Influence of monsoon low pressure systems on South Asian disasters and implications for disaster prediction

Transient atmospheric vortices called monsoon low pressure systems (LPS) generate a large fraction of total rainfall over South Asia and often produce extreme precipitation. Here, we assess the influence of these storms on the occurrence of disasters, using information from the Emergency Events Database (EM-DAT) that we geocoded and then associated with LPS tracks. We show that more than half of hydro-meteorological disasters over South Asia during summer are associated with these LPS events. Weaker LPS (which are called monsoon lows) occur more frequently than stronger LPS (called monsoon depressions), but the stronger LPS produce a larger number of disasters. Furthermore, although many prior studies have shown that the peak rainfall in LPS falls southwest of the vortex centre, the disasters are concentrated on the northern edge of the LPS tracks, along the Himalayas and upper basins of the Ganga and Brahmaputra rivers. Observations show a sharp peak in rainfall on the day of


Introduction
Globally, natural disasters affect many millions of people each year and cause tens to hundreds of thousands of deaths annually [1,2]. While the frequency of disasters seems to have increased over the last few decades [3,4], improvements in warning systems and mitigation strategies have reduced some of their impact [5][6][7][8]. This motivates efforts to not only understand the factors responsible for disasters, but to better forecast and prepare for these events.
South Asia is especially vulnerable to natural disasters, due to its abundant meteorological and seismic hazards, its high population density, and its lower level of economic development [6,9]. Floods and storms are especially prominent in South Asia, producing 60%-80% of the total number of natural disasters there [10,11]. Compared to geological disasters, such hydro-meteorological disasters also have a higher risk of transforming into a large-scale catastrophic disaster [4,12].
Monsoon low pressure systems (LPS) are the principal rain-bearing systems observed over South Asia during boreal summer, forming and propagating within the larger, continental-scale monsoon circulation. These atmospheric vortices have a horizontal scale of about 2000 km and last for about 3-6 days [13,14]. They form most often over the Bay of Bengal and propagate northwestward, with peak precipitation falling southwest of the vortex centre [15,16].
Individual LPS are typically classified as weaker monsoon lows or stronger monsoon depressions, based on the strength of their winds [17][18][19]. The total population of LPS is collectively responsible for a large fraction of monsoonal precipitation over South Asia [13,20,21], and individual LPS are often associated with extreme rainfall events [18,22,23]. A recent study showed that over 80% of the extreme precipitation events over Central India occur on days when an LPS exists [24]. While many studies have explored the association between monsoon LPS and extreme precipitation events [18,24,25], none have yet determined how many natural disasters in South Asia are caused by LPS and their associated rainfall. This task is undertaken here. Specifically, we geocode hydro-meteorological disasters in South Asia from a mass-casualty disaster dataset, then associate these with LPS tracks obtained from an automated tracking algorithm applied to atmospheric reanalysis data [19]. We examine the geographic distribution of these disasters along with the associated time series of rainfall. We also briefly analyze the skill of short-term forecasts of past rainfall associated with disasters, and discuss the utility of such forecasts in disaster preparation. The results are presented in Section 2 and conclusions are provided in Section 3. Data and methodology employed in our study is outlined in Section 4.

Disaster association with LPS tracks
According to EM-DAT, South Asia has experienced 344 disaster events, affecting 2394 disaster locations during June to September of 1990-2019 (Table 1).
More than 90% of these disaster events are of hydrological or meteorological origin. On associating the hydro-meteorological disasters with LPS tracks, we find that more than half of the disaster events can be tied to LPS (Table 2).  Here, we associate the disaster event with an LPS if at least half of the disaster locations constituting a given disaster event are within 800 km of the LPS track. More details regarding this are discussed in the Methods section.
One important result can be noted immediately: although monsoon lows occur more frequently than depressions, the depressions are responsible for more disasters. There is some overlap between the lows and depressions, as a single disaster event can be associated with multiple LPS; this is why the sum of the disaster events associated with lows and those associated with depressions is about 25% larger than the number of events associated with all LPS.
There is a spatial offset between the regions of peak LPS track density and the regions in which LPS-related disasters are most frequent. The precipitation associated with monsoon LPS has a peak over central and western India, with heavy rainfall also observed over eastern India and the foothills of the Himalayas [24]; this is consistent with the location of LPS tracks (brown lines in Figure 1a). In contrast, the hydro-meteorological disasters associated with monsoon LPS are mostly concentrated along the Himalayas and the upper basins of the Ganga and Brahmaputra rivers, with a substantial number also present in other river basins (red dots in Figure 1a). In fact, about 60% of these disasters are due to riverine floods, according to EM-DAT, and such floods can be caused by both local and upstream rainfall, with some temporal offset. For example, the Ganga and Brahmaputra basins are over 1000 km in horizontal extent, and it can take three weeks for water to travel from the uppermost parts of those basins to the river mouth [26]. Our analysis does not account for large temporal offsets between LPS precipitation and resulting floods, but given typical LPS propagation speeds of 2-3 m s −1 and the 800 km radius we use to associate a disaster with an LPS, this allows for about four days of offset between the peak precipitation near the centre of a typical LPS and the occurrence of a disaster. Larger temporal and spatial offsets between precipitation and flooding are expected to shift disasters relative to precipitation for some of the riverine floods; these are not included in our analysis, but a relevant example is discussed in Section 2.3.
A storm-centred composite of disaster frequency confirms that the disasters associated with LPS predominately lie along the northern edge of the LPS tracks ( Figure 1b). This is true even though the peak precipitation in LPS is well-known to typically fall southwest of the vortex centre [13,27,28] (Supplementary Figure 1). Since disaster risk depends on human exposure and vulnerability, as well as geomorphic factors that control the likelihood of a given rainfall event to produce a flood [29], this spatial offset may result from  Figure 2b). This alone shifts the human exposure to LPS rainfall northward relative to the LPS tracks. Geomorphic factors may also be responsible for shifting the frequency of disaster occurrence toward Nepal, as the Himalayas are prone to floods and landslides during extreme precipitation events [32,33].
We now test whether the association between LPS and disasters may occur . Relative risk greater than 1 indicates that the risk of disasters is Influence of monsoon LPS on South Asian disasters increased when LPS are located within 800 km of the grid cell [34]. Most of the domain has a relative risk much greater than 1, with many of the grid points being statistically significant, as inferred from the 95% confidence interval of relative risk being above unity [35].
We next examine the fraction of disasters associated with stronger and weaker LPS (depressions and lows, respectively). The fraction of disaster locations associated with all monsoon LPS in each of the individual 2 • × 2 • subregions is shown in Figure 3a. Over a large part of the domain, more than half of the disaster locations are associated with LPS, even though these regions are under the influence of monsoon LPS for less than 50% of the monsoonal season (Figure 2b). This is consistent with our relative risk calculation above. The contribution of monsoon depressions is above 50% over most of the domain, while disasters associated with lows are more concentrated over Nepal, Bhutan, and northeastern India as well as along the Ganga and Brahmaputra river basins (Figure 3b, c), as these regions experience higher track density of lows than depressions (Supplementary Figure 3).
These results have a few caveats. First, although a large fraction of disasters over central and peninsular India is associated with LPS, the number of disasters occurring over these regions is comparatively low (e.g., Figure 1a).
Second, some disasters over the west coast of India might be caused by precipitation produced by mid-tropospheric cyclones [36]; some of these disasters may be classified here as being associated with LPS because mid-tropospheric cyclones can evolve from LPS [37], and the extension of those cyclones into the lower troposphere would allow those storms to be included in our LPS dataset.

Observed precipitation
The algorithm that we used to detect LPS identified these storms using patterns of winds in the lower troposphere, so the physical link between LPS and the precipitation involved in hydrological disasters still needs to be estab- This is less than the typical peak LPS rainfall [38], but here we are examining the precipitation at the location of disasters, which need not align with the location of the precipitation peak. The mean rainfall on the day of the composite mean disaster (day 0) is about 50% higher for disasters associated with LPS than for non-LPS disasters. This suggests a physical link between LPS, the extreme precipitation produced by LPS, and the occurrence of disasters. Furthermore, peak rainfall on day 0 is higher for disasters associated with depressions as compared to those associated with lows (not shown). Since the EM-DAT dataset from which disaster locations and times are obtained contains no physical environmental variables, it is notable that these precipitation time series result from composites based only on our geocoded EM-DAT coordinates. It is also consequential that precipitation peaks on the day of disasters and not a day or two before, at least in the composite mean, and that precipitation rates are higher for LPS-related disasters.

Precipitation forecast skill
We now assess the skill with which precipitation at the location of LPSassociated disasters can be forecast. We assess this for short-term (1-5 days lead time) predictions made with the ECMWF model, as described in the Methods section. Figure 5a shows the time series of predicted daily rainfall for the disasters associated with monsoon LPS, with forecasts made 24, 72 and 120 hours before each day in the composite time series. Figure 5b shows a similar time series for forecasts made at various times before the day of disaster.
Results from these two methods are shown to provide two distinct perspectives on the predictive skill. For example, Figure 5a shows forecasts made 72 hours before each day in the disaster time series, while Figure 5b shows the forecast made only once 72 hours before the day of the disaster. In both cases, a peak in rainfall occurs at day 0 for various forecast lead times, although the peak decreases in magnitude with increasing lead times. And similar to observed precipitation, the mean rainfall on the day of the disaster (day 0) is about 50% higher for disaster events associated with LPS than for those not associated with LPS (Supplementary Figure 4). The predicted precipitation peak on the day of the disaster is about 20% lower than the observed peak (compare Figures 5 and 4a). The underestimation in the ECMWF model can also be seen in storm-centred composites and a scatter plot of spatially averaged precipitation ( Supplementary Figure 1), where the ECMWF model underestimates the magnitude of peak precipitation. But the model forecasts are, if anything, slightly underdispersive compared to the observations (as indicated by the error bars in Figures 4 and 5), as is also seen in many of the Subseasonalto-Seasonal (S2S) model forecasts [39]. Nevertheless, a clear peak in rainfall with forecasts made even five days before disasters suggests that short-term forecasts can be used to improve disaster warning and preparation.
We now discuss brief case studies of two well-known disaster events: flood-   We found that disasters occur mostly along the northern edge of the main cluster of LPS tracks, while peak rainfall in LPS is well-known to lie southwest of the LPS centre. We confirmed this location of peak rainfall in our track dataset (Supplementary Figure 1), so this discrepancy is not caused by anomalous behavior of the LPS in our dataset. Given that geomorphology, fluvial transport, and groundwater hydrology all control how precipitation is related to flood characteristics, it is not surprising that there is a spatial offset between the peak LPS rainfall and the disaster location. Furthermore, we illustrated how human exposure may contribute to such offsets, with the peak population density located hundreds of kilometers to the north of the peak precipitation in a storm-centred composite of LPS (Supplementary Figure 2).
Observations obtained from GPM show a sharp peak in rainfall on the day of disasters. This peak is more prominent for disasters associated with LPS than for non-LPS disasters, which suggests a physical link between LPS and hydro-meteorological disasters. Numerical weather prediction model forecasts also show a similar peak in precipitation for disasters associated with LPS.
This peak is clearly visible for forecasts made even up to five days before disasters, although the peak rainfall decreases with increasing forecast lead time. These short-term forecasts of LPS tracks and precipitation might thus be useful in designing disaster early warning systems and in improving disaster preparedness.
Although it was not a focus of our analysis, we briefly examine the occurrence of fatalities in LPS-associated disasters. The mean number of fatalities in disasters associated with monsoon depressions is higher than that in disasters associated with lows (Supplementary Figure 5), but the uncertainties are large enough that these differences are not distinct when considering the 95% confidence interval of the means. A distinct difference is found between the average fatalities in disasters associated with multiple LPS and the average in single monsoon lows. This highlights the importance of considering the impacts of LPS traveling through regions that were recently traversed by another LPS, though more work is clearly needed to understand the relevance of geophysical hazards such as elevated river and groundwater levels and the time-evolution of social vulnerability. Future work could also assess the possible influence of LPS characteristics on the severity of disasters as well as the influence of cooccuring phenomena, like the boreal summer intraseasonal oscillation (BSISO), that are thought to alter LPS and their associated precipitation [40][41][42].

Datasets
Disaster information is obtained from the Emergency Events Database [EM-DAT; 43], and is then geocoded using the Wrangler for Emergency Events Database (WEED; https://github.com/rammkripa/weed) package [44]. As our main focus is South Asia, we only consider disasters occurring in India, Pakistan, Bangladesh, Bhutan, Nepal, and Sri Lanka. We also restrict our analysis to the months of June-September, as monsoon LPS are mainly active during this summer monsoon season. EM-DAT classifies disasters into categories based on the primary triggering event; we only consider hydrological and meteorological disasters, thus excluding disasters caused by earthquakes, drought, epidemics, and other events unrelated to precipitation. A single disaster event often affects multiple locations, so here we refer to each potentially large-scale event as a "disaster event", and the constituent locations as "disaster locations". As disaster information over South Asia is more sparse in the earlier part of the twentieth century, we limit our analysis to 1990-2019.

Association of disasters with LPS
A disaster location is associated with a monsoon LPS if the LPS track passes within 800 km of the disaster location anytime during the duration of the disaster. The typical horizontal scale of monsoon LPS is about 2000 km and this 800 km radius was chosen for consistency with prior work that found this value to be near optimal for attributing rainfall to LPS [21,24,49]. Here, we are further assuming that the hydro-meteorological disasters occurring within this 800 km radius are due to the precipitation associated with monsoon LPS; since our identification and tracking of LPS is based on lower-tropospheric winds and not rainfall, this allows us to test this assumption (see Section 2.2).
It should be noted that the use of a fixed radius can include contributions from non-LPS events like unrelated small-scale convective activity [21], but it can also exclude disaster events that occur far from the region of precipitation, like downstream riverine floods. However, the use of a more complex method like catchment analysis is beyond the scope of this study, and for simplicity we use a fixed radius of 800 km to associate disasters with monsoon LPS.
Usually, a single disaster event affects multiple locations. We classify a disaster event as associated with an LPS if at least 50% of its locations are individually associated with the LPS. Although this attribution fraction of 50% as well as the attribution radius of 800 km were chosen somewhat subjectively, our conclusions remain qualitatively unaffected by changes in these parameters. We varied the attribution radius from 500 km to 1000 km and considered attribution fractions ranging from 1% to 100%, verifying that our results did not qualitatively change. Another caveat is that the mapping of disasters to LPS is not always unique, as some disaster events can be associated

Supplementary Discussion
Skill of ECMWF precipitation forecasts LPS storm-centred precipitation composites show that the ECMWF forecasts are able to capture the observed precipitation peak to the southwest of the vortex centre, but it underestimates the peak magnitude (Supplementary Figure   1). Scattered plots of spatially averaged precipitation (within 5 • of the LPS centre, rather than 8 • , to better capture the highest rain rates) shows the ECMWF model having issues forecasting heavy precipitation, especially for rain rates exceeding 20 mm/day (Supplementary Figure 1d). Mean skill score    Average fatalities per disaster event, with disasters stratified by the strength and number of LPS associated with each disaster. Fatalities from disasters associated with only a single depression, only a single low, and with multiple LPS are shown. Means are shown by blue horizontal bars, while error bars mark the 95% confidence interval for the mean. Only the difference in mean fatalities for disasters associated with a single low and multiple LPS systems is statistically significant, evaluated using a t-test at a 95% confidence level. The plots are for the years 1990-2019.