity in La-Niña. In addition, ENSO has an impact on T1 and 728 T3 storms' translation speed during the warm season with a positive correlation (faster in El Niño 729