Climatological analysis of tropical cyclone impacts on hydrological extremes in the Mid-Atlantic region of the United States

The TCP analysis was performed at the HUC8 subbasin level from 1950 to 2013, when both HURDAT2 and gridded precipitation records are available. A TC event is considered related to the co-occurring precipitation in a HUC8 subbasin if the basin is within the 500 km radius from the center of the TC track. The use of 500 km radius is based on recommendations from a large number of existing literature (Englehart and Douglas 2001, Barlow 2011, Nogueira and Keim 2011, Kam et al 2013, Prat and Nelson 2013, Brun and Barros 2014). For each documented TC event, the HUC8-level TCP was calculated as the total precipitation over the storm duration averaged over all (1/16th degree) precipitation grids within a HUC8 subbasin. Several metrics were computed to describe the characteristics of TCP including the TCP contribution to annual precipitation, TC frequency and duration.

If a TC event produces precipitation in a HUC8 subbasin (based on the above TCP analyses) that contributes flow to a gage location, streamflows from this gage over the event duration plus a lag time of 3 d is considered as TC related flows (TCFs). The choice of 3 d is based on gage observations showing peak discharge generally lagging peak precipitation by approximately 1–3 d. Note that a flow gage can be related to one or more HUC8 subbasins depending on the hierarchical river networks. For example, a mainstem gage is related to all upstream HUC8 subbasins that contribute flows to the gage location.

We constructed for each flow gage two sets of annual maximum daily flow (AMF) series based on: (a) daily flow records including TCF (denoted by AMF_TC) and (b) daily flows excluding TCF (denoted by AMF_NOTC). The AMF series were developed for the water years 1951–2019, when gage flows and HURDAT2 records are available. Water year was used to better capture cyclical patterns of streamflows. We quantified the impact of TC on what we define as large and extreme AMF events. Here large (or extreme) AMFs are defined as the top 10% (or 5%) of AMF events, i.e. top 7 (or 3) AMF events, ranked by the magnitude of flows.

Flood frequency analysis was conducted at each gage based on AMF_TC and AMF_NOTC, respectively; the difference in the magnitude of a design flood was compared to quantify the TC impacts. We first examined the trend in the AMF series using the nonparametric Mann–Kendall test (Kendall 1975; Mann 1945). Where trends exist at the 5% significance level, Sen's slope (Sen 1968) was used to detrend the AMF series before fitting the generalized extreme value (GEV) distribution (Perica et al 2013, Yan et al 2018). The Mann–Kendall test for monotonic analysis of trend combined with the Sen's slope is a widely applied trend analysis approach in hydrologic research (Collaud Coen et al 2020). The GEV distribution combines the Gumbel, Frechet, and Weibull distributions, and uses three parameters (location, scale and shape parameters) for distribution fittings based on L-moments statistics (Hosking and Wallis 1997). The trend analysis performed here used the 'trend' package in R and the L-moments analysis used the 'lmom' package in R. Given the daily resolution of streamflow measurements, we constructed flood frequency curves for the 24 h duration and assessed associated infrastructure design risk given by equation (1) (Chow et al 1988):

$$\begin{equation}R = 1 - {\left( {1 - p} \right)^T},\end{equation} \tag{ 1 }$$

where $R$ is the risk hazard associated with the likelihood of exceeding at least one design event over the infrastructure design life, which is defined by the event return period $T$, $p$ is the annual exceedance probability of a design event. A risk value closer to 0 indicates lower risk of failure. We compared $R$ with and without including TCF for $\,T$ = 100 year. The difference in $R$ results from the difference in the flood frequency curve with and without including TCF.

Common drought indices such as Palmer Drought Severity Index (PDSI; Palmer 1965) focus on droughts at the monthly or weekly timescales, which are not appropriate for capturing the event-scale TC effect on droughts. Similar to Kam et al (2013), here we classified droughts into six categories D1–D6 based on their duration and severity defined by streamflow conditions (table 2). Specifically, a drought is defined when streamflows are continuously below the low flow thresholds, i.e. extreme droughts if <10th percentile flow, severe droughts if <15th percentile flow, and moderate droughts if <20th percentile flow based on daily flow records over 69 water years. A long-term drought is defined if it lasts for over 90 continuous days, and a drought lasting for 30–90 continuous days is defined as a short-term drought. To characterize the impact of TCs on drought, we compared the differences of drought frequency and duration for each drought class based on (a) daily flow records including TCFs, and (b) daily flow records where TCFs were replaced by the average streamflow over the week prior to each TC event.

A total of 211 landfalling TC events occurred within the Mid-Atlantic region over 1950–2013. About 62% of these events lasted for up to 7 d, and about 30% lasted between 7 and 14 d. The most active hurricane season occurred in year 2004 with eight TCs (see figures S1–S3 (available online at stacks.iop.org/ERL/16/124009/mmedia)). The average TC contribution to the annual precipitation (i.e. TCP/P) ranged from 4.2% to 8.7% across the HUC8 subbasins (figure 2). However, there is a remarkable inter-annual variability. For example, the maximum single-year TCP/P ranged from 36.2% to 57.2%, which occurred in 2004 for the majority of SRB and the upper DRB. At the monthly scale, about 70.7%–84.5% of annual TCP occurred during August and September over the region. The mean monthly TCP/P ranges from 16.2% to 29.4% in August and 18.2%–38.4% in September. Within DRB, the annual TCP was relatively higher in the western portions of the basin (e.g. the East Branch Delaware River, Lackawaxen, Lehigh, and Schuylkill River). Within SRB, the annual TCP was relatively higher in the northern and western parts of the basin such as the West Branch Susquehanna River and Sinnemahoning River.

Zoom In Zoom Out Reset image size

Trend analysis detected a statistically significant increasing trend in the annual precipitation in 15 (out of 32) HUC8 subbasins, and the increasing trend was most significant (>3 mm per year) in the northern areas of DRB and SRB (figure 3). However, no statistically significant trend was detected in the frequency, magnitude, or duration of TCP, or TCP/P.

Zoom In Zoom Out Reset image size

4.2.1. Annual maximum flows (AMF)

Despite a significant increasing trend in annual precipitation in 15 (out of 32) HUC8 subbasins, the Mann–Kendall test suggested no trend in AMF_TC at most gages but a significant declining trend at two gages in SRB (figure 3(b)), where the Sen's slope was −3.9 m³ s⁻¹ and −6.4 m³ s⁻¹ per year, respectively. For non-TC driven AMF events (i.e. AMF_NOTC), five gages in SRB showed a significant declining trend (figure 3(c)), ranging from −10.7 m³ s⁻¹ to −3.8 m³ s⁻¹ per year, and three gages (one in SRB and two in DRB) showed a significant increasing trend ranging from 0.7 m³ s⁻¹ to 1.3 m³ s⁻¹ per year. Note that non-TC driven AMFs refers to AMF events dominated by physical mechanism not related to TCs, such as snowmelt and soil moisture excess.

Comparisons of flow magnitudes between AMF_TC and AMF_NOTC (figure 4(a)) suggested that TC was the driver of about 1.4%–15.9% of all AMF events over analyzed 69 water years across the region. For gage locations, the total number of TC-driven AMF events generally decreased as gage elevations increased with a Pearson's r of −0.7 (figure 4(b)). Despite limited contribution of TC to regional precipitation, TC was the dominant driver for extreme AMFs (top 3 AMF events) in the southern DRB (i.e. tributaries of the Christina River and Schuylkill River) and the southwestern portions of SRB (i.e. tributaries of the Lower Susquehanna and Junita River). TC also dominated extreme AMFs in the West Branch River that received the highest amount of TCP in SRB (figure 4(c)). Overall, TC caused at least 67% (2 out of 3) of extreme AMFs for over 50% of the SRB gages, and 25% of the DRB gages. In contrast, large AMFs (top 7 AMF events) were dominated more often by non-TC mechanisms than by TCs. No gage in SRB and only two gages in DRB had over 50% of their large AMFs caused by TC.

Zoom In Zoom Out Reset image size

4.2.2. Design floods and associated design risk

The differences between flood frequency curves derived from AMF_TC and AMF_NOTC were used to quantify the impact of TC on design floods with varied returned periods (figure 5). At the basin level, the impact of TCs on extreme floods was higher in SRB relative to DRB based on the basin-wide average of the differences in design flood magnitudes. The basin-average TC-driven change of the 10-, 50- and 100-year flood was 8.4%, 12.4%, 14.3% in DRB, and was 10.1%, 19.1%, 22.8% for SRB, respectively. The impact of TC tends to increase for floods with a longer return period, which is consistent with the previous analysis showing TCs' dominant role in extreme AMFs. Analysis of TC impacts on the design risk suggested that not accounting for TC-driven floods could increase the basin-average 10-, 50- and 100-year design risk by 18.9%, 37.1%, and 48.0% in DRB, and 27.2%, 57.9%, and 57.6% in SRB, respectively.

Zoom In Zoom Out Reset image size

Strong spatial variability of TCs' impacts on extreme floods and design risk was observed within both basins. TC changed the magnitude of the 10-, 50- and 100-year flood by 1.2%–14.1%, −0.4%–19.6%, and −1.1%–23.3% across all DRB gage locations, and by 1.0%–21.5%, −0.4%–44.4%, and −1.2%–53.0% across SRB, respectively. Taking the 100-year flood as an example (figure 6(a)), greater TC impacts was found at the tributaries in the southern DRB and the southwestern portions of SRB. For the design risk, TC changed the 10-, 50- and 100-year design risk by 0%–33.2%, −1.2%–61.4%, and −3%–66.5% in DRB, and by 0%–70.9%, −1.2%–85.0%, and −2.4%–71.4% in SRB, respectively. Impacts of TCs on design risk were generally higher towards southern areas of the basin (figure 6(b)).While the TC-driven change in design risk was positively correlated with the change in design flood across all gages (Pearson's correlation coefficient >0.85), their relationship is nonlinear. As a result, for over half of the SRB gages and one DRB gage, TC imposed a higher design risk associated with the 50 year event comparing to the 100-year event.

Zoom In Zoom Out Reset image size

Among the mainstem gages in DRB, the greatest impact of TC on extreme floods and design risk was found at the most downstream gage (USGS 01463500 at Trenton, NJ), where TC increased 10 year, 50 year and 100 year floods (design risk) by 6.2% (19.6%), 10.4% (50.2%), and 12.2% (62.6%), respectively (figure 6(c)). For the mainstem gages in SRB, TC showed greatest impact on an upstream gage (USGS 01545500 at Renovo, PA), where TC increased 10-, 50- and 100-year floods (design risk) by 10.9% (35.9%), 23.9% (83.3%), and 29.4% (71.4%), respectively (figure 6(d)).

Analysis of drought frequency for each drought classification (table 2) over the analysis period showed that both DRB and SRB were prone to short-term droughts (D1–D3) lasting from 30 to 90 d, relative to long-term droughts with a duration over 90 d (D4–D6). Among all six drought categories, D3 (short-term extreme drought) was the dominant type of hydrological drought in both basins (figure 7). The following analysis focuses on the most frequent D3 and most extreme drought D6.

Zoom In Zoom Out Reset image size

At the basin level, D3 and D6 were more frequent and widespread in SRB compared to DRB. Over the analysis period, there were 18–43 D3 events across all gage locations in DRB and 23–52 D3 events in SRB. D6 occurred 0–7 times in DRB and 1–12 times in SRB. Spatially, the northern and middle SRB were most susceptible to D3, and D6 occurred most often in the northern SRB (figure 7). Evaluated by the spatial extent of drought, the water year 2002 featured the most widespread D3 drought that covered 90% of the gage locations over the Mid-Atlantic region, and the most widespread D6 drought occurred in the water year 1964 covering 74% of the region (see figures S5–S8). Seasonally, 56%–95% of D3 events occurred between June and September, and at least 50% of D6 events occurred during this period. Whereas the peak months of drought coincided with the hurricane season, they often occurred in different years based on historical records.

Comparisons of drought frequency with and without including the TC effect (figure 8) suggested that up to 23% of D3 droughts and up to 100% of D6 droughts were ended by TC over the Mid-Atlantic region. The total duration of D3 and D6 was reduced by TC by up to 25% (or 408 d) and 100% (or 310 d), respectively. TC impacts on drought exhibited marked spatial variability across the region. In general, the TC effect on alleviating D3 droughts was more pronounced in DRB relative to SRB. Specifically, TC reduced the D3 frequency by an average of 12% in DRB, compared to 7% in SRB; TC reduced the D3 duration in DRB by 11%, relative to 9% in SRB. Spatial analysis (figure 9) revealed that the locations showing greater TC impacts on D3 alleviation (e.g. the southern DRB and southwestern SRB) also had higher TC impacts on extreme floods (as shown in figure 6). For much of both basins, TC was able to alleviate or terminate some long-term drought events (D4–D6). For example, TC terminated the D6 drought that occurred in the water year 1964 for 33% of DRB and 13% of SRB locations, respectively. In some part of SRB such as the West Branch River, TC was able to terminate long-term droughts (D4–D6) months earlier, turning them into short-term droughts (D1–D3), as evidenced by the decreased frequency and duration of D4–D6 and slightly increased D1–D3 (figure 8).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The Mid-Atlantic region is characterized by very complex topography and heterogenous land covers, which interact with TC storm characteristics (e.g. size, intensity, speed of movement), large-scale and regional climate, and local antecedent soil moisture to modulate local-scale hydrological responses to TCs (Konrad and Perry 2010, Barlow 2011, Sun and Barros 2012, Titley et al 2021). The relative importance of each control factor can be highly variable depending on the characteristics of storms and local conditions. For example, while slow-moving TCs are more likely to generate heavy rainfall than fast-moving TCs (Lai et al 2020), fast-moving TCs can lead to more severe flooding over saturated areas than slow-moving TCs over areas with dry antecedent soil moisture conditions (Titley et al 2021). Similarly, as demonstrated in this analysis, the TC effect on floods is not necessarily greater for locations receiving a higher amount of TCP, as nonlinear processes are integrated and propagated downstream in space and time. In the context of future climates, while existing projections converge on an increase in global average TC intensity, there is much less agreement on the magnitude of increase, which varies with the model resolution and hypotheses (Knutson et al 2010, 2020, Yoshida et al 2017, Zhang et al 2020). Substantial discrepancies also exist among model projection of regional variables such as the structure of SST and vertical shear, which are important for TC genesis and characteristics (Knutson et al 2020). Hence, regional TC projection as well as TCs' hydrologic impacts remain major modeling challenges, due to complex nonlinear process dynamics. In this regard, knowledge of local-scale climatological characteristics and hydrological impacts of TCs derived from observations is essential for evaluating models' ability to simulate realistic spatial patterns of variability across a hierarchy of scales and guiding future model development. For instance, simulations of process-based responses in Earth system models (ESMs) are very uncertain at fine spatial and temporal scales. The climatological analysis of TCs at the local scale can be used for evaluations of subgrid level improvements in ESMs through explicit representation of subgrid interactions and/or new parameterizations.

To understand the TC influence on flood generation relative to non-TC mechanisms, flood seasonality was analyzed as a surrogate for distinct flood driving mechanisms, given the seasonality of climate drivers that produce flood-generating precipitation (figure 10). For example, floods occurring between March and April were considered to be driven by snowmelt potentially accompanied by rain associated with extratropical systems, while June–August floods were driven by meso-scale convective systems. Result of flood seasonality (figure 10), as quantified by the circular statistics (described in SI), suggested poor flood seasonality (i.e. considerable inter-annual variability in the timing of flood occurrence) in the southern SRB and over the entire DRB. Strongest flood seasonality was found in the high-elevation northern SRB, where the mean date of floods fell mostly within March that are likely associated with large rain-on-snow events. When considering future flood potentials in the Mid-Atlantic region, flood generation mechanisms may change under changing climate conditions. In projected warmer conditions, floods historically driven by snowmelt or rain-on-snow could shift to rain dominated, implying a changing flood seasonality but not necessarily greater flood potential, as discussed above.

Zoom In Zoom Out Reset image size

Lastly, uncertainties in the observational datasets used for this analysis warrant brief discussions: (a) due to typically lower density of COOP stations in complex terrain and no adjustment for undercatch of snowfall, precipitation underestimation is expected in snowy region (Maurer et al 2002, Livneh et al 2013, Sun et al 2019); (b) given the daily resolution of precipitation and USGS streamflow data, temporal variability within 24 h in TC-related precipitation or flood/drought responses is not captured, although the intraday variability in flood magnitude could be substantial for highly urbanized watersheds.