Meteorological characteristics of line-shaped rainbands in northern Japan and its surrounding seas under climate change

In Japan, torrential rainfall frequently occurs during the warm season, from summer to autumn. Flooding and sediment disasters associated with torrential rainfall pose serious issues. Imada et al. (2020) show that some historical disasters have increased in risk of torrential rainfall due to climate change. Approximately 50–60% of meso-β-scale rainfall events that are not directly affected by tropical cyclones, including typhoons, are line-shaped (Shimura et al. 2000; Tsuguchi & Kato 2014). According to Bluestein & Jain (1985), most line-shaped torrential rains are of the broken-line type or back-building type. However, Ogura (1991) pointed out that most torrential rainfalls occurring in Japan are of the back-building type. Such line-shaped torrential rains are called line-shaped rainbands (hereinafter referred to as ‘LRBs’). Back-building LRBs maintain a quasi-steady state by moving the precipitation cell and generating new ones behind it (e.g., Kato 1998, 2006; Yoshizaki et al. 2000). The locations where LRBs occur depend on topography and convergence zones in the lower atmosphere (e.g., Takemi 2018; Kato 2020); rainfall areas of LRBs are characterized by rainfall amount, shape, and stagnation, and many of those satisfying threshold values have meso-convective system features (Yamada et al. 2012; Unuma & Takemi 2016; Ohya & Yamada 2023).

The number of LRBs has been increasing since 1990 in the Hokkaido region (Yamada et al. 2012), the northernmost island of Japan, which was the target of this study. In addition, it has been shown that years with high or low LRBs occurrence are characterized by warm and cold atmospheric circulation conditions with distinct climatological features during July and August (henceforth July–August), which will be discussed in detail later. A related study notes the following weather patterns present on days of LRBs occurrence: stagnant fronts, approaching typhoons, and low-pressure systems passing through the area. The target cases for this study are defined in accordance with Ohya & Yamada (2023). The definition involves smaller rainfall amounts than in other regions of Japan for LRBs. This definition has identified cases with typical back-building events have been identified, such as those that occurred in August 2010 and September 2014 (Yamada et al. 2012; Ohya & Yamada 2020). Understanding the specific characteristics of meteorological fields conducive to LRBs occurrence within this region holds critical significance. Moreover, investigating the future shifts in these meteorological attributes due to climate change remains imperative.

This study aims to statistically examine the synoptic meteorological patterns of LRBs during the summer season. First, we verify the characteristics of LRBs previously described by Yamada et al. (2012) by utilizing a large ensemble climate dataset. Note that July–August, when LRBs are concentrated, is the target season. In recent years, the probability of occurrences of some most significant torrential rainfall events with LRBs has increased due to climate change. Osakada & Nakakita (2018) suggest that under future climate conditions, torrential rainfall associated with rainy seasonal fronts may occur in the Hokkaido region, which has not been the case in the past climate simulations. In view of the frequent occurrence of floods and landslides caused not only by typhoons and fronts but also by LRBs, and the further escalation of torrential rainfall associated with climate change, Japan is now considering flood control measures in response to climate change. Aiming to improve flood control safety in river basins, the Ministry of Land, Infrastructure, Transport and Tourism (MLIT) has proposed that the average change in 100-year probable daily rainfall will increase by a factor of 1.1 nationwide and 1.15 regionwide in Hokkaido when the global average temperature increases by 2° in comparison to the pre-industrial era. The above study was conducted using ensemble climate change prediction data from the database for policy decision-making for future climate change (d4PDF; Mizuta et al. 2012; Mizuta et al. 2017) with coarse horizontal resolutions and the dynamic downscaling of the same data to a horizontal resolution of 5 km. By using high-resolution datasets in this study, we extracted LRBs in the target area and compared them with observations, and future changes in LRBs in the target area were also shown. Then, the characteristics of the seasonal mean synoptic meteorological fields for the years of high and low LRBs occurrence under each climatic condition are described, and the characteristics of the daily synoptic meteorological fields that consist of the period are clarified. Characteristics of the seasonal mean and daily synoptic meteorological fields, which are associated with LRBs, are discussed for both past and future climates, thus making this study a significant novelty.

In this study, grid-based rainfall data generated from observations and numerical simulations are used. The historical rainfall data are the Radar/Raingauge-Analyzed Precipitation product (Nagata 2011), which is based on radar rain intensity and corrected by ground rain gauges. For past and future climate simulation data, we used 5-km-resolution downscaling datasets from d4PDF (d4PDF-5 km; Yamada et al. 2021). d4PDF-5 km is derived from a 20-km-resolution regional experiment (d4PDF-20 km; Mizuta et al. 2017). Likewise, d4PDF-20 km is derived from a 60-km-resolution global experiment (d4PDF-60 km; Mizuta et al. 2012). Originally, d4PDF-60 km is the simulation data from a global climate model created by adding perturbations to the observational value of sea ice and sea surface temperatures over a 60-year period from 1951 to 2010. The past climate simulations (henceforth the PAST experiments) are based on the observed sea surface temperature (henceforth SST) for 60 years, and they are produced by perturbing of 100 ensemble members, thus producing the equivalent of 3,000 years of data. For the future climate, the observed SST without the interannual trend component is used to add to the global warming effect. The warming effect gives a global mean temperature increase of 2° from the pre-industrial era (the +2 K experiments), corresponding to 2040 in the RCP8.5 scenario. Similarly, a +4 K experiment (i.e., a 4°C increase from pre-industrial times) corresponding to 2090 in the RCP8.5 scenario is also produced. The +2 K/ + 4 K experiments are based on the six representative models (CCSM4, HadGEM2-AO, GFDL-CM3, MRI-CGCM3, MIRM-, MIRM-, and MIRM) selected by cluster analysis of future SST changes in the global atmosphere-ocean coupled models of the Fifth Coupled Model Intercomparison Program (CMIP5). The above 6 SST patterns in the future climate with nine perturbations for +2 K experiments and 15 perturbations for +4 K experiments are applied to the bottom boundary conditions. Since this is also based on 60 years of SST, the data are equivalent to 3,240 and 5,400 years, respectively. The d4PDF-20 km is the result of dynamic downscaling using NHRCM (non-hydrostatic regional climate model; Sasaki et al. 2011) of the Meteorological Research Institute in the East Asia region with the above d4PDF-60 km as the lateral boundary condition. Similarly, d4PDF-5 km is the result of dynamically downscaling using NHRCM with d4PD4-20 km as a preliminary boundary condition in Hokkaido and its surrounding waters (Hoshino et al. 2020; Yamada et al. 2021). For the d4PDF-5 km high-resolution data, 777 years (the PAST experiments), 441 years (the +2 K experiments), and 1,592 years (the +4 K experiments) are available due to the availability of computing resources. The physical processes are the Kain–Fritsch convective parameterization (Kain & Fritsch 1993), and the Mellor–Yamada–Nakanishi–Niino atmospheric boundary layer scheme (closure level: 3; Nakanishi & Niino 2004) was adopted, respectively. The improved MRI/JMA Simple Biosphere (iSiB; Hirai et al. 2007) model was used for the land surface processes. The 5-km mesh elevation data are based on the area-averaged elevation. The data have been confirmed to have a high reproducibility of annual maximum rainfall including topographical influences. Although it cannot resolve individual cumulus clouds, the shape and scale of rainfall areas can be inferred from the downscaling results.

In this study, LRBs were defined as linear and stagnant rainfall areas based on Ohya & Yamada (2023) criteria and were selected from the historical rainfall data and d4PDF-5 km rainfall data if the accumulative rainfall amount in the previous 3 h exceeds 40 mm within a linear stagnant rainfall area, and at least one grid within the rainfall area has rainfall amount of at least 35 mm/h. This rainfall criterion is generally equivalent to related studies in Europe, the U.S., and the tropics (e.g., Nesbitt et al. 2006; Fairman et al. 2017). However, it includes weaker rainfall events than previous studies in Japan (e.g., Unuma & Takemi 2016; Hirockawa et al. 2020). The extraction criteria in previous studies have been found to be insufficient for the Hokkaido region, which is the target of this study. For example, the LRBs that occurred from the Sea of Japan to central Hokkaido in August 2010 (Yamada et al. 2012) were not included. Ohya & Yamada (2023) confirmed that this definition can capture the isolated and stagnant linear rainfall systems that have recently occurred in the Hokkaido region, including at least six disaster LRBs. However, the same method sometimes extracts rainfall along the topography of the Hidaka and Daisetsu Mountain ranges as linearly stagnant rainfall. To improve this situation, when 30% of the rainfall area is covered above 700 m elevation, the rainfall is considered topographic rather than LRBs and is removed.

The classification method for the meteorological fields was based on self-organizing maps (hereafter SOM; Kohonen 1982; Kohonen et al. 2000), combined with principal component analysis (hereafter PCA) and k-means clustering (e.g., Nguyen-Le et al. 2017). SOM has proven to be a suitable match for synoptic meteorology (e.g., Sheridan & Lee 2011). The input information consists of six variables: sea-level pressure (hereafter SLP), geopotential height at 500 hPa, specific humidity at 850 hPa, zonal wind speed, meridional wind speed at 850 hPa, and temperature at 850 hPa. These variables follow Yamada et al. (2012) and Nguyen-Le et al. (2017), but allow for consideration of the pressure configuration, temperature in the lower layers, and water vapor flux. For the analysis, all grid data for the area in the red box in Figure 1 were used. Therefore, the input dimension of the data is 27,060, which is the number of meteorological variables (6 variables) multiplied by the number of grids (82 east–west grids; 55 north–south grids). The number of input data is based on every 6-hourly output value for each ensemble during the summer (July–August): That is, 4 outputs per day × 62 days in July–August per year × 777 years = 192,696-time steps. Since this method is unsupervised learning, all time steps are used as training data. Similarly, there are 109,368-time steps for the +2 K experiments and 394,816-time steps for the +4 K experiments. However, since the meteorological variables or grids are not independent of each other, PCA was first used to reduce the number of dimensions. Furthermore, since the variables used have different dimensions and units, the input information was normalized from 0 to 1 for each grid and each variable. The SOM training information was the top principal component that contributed 85% of the variance of the input data.

The SOM is a nonlinear method for converting the similarity of high-dimensional input data into geometric relationships, and its most important principle is ‘learning while preserving the topological relationship between a neuron and its spatial neighbors (neighborhood learning)’. Machine learning was performed by the following process. First, an initialization process randomly assigns initial values to reference vectors to a vector space with a 20 × 20 geometric arrangement. Next, a competitive process selects the node with the closest distance to the optimal reference vector. Then, a cooperative process calculates the learning weights for the closely located neighboring nodes, weighted by distance and time. Finally, an adaptive process updates the node with the learning amount, and the process is repeated until learning converges. After sufficient learning iterations until convergence, classification into 400 nodes (20 × 20) was performed, and the composite average of each node was used to classify the meteorological fields into 10 patterns by k-means clustering. This method has been shown to be suitable for analyzing meteorological data because it can adequately represent nonlinear relationships. (e.g., Ohba et al. 2016; Nguyen-Le et al. 2017). In the case of datasets with a large number of grids and variables (high-dimensional data), such as meteorological data, the computational cost of training and running a machine-learning model can be exorbitantly high. Using only the top few principal components of the PCA as training data, the computational cost can be reduced while efficiently preserving the data variance. Thus, in this paper, this approach has reduced the computational cost to less than 1%.

In this study, we examined the relationship between the number of LRBs and meteorological fields in Hokkaido and its surrounding areas during July–August using the historical rainfall data and the past and future climate simulations (the +2 K experiments and the +4 K experiments) from the ensemble climate dataset. We found that the number of LRBs in the target area increased approximately 1.51 times at the +2 K experiments and about 2.07 times at the +4 K experiments compared to the PAST experiments. The potential implication of a significant increase in the number of LRBs is a serious hydrological problem. Specifically, extreme heavy rainfall has been relatively rare in the past, but they become more frequent in recent years. The impact of this is to include torrential events of a previously unexperienced magnitude in the huge number of LRBs. The d4PDF dataset used in this study is based on CMIP5 as the sea surface temperature data. There is currently no apparent difference to CMIP6 in climate repeatability. It is not expected to be a major issue in this study, but it is possible that some differences may be seen in areas prone to precipitation when updated to CMIP6. During many LRB years in each climate scenario, the 2-month mean synoptic field of water vapor flux increases from the south to Hokkaido, which correlates with positive SLP anomalies offshore of the Pacific Ocean. This result is revealed to have statistical significance. These findings are consistent with observations.

Moreover, we analyzed the occurrence characteristics of daily meteorological fields for many LRB years, less LRB years, and all years. Since this method is just a classification of meteorological fields based on the spatial distribution of synoptic scale using key properties, it does not allow us to identify regions or conditions of LRBs occurrences or to mention the dynamical processes in detail. However, as described later, the results are sufficient for a statistical understanding of the frequency of synoptic-scale meteorological patterns. The results revealed that (1) low-pressure patterns, (2) frontal patterns, and (3) typhoon/tropical cyclone patterns, which are more likely to cause LRBs, occurred 1.3 times more frequently in many LRB years compared to all years. The expected number of LRBs occurrences associated with these three weather patterns in many LRB years is more than 1.5 times larger than that in all years. Consequently, the combination of weather patterns favorable for LRBs occurrences and the high number of LRBs occurrences in those weather patterns contribute to the increased number of LRBs in many LRB years. The +2 K and +4 K experiments also demonstrated a similar relationship between the number of LRBs occurrences and the daily weather patterns. Therefore, understanding the weather patterns likely to cause LRBs, as derived from historical observations, will be valuable for predicting future climate scenarios.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.