Extreme Compound and Seesaw Hydrometeorological Events in New Zealand: An Initial Assessment

Attention is increasingly being turned toward land atmosphere interactions within the wider hydrological cycle when investigating extreme hydrometeorological events. This is particularly the case with the identification of compound and seesaw events. To do so, accurate soil moisture data are essential. Here, soil moisture from three reanalysis products (ERA5‐Land, BARRA and ERA5) is compared to station observations from 12 sites across New Zealand, for an average timespan of 18 years. Soil moisture data from all three reanalyses are used to investigate land‐atmosphere coupling with gridded (observational) precipitation and temperature. This enables compound (co‐occurrence of hot and dry) and seesaw (rapid transitions from dry to wet) events to be identified and examined. No best performing reanalysis data set for soil moisture is evident (median Pearson's r range: 0.78–0.81). All reanalyses successfully capture the seasonal and residual component of soil moisture, but not the observed soil moisture trends at each location. Strong coupling between soil moisture and temperature occurs in all three reanalyses across the predominately energy‐limited regions of the lower North Island and entire South Island. Consequently, these regions reveal a high frequency of compound event occurrence and potential shifts in land states to a water limited phase during compound months. A series of seesaw events is also detected for the first time in New Zealand (terminating approximately one‐fifth of drought events), with a high frequency of seesaw event occurrence detected in previously identified areas of atmospheric river (AR) activity.

In exploring this more holistic approach to extreme hydrometeorological events, the role of soil moisture emerges as a key component due to the feedback loops present in the interaction between land and atmosphere (Seneviratne et al., 2010), requiring data which accurately portrays this process.Similarly, an important first step in investigating extreme hydrometeorological events is to first gain a broader understanding of the land-atmosphere interactions (i.e., coupling) and dependence structure between hydrometeorological variables (e.g., soil moisture and temperature/precipitation) (Tootoonchi et al., 2022).In doing so, a more refined focus is able to be developed to target specific event types, such as compounding and seesaw behavior.
Representation of soil moisture on large spatial scales is often performed via satellite imaging, which are typically of a coarse resolution (Gruber et al., 2020), and as such lack the resolution required for heterogeneous landscapes such as those found in New Zealand.With the improved spatial resolution offered by current generation reanalysis products across large spatial scales (Ling et al., 2021;Muñoz-Sabater et al., 2021), the representation of soil moisture within these models is of key interest (Gevaert et al., 2018).Further, Li et al. (2020) identified a need for more regional performance assessments involving fine scales and diverse topography.New Zealand, displaying a complex topography and varied climate (Macara, 2018), is an ideal candidate for such an assessment.The primary and most commonly employed data set for soil moisture analysis in New Zealand involves a simple water balance approach (Porteous et al., 1994) driven by a high-resolution precipitation and potential evapotranspiration (PET) data set based on statistical interpolation of station observations (the Virtual Climate Station Network (VCSN; Tait et al., 2012;Tait & Woods, 2007)).Such an approach, while computationally simple and available on a fine resolution, is less able to accurately mimic the soil-vegetation-atmosphere coupling compared to climate model simulations of the terrestrial water cycle (Berg & Sheffield, 2018).For example, PET becomes increasingly misrepresentative of actual evapotranspiration (AET) as the land surface moves toward drought conditions (Swann et al., 2016).As a result, Berg and Sheffield (2018) recommended the use of model outputs rather than offline proxy metrics for analysis of soil moisture.Therefore, despite the apparent greater accuracy in the representation of driving variables for soil moisture within the VCSN, the resultant soil moisture data set may be inappropriate for examination of extreme hydrometeorological events across the country, particularly under a changing climate.
New Zealand is exposed to extreme hydrometeorological events, including droughts (Bennet & Kingston, 2022), heat waves (Harrington, 2021;Salinger et al. (2019)) and extreme precipitation (Reid et al., 2021), which are likely to be exacerbated under a changing climate (IPCC, 2021).However, the role that land-atmosphere coupling plays in the development and magnitude of these events remains unexplored.With New Zealand covering multiple climate zones, understanding the characteristics and variation of extreme hydrometeorological events across this mosaic of climates is vital.
New evidence has highlighted a more realistic simulation in the most recent generation of reanalysis data sets in the presentation of precipitation and temperature for New Zealand (Pirooz et al., 2021).If similar performance is found in the representation of reanalysis soil moisture, then land atmosphere interactions would then be able to be explored in more detail than previously possible for New Zealand.Such an examination would not only be insightful for the understanding of existing, well studied extreme hydrometeorological events (heat waves, droughts), but would also enable for the first time an examination of compound and seesaw activity for New Zealand.For example, drought conditions in 2021 on the east coast of the South Island were broken by an atmospheric river event, resulting in heavy flooding (NIWA, 2021).Similarly, the role that the late spring heat wave in 10.1029/2022JD038346 3 of 20 2019 (Harrington, 2021) had on the onset of drought conditions across the North Island the following summer remains unexplored.
The primary aim of this study is to examine the land-atmosphere coupling across New Zealand, and its role during compound and seesaw events.This land-atmosphere coupling is investigated using soil moisture as a proxy, given the controlling nature of soil moisture and its role as a critical variable in land-atmosphere exchanges, with the strength of coupling defined by the correlation of soil moisture (land) with precipitation and temperature (atmosphere).In doing so, the role of soil moisture and land-atmosphere coupling during these compound and seesaw events is explored in a New Zealand context.As a precursor to meeting this primary aim, the relative performance of soil moisture simulation in the current generation reanalysis products will be compared.

Reanalysis Data Sets
Hourly soil moisture data were obtained from the European Reanalysis 5th Generation (ERA5; Hersbach et al., 2020), European Reanalysis 5th Generation Land Component (ERA5-Land;Muñoz-Sabater et al., 2021) and the Bureau of Meteorology (BOM) Atmospheric high-resolution Regional Reanalysis for Australia (BARRA-R; Su et al., 2019), for the period 1 January 1990 to 31 December 2018.Hourly data were first aggregated into daily and then monthly means, before conversion to mm of water.Monthly mean sensible and latent heat flux data were also obtained from each data set.
ERA5 is available at a 0.25° × 0.25° resolution at hourly intervals (Hersbach et al., 2020), while ERA5-Land is available at a resolution of 0.1° × 0.1° and at an hourly temporal resolution (Muñoz-Sabater et al., 2021; Table 1).In contrast to ERA5 and ERA5-Land, BARRA assimilates additional land-surface observations for New Zealand from the National Climate Database (CliFlo; CliFlo, 2021), with the resulting model output from BARRA performing better for precipitation and temperature than both ERA5-Land and ERA5 (Pirooz et al., 2021).BARRA is available on a 0.12° × 0.12° resolution at 10 min to hourly intervals (Su et al., 2019).

Precipitation and Temperature Gridded Data Sets
The VCSN, complied and hosted by the National Institute of Water and Atmospheric Research (NIWA), was selected to provide precipitation and temperature data.VCSN provides daily estimates of climatic data on a 5 km grid covering New Zealand (Tait & Turner, 2005).
VCSN data were accessed for 1 January 1990 to 31 December 2018.Daily estimates are produced based on the daily interpolation of actual data observations made at climate stations located across the country (Tait & Turner, 2005).Temperature was available as daily minimum and maximum values.Monthly means of both minimum and maximum temperature were first calculated, before monthly mean temperature was obtained as the average of the monthly minimum and maximum temperature.Daily precipitation data were summed across each month.
Due to the different grid cell resolutions of the reanalysis products, VCSN monthly total precipitation and mean temperature were regridded (aggregated) to the native resolution of each reanalysis data set (Table 1).Aggregation was performed using the nearest neighbor method.Because analysis was performed on a monthly time step, the ability to capture the statistical properties at fine resolutions was not a dominating consideration (Rajulapati et al., 2021).

Station Observations
To enable comparisons against specific locations, ground station observations of soil moisture were obtained from the NIWA Automatic Weather Station (AWS) network (CliFlo climate database; CliFlo, 2021).Mean monthly soil moisture was used.
Twelve locations were selected as ground station observations (Figure 1), to best represent the complex and varied climate across New Zealand.Locations were first selected based on the seven station temperature series (7SS) of Mullan et al. (2010), originally designed to sample from most parts of New Zealand and which is often used as basis for understanding the national temperature response to climate change.Reefton replaced the Hokitika 7SS location, Paraparaumu replaced Wellington, Martinborough replaced Masterton and Hamilton replaced Auckland, all due to the lack of consistent soil moisture data at the original locations.Additional stations have been added to capture a greater variety of climatic regions throughout New Zealand (Kaitāia, Gisborne, Stratford, Invercargill and Lauder) (Figure 1).The longest station record was Kaitāia (November 1999), with the shortest at Hamilton (December 2005), with an average length of record across all 12 sites of 18 years (212 months) (Table S1 in Supporting Information S1).
A missing monthly value is outputted from CliFlo if there are more than 10 (or 5 consecutive) missing daily observations within a selected month, which numbered n = 34 (1.34%) in the current work.For missing values, the average monthly value for the month concerned was taken across the entire time series of that selected station (i.e., a mean of all January's for the relevant station across the entire time series), to ensure an uninterrupted time series.The CliFlo database returns soil moisture as a percentage of the total soil volume (soil profile depth of 20 cm; with conversion to mm of water being performed by multiplying the percentage of total soil volume by the soil profile depth).

Analysis of Soil Moisture Observations to Reanalysis Data Sets
The closest grid cell at each observation location was identified from each reanalysis data set (Figure 1).Subsequent analysis was then performed between these ground station measurements and grid cell values, with the time series length stipulated by the length of the station record (Table S1 in Supporting Information S1).
Annual cycles at each location were calculated as the mean of each month for all data sets (observations, ERA5-Land, BARRA and ERA5), thereby creating a 12 station series of soil moisture for New Zealand.A single time series for each data set was also constructed by integrating the data across all 12 locations (i.e., mean of all locations; 12 stations).These data set mean time series were then further analyzed by performing seasonal trend decomposition using the Seasonal and Trend decomposition using the Loess method (STL; Cleveland et al., 1990), following the best practice recommended by Gruber et al. (2020).The STL algorithm smooths a time series using locally weighted scatterplot smoothing in two loops: an inner to iterate seasonal and trend smoothing and an outer to reduce residuals or outliers (Cleveland et al., 1990).The resultant outputs include a decomposition of the original time series into its individual constituent components: trend, seasonal and residual.These underlying components were analyzed using Root Mean Square Error (RMSE), standard deviation (SD) and correlation (Pearson's r;Pearson, 1895), with the trend component further analyzed by applying ordinary least square regression on each data set, following Li et al. (2020).All statistical results are reported here as the mean of all bootstrapped samples (10,000 iterations).
At each location, Pearson's correlation coefficients and standard deviation were calculated between the observational data and the corresponding reanalysis grid cells.Comparison between observations and the corresponding reanalysis grid cells was also visualized using scatterplots and marginal distributions (Figure S2 in Supporting Information S1).

Soil Moisture and Precipitation/Temperature Coupling
The representation of land-atmosphere coupling across New Zealand was also investigated, via a simple correlation (Kendall's τ;Kendall, 1938) between monthly mean soil moisture and total precipitation/mean temperature.While correlation cannot demonstrate causality, it can provide an indication of possible physical relationships, especially where causality has already been established (Seneviratne et al., 2010), and has been used successfully to evaluate land-atmosphere coupling (Knist et al., 2017;Li et al., 2017).Monthly total precipitation and mean temperature data from the VCSN were aggregated to the native resolution of each individual reanalysis soil mois ture data set (ERA5-Land, BARRA and ERA5).The VCSN data set was selected to set a consistent representation of precipitation and temperature, allowing any differences in land-atmosphere coupling to then be attributed to the representation of soil moisture within each data set.To investigate land-atmosphere coupling, observational data were removed, allowing the study period to be extended back to the length of the shortest reanalysis data set (1990-BARRA; see Table 1).These extended time series were again decomposed to exclude the seasonal component using STL (Cleveland et al., 1990), before restricting the data sets to the growing season, herein defined as November -March (Salinger, 1987).The focus on growing season was made because of the stronger land-atmosphere coupling typically experienced during the period (Chen & Dirmeyer, 2020).Seasonality was removed to enable more rigorous evaluation of the coupling in mean soil moisture and total precipitation/mean temperature (Li et al., 2020), on the knowledge that seasonal cycles are well captured in reanalysis products (Jiao et al., 2021).
Aggregated, deseasoned mean soil moisture for the growing seasons from 1990 to 2018 from each of the reanalysis data sets was compared to the aggregated, deseasoned total precipitation and mean temperature for the growing seasons from 1990 to 2018, using the Kendall's τ correlation metric.
The aggregated, deseasoned total precipitation, mean temperature and mean soil moisture (from each reanalysis product), were also interrogated for the entire time period; January 1990 to December 2018 (i.e., no growing season restriction).The data were first filtered into dry and wet classifications, representing the lowest/highest third of monthly mean soil moisture (n = 116).Sensitivity testing was also performed using a 20% threshold (Figures S3 and S4 in Supporting Information S1), where the dry/wet month detection showed good agreement between each data set classification (>80%; Table S2 in Supporting Information S1).Monthly soil moisture from each data set was first ranked from highest to lowest, before selecting the top and bottom third to obtain the wet and dry classification.The relationship between this wet and dry classification to the previously utilized growing season (November-March) classification is shown in Table S2 in Supporting Information S1.Total precipitation and mean temperature were then also restricted to these same monthly dates and coupling strength (Soil Moisture-Precipitation (SM-P); Soil Moisture-Temperature (SM-T)) then calculated using Kendall's τ.
Additional analyses were employed utilizing the latent and sensible heat flux data obtained from each data set.
Evaporative fraction (latent/sensible + latent) was calculated using monthly data, after pretreating the data in the same manner as above (e.g., deseasoned, growing season restriction).Following Dirmeyer et al. (2021), critical soil moisture thresholds were then identified for each data set and grid cell using segmented regression.Changes in the slope of the regression lines either side of this threshold level were then obtained, before filtering the results.Namely, these filters included both slopes being significantly different, the dry side of the slope being positive and of larger magnitude than the wet side, with 10 data points existing either side of the threshold.
Significance was tested at the 5% level, after adjusting the degrees of freedom to account for the autocorrelation in the soil moisture data set.
Critical soil moisture thresholds (crit_full) were also obtained using the full evaporative fraction (i.e., no deseasoned or growing season restriction), again using segmented regression.Time series of growing season (November-March) soil moisture were analyzed in relation to this crit_full threshold for the 12 station locations.Wilting levels (wilt) were obtained from both the BARRA-R (Su et al., 2019) and ERA5 (ECMWF, 2018) data sets, while ERA5 Land wilting levels were obtained as a nearest neighbor remapping from the ERA5 wilting level.Monthly soil moisture (SM), for each data sets and each grid cell, was then classed as either dry (SM < wilt), transitional (wilt < SM < crit_full) or wet (SM > crit_full) as noted by Seneviratne et al. (2010).
The evaporative regime for each grid cell (and data set) was then identified as the dominant class (Figure S5 in Supporting Information S1).

Compound and Seesaw Events
The raw monthly total precipitation and monthly maximum temperature for each aggregated VCSN data set was first standardised to a normal distribution, with a mean of zero and standard deviation of one.A one-month accumulation period was utilized, while 12 distributions were fitted (i.e., one for each month) to account for seasonal differences (Kao & Govindaraju, 2010).Standardization was achieved via the Gamma distribution (precipitation; Standardised Precipitation Index, SPI) (McKee et al., 1993), the normal distribution (temperature; Standardised Temperature Index, STI) (Zscheischler et al., 2014), and the Beta distribution (Standardised Soil Moisture Index; SSMI) (Hao & AghaKouchak, 2014;Sheffield et al., 2004).
After transformation to the standard normal distribution, compound events were defined as the co-occurrence of soil moisture (SSMI) below −1, and maximum temperature (STI) above 1 (i.e., bottom/top 32%) at each grid cell to describe the joint dry (soil moisture) and hot (temperature) conditions.This co-occurrence of extremes was examined both as counts of the number of occurrences (months) across the time series , and by applying a Mann-Kendall test (Mann, 1945) at each grid cell to identify any trend in the co-occurrences of hot and dry conditions (Feng et al., 2021).The percentage of compound events during growing seasons (November-March), was also calculated.Similar, the percentage of compound events during dry months (bottom third of ranked soil moisture) and wet months (top third of ranked soil moisture) was also calculated (Figure S6 in Supporting Information S1).In real terms, the co-occurrence of hot and dry conditions can be alleviated by human intervention that is, irrigation, which is expressed within the data sets here via satellite assimilation of soil moisture (ERA5).For New Zealand, irrigation primarily occurs in the mid east coast of the South Island and dominates on the sub-grid scale compared to the reanalysis resolution.Both ERA5-Land and BARRA do not contain satellite assimilation of soil moisture.
Seesaw events were defined and examined following He and Sheffield (2020): an Event Coincidence Analysis (ECA) (Siegmund et al., 2017) was undertaken to identify how frequently droughts (dry periods) are followed by pluvials (wet periods), with a mutual delay of 1 month to capture rapid transitions in hydrometeorological states.The use of a 1 month delay differs to that of He and Sheffield (2020) who employed a 3 months delay to capture seasonal scale transitions.In simple terms, the 1 month delay reflects a change from drought conditions to pluvial conditions during the following month, thus capturing more abrupt endings to dry phases than the 3 months delay imposed by He and Sheffield (2020).Poisson based significance tests were also applied to each land grid cell to identify if the estimated seesaw event occurrence was significant or not.Further in-depth details of the process are contained in the work of He and Sheffield (2020) and Siegmund et al. (2017).
For seesaw events, droughts were defined as any month below the −1 threshold in the SSMI data set, with pluvials identified as those months above the +1 threshold in the SPI.The occurrence of both droughts and pluvials, defined by exceedance of precipitation at the −1/1 level that is, SPI (aggregated VCSN data to the native resolution of each data set; Table 1) was also performed.Periods were also restricted to extended seasons, defined as summer half year (October-March) and winter half year (April-September), with seesaw events identified if a transition from drought to pluvial occurred during these periods.This period selection differs to that used in the analysis of compound events (growing season), and follows the methodology of He and Sheffield (2020) of utilizing summer and winter half year periods.Period restriction was also performed for the growing season (November-March), again identifying any seesaw event as a transition from drought to pluvial during this period (Figure S7 in Supporting Information S1).

Soil Moisture Comparison
Observational data shows a clear seasonal cycle at all sites (Figure 2).Peaks in soil moisture occur in late winter (July/August), with the lowest values recorded in late summer or early autumn (February/March).The highest average soil moisture is recorded at Nelson (123 mm, August), while the lowest average soil moisture is recorded at Paraparaumu (17 mm, March).Annual cycles at each site are captured to varying degrees across the reanalysis data sets, with no one data set emerging as better performing (median correlation of 0.79).Martinborough (average intra range difference of 5 mm to observations range) shows the smallest deviation in annual cycles to observations, while Nelson shows the largest (average intra range difference of 48 mm to observations range).
There is no clear best performing reanalysis data set when assessed on the correlation of the entire time series at each station (Table 2), although median correlation is slightly higher for ERA5 (0.81).Lowest correlation scores (Invercargill, Reefton, Stratford) are present for temperate locations exposed to prevailing westerly or southerly air movement, while the highest correlation scores (Gisborne, Lincoln) occur in locations on the east coast of both islands (with the exclusion of southerly exposed Dunedin).Gisborne has the strongest average correlation across the data sets (0.88), while Reefton has the lowest (0.37).Martinborough, Stratford, Hamilton and Kaitāia are all in close agreement in correlation coefficients, while Gisborne has the largest difference (range of 0.11).
Reanalysis data sets show comparable standard deviations at all sites, with similar median scores (Table 2; range of 0.88).The largest difference in standard deviations between observations and data sets occurs at Nelson (ERA5; 23.37), while ERA5-Land shows the smallest difference to observational standard deviation at Martinborough (0.07).The closest agreement in standard deviation scores to observations exists for east coast locations on both islands (Dunedin, Lincoln, Martinborough, Gisborne), while the poorest agreement relative to observations predominately occur at South Island locations away from the east coast (Invercargill, Lauder, Reefton, Nelson).
Integrated time series (mean of all 12 locations) highlights moderate to strong correlations between the decomposed time series components of observations and reanalysis data sets (Figure 3; correlations of 0.67-0.98 on individual components).Strong variation is present in the observations, with no data set able to adequately capture this variation (Table 3).ERA5-Land shows the greatest agreement in variation (standard deviation 8.94) and magnitude terms (smallest RMSE, 8.13), with ERA5 revealing a consistently smaller magnitude than observational data.Observational data reveal a statistically significant increasing trend in soil moisture (0.56 mm yr −1 ), but this is not captured by any reanalysis data set.Correlation in the trend components to observations (after STL  S1 in Supporting Information S1. decomposition) is similar across all data sets (BARRA (0.68), ERA5 and ERA5-Land (0.67)).RMSE is largest between ERA5 and observations (12.42), and smallest with BARRA (4.18).
Correlations of the seasonal component of the integrated time series demonstrate ERA5 and ERA5 Land as the best performing (0.98), followed by BARRA (0.97) (Figure 3; Table 3), although such differences may remain as products of random noise.Residual range is best captured by ERA5 and ERA5 Land followed by BARRA, with RMSE similarly ranging from 2.86 to 3.26.All reanalysis data sets capture anomalous conditions present in the observational data set, such as the unusually dry summers of 1999/2000 and 2017/2018 (Figure 3d).

Land-Atmosphere Coupling
SM-P correlation (Kendall's τ) shows good agreement across all reanalysis data sets, with statistically significant positive correlations across the entire country (Figure 4; country average of 0.42 across all three reanalysis data sets).SM-T correlation also shows broad agreement between data sets.Significant negative correlations are Note.Note the differences in time periods at each site (Table S1 in Supporting Information S1).Statistical significance (p = 0.01) is indicated by an asterisk.

Table 2 Summary Statistics of Soil Moisture (Correlation, Standard Deviation) at Each Location Between Observations and the Corresponding Grid Cell From Each Reanalysis Data Set
found across all reanalysis data sets for much of the South Island and the lower North Island, matched by mid to high evaporative fractions (Figure 5).The strongest coupling is found throughout the inner montane regions of the lower South Island, with the exception of the BARRA data set which similarly identifies a lower evaporative fraction in these inner montane regions.Good agreement in correlation strength is found amongst the data sets for both the dry and wet months.SM-P correlation during dry seasons indicates widespread significant coupling across the entire country (country average of 0.32 across all three reanalysis data sets), with the strongest correlations across the south and west coast of the South Island (ERA5 and ERA5-Land) and lower east coast of the North Island (Figure 6).Such a pattern is similarly replicated during wet months (Figure 7; country average of 0.30 across all three reanalysis data sets).Significant negative SM-T correlations are again present across much of the country during both dry and wet months, with the exception of the upper South Island and most of the top half of the North Island, similar across all reanalysis data sets.BARRA highlights positive SM-T correlation across these areas during dry months.The emergence of these regions with positive SM-T correlations is stronger (and in agreement across all data sets) during wet months, excluding the upper South Island.

Compound and Seesaw Events
The co-occurrence of hot and dry extremes agrees strongly across the reanalysis data sets, occurring during the growing season for 10-15 months (out of 348 months in the entire time series) for much of the country (Figures 8a-8c).
Lower occurrence occurs during the growing season (<10 months) across the east and north of the North Island, with high occurrence (>15 months) across parts of the east coast of the South Island.This variation between the South Island east coast and northern North Island is reflective of the correspondingly high/low total number of events in these regions (Figures 8d-8f).
As an exception, the east coast of the North Island reveals a relatively high occurrence of total compound months (20-25 months), while the occurrence during growing seasons remains relatively low (5-10 months).Areas of the lower and upper South Island reveal the most frequent occurrences of hot and dry conditions, with a maximum of 35 months across the entire time series (10%; Figures 8d-8f).BARRA also shows a large number of occurrences around the lower middle reaches of the North Island, which is not replicated in ERA5-Land and ERA5, one of the few deviations between data sets, and which is not present amongst growing season event occurrences for the region.Relatively few occurrences of hot and dry conditions exist across the upper and middle sections of the North Island (Figures 8d-8f).
Strong statistically significant increases in the co-occurrence of hot and dry months are present across the west coast, south and lower inner montane regions of the South Island, with significant increases also found across much of the west coast and middle reaches of the North Island (Figures 8g-8i).This spatial coverage agrees across all reanalysis data sets.All reanalysis data sets agree in direction with regards to decreasing trends in hot and dry months in the north east regions of both islands, although this is not statistically significant for BARRA across the north east of the South Island.
Agreement in the representation of seesaw events is present across all reanalysis data sets in the lower east coast regions of the South Island (25%-35%) and the Southern Alps (15%-25%) during the summer half year period, while during the winter half year period agreement is present throughout the lower South Island (25%-35%) (Figure 9).Significant event occurrence (Poisson-based) across the upper east coast of the South Island agrees across all data sets during the winter half year period, although this is weaker in BARRA, with ERA5 and ERA5-Land also being significant during the summer half year period and the full time series.The middle reaches of the North Island contain scattered significant event occurrences throughout all data sets and periods (15%-35%), with the strongest occurrence during the summer half year period.
In comparison to seesaw events defined by SSMI droughts, seesaw events defined by SPI droughts show agreement across the south and north east of the South Island during winter half year periods (Figure 9).In contrast, the south of the South Island reveals significant seesaw event occurrence when droughts are defined using the SPI during summer half year periods and across the full time series, which is not present with SSMI defined droughts.During winter half year periods, the west coast of the North Island reveals a similar contrast between SSMI and SPI defined droughts.

Comparison to Soil Moisture Observations
No substantial differences are detected between the three reanalysis soil moisture data sets (ERA5-Land, BARRA and ERA5).In particular, no data set offers a better performance when compared to station observations (median correlation range of 0.03), nor does any spatial agreement become apparent (Figure 2; Table 2).The similar performance of BARRA to both ERA5-Land and ERA5 indicates that the assimilation of local station observations into the model does not result in significant improvements in the representation of soil moisture, despite the good skill in soil moisture representation within the underlying JULES land surface model (Yang et al., 2014).This absence of any significant improvement in BARRA indicates that the increased resolution (ERA5 Land) and satellite assimilation (ERA5), resulting here in minor performance increases, may be of more significance to increased soil moisture representation skill than the assimilation of primary variables from local station observations.While no substantial differences are detected between data sets, minor improvements in the representation of soil moisture compared to observations (mean of all locations) across New Zealand are apparent in ERA5 and ERA5  Land compared to BARRA, particularly relating to the ability to capture the complete time series and anomalies (Table 3) (average correlations of 0.91 and 0.80 respectively).Within the ERA5 land-surface model, soil moisture is corrected via the use of assimilated satellite observations (de Rosnay et al., 2014), resulting in improvements compared to previous generation reanalysis products globally (Li et al., 2020).Of note, the ERA5-Land data set does not benefit from this assimilation process (Beck et al., 2021).
The lack of significant improvement in soil moisture representation of observation data between ERA5 and ERA5 Land in the current work stands in contrast to the improvements found between ERA5 and ERA5 Land that was achieved via an increase in model resolution (Beck et al., 2021;Muñoz-Sabater et al., 2021).The performance of ERA5-Land in capturing the complexity in soil moisture characteristics and terrain for New Zealand (Hewitt, 2010;Salinger & Mullan, 1999) when downscaled to a fine resolution, itself embedded within the uncertainties of comparing point based with grid scale measurements (Li et al., 2020), may explain these minor differences.Therefore, the relatively minor improvements in soil moisture representation via assimilated satellite observations (ERA5) revealed here (Figure 3; Table 3) provide important findings for the continued advances in regional scale reanalysis products (Su et al., 2021) and the proposed New Zealand Reanalysis (NZRA; Pirooz et al., 2021).Despite the inability of the three reanalysis data sets to capture the observed soil moisture trend (0.56 mm yr −1 ), the accurate portrayal of extreme events and the seasonal cycle in soil moisture, emphasized as the true value in soil moisture representation by Koster et al. (2009), make all three reanalysis soil moisture data sets worthwhile additions to any investigation of extreme hydrometeorological events (Figure 3; Table 3).

Land-Atmosphere Coupling
The strong correlation between soil moisture and precipitation (Figure 4) is typical of a maritime climate (Sehler et al., 2019) and indicative of a strongly responsive hydrological regime, identifying areas potentially exposed to strong seesaw activity (He & Sheffield, 2020).Areas of positive SM-T correlation exist across the upper North Island in the BARRA data set, while during wet months these areas become significantly positively correlated (SM-T) within all data sets (Figure 7).The positive SM-T correlation highlights the strong latent heat flux in the upper North Island, particularly during wet months when an energy limited environment dominates (Figure 5).Dry months, with a weak to no correlation present in the ERA5 and ERA5 Land data sets, indicate a phase change into a transitional regime, further evidenced by the hypersensitive thresholds in the upper North Island.Interestingly, despite a domination of a transitional environment (driven by larger threshold levels for evaporative fluxes), a weaker dominance of latent heat, and hypersensitive threshold levels in the upper North Island, the BARRA data set still reveals positive SM-T correlations during dry periods, indicating a weaker SM-T coupling across the region expressed by BARRA more generally (Figure 6).
The strong SM-T coupling during the growing season across much of the country indicates a change in land states for these typically wet, energy limited regions during dry months, revealing hypersensitive shifts (Figure 5) into transitional or water limiting regimes dominated by soil moisture control and a greater proportion of sensible heat flux (Seneviratne et al., 2010).Thus potential "hot spot" areas of land-atmosphere coupling like that witnessed during the 2018 summer drought and heatwave across the wet, energy-limited regions of Northern Europe and the United Kingdom (Dirmeyer et al., 2021;Orth, 2021) are identified for much of New Zealand during dry months.The hypersensitive threshold levels across the east coast of both islands and north of North Island (Figure 5), and the shift into transitional or water limiting regimes during dry months, provides further evidence that even maritime climates can be impacted by land-atmosphere feedbacks (Orth, 2021), and that the current generation of reanalysis data sets (with their inherent land-atmosphere models) are able to identify this coupling in typically wet climates.Importantly for New Zealand, such land-atmosphere coupling appears to become stronger during dry (drought) months (Figure 6).In New Zealand, drought analysis is frequently performed using offline proxy metrics (Mol et al., 2017), with Berg and Sheffield (2018) noting that global drought projection discrepancies exist because of differences in these offline metrics against land atmosphere model outputs.Such discrepancy is attributed to the inherent coupling in land atmosphere models being able to capture the complex soil moisture-vegetable-atmosphere feedbacks that exist during drought events, and recommend the use of these coupled products for drought assessment (Berg & Sheffield, 2018).In the present work, this land-atmosphere coupling has been shown to exist across New Zealand, particularly during dry months and growing seasons, with previous assessment of drought likely to benefit from a careful reevaluation using these coupled soil moisture products and the land-atmosphere interactions they can portray.

Compound and Seesaw Events
Locations revealing strong negative SM-T correlation simultaneously reveal high occurrence of compound events (Figures 8d-8f).This is further evidenced by the relatively high occurrence of compound events during the growing season (Figures 8a-8c), whereby SM-T relationships are similarly strengthened across much of the country.This spatial agreement suggests soil moisture drought plays some combination of roles as a driver and/or outcome of heat wave occurrence.An ever-growing body of research internationally (Hao et al., 2020;Wu et al., 2021;Zscheischler et al., 2018) indicates how widespread and substantial the impacts these co-occurring, or compounding, events can have.Given the uniform probability distribution of the SSMI (∼60 months in drought conditions; −1 SSMI), hot conditions occur concurrently between 8% (upper North Island) and 58% (lower South Island) with droughts, or a maximum of 10% of total months between 1990 and 2018.Therefore, with the current work revealing that compounding effects are present throughout New Zealand, further work is urgently required in exploring the role heat waves may play in the onset of flash droughts (Mo & Lettenmaier, 2015), or the role drought may play in priming the land surface for heat wave onset (Dirmeyer et al., 2021).
While a relatively cool climate, heat waves in New Zealand have recently come under increased scrutiny, with developments highlighting the importance of relative heat (Harrington, 2021) and the role of sea surface temperatures on atmospheric conditions (Salinger et al., 2019).In particular, heat wave risk has shown to have strong regional variation under temperature increases (Harrington & Frame, 2022).There is a low occurrence of compound hot and dry conditions across the upper north and northeast of the North Island (Figures 8a-8c), although with an increasing trend, consistent with the upward trend in hot days found by Harrington and Frame (2022).Similarly, the strong compound event occurrence and increasing trends reflects the increased risk of hot days shown by Harrington and Frame (2022) in the middle reaches of the North Island and lower South Island (Figures 8g-8i).The low occurrence of compound hot and dry conditions in the upper north and northeast of the North Island suggests that soil moisture plays a less important role in compound events which results in a more stable land state during dry or growing months (Orth, 2021), particularly when viewed collectively with the weak to positive covariation in SM-T and dominance of latent heat fluxes throughout these typically wet regions (Figure 5).
Modest frequency of seesaw event occurrence (i.e., on average 17% of droughts (or one-fifth) are followed by pluvial activity the following month) is found in the present work, similar to that found globally by He and Sheffield (2020).This modest occurrence may in part reflect the approach of He and Sheffield (2020) in creating binary event occurrence for seesaw event detection, resulting in a loss of information because of the strict detection criteria.SPI-defined drought identifies a greater occurrence of seesaw events than SSMI-defined drought throughout the west coast of the North Island (winter half year) and lower South Island (summer half year) (Figure 9).This may be due to the relatively rapid time series variations in SPI, resulting in more frequent drought occurrence during the 2000s (west coast of North Island and lower South Island) compared to the more slowly-varying SSMI (Figure S8 in Supporting Information S1).Similar differences in the ability of the SPI versus SSMI to capture drought conditions were identified by Hao and AghaKouchak (2013) in California and North Carolina.In contrast, greater seesaw event occurrence in the SSMI time series for winter half year periods in the north-east of the South Island is here linked to the greater persistence of drought phases within the SSMI (vs.SPI) and the tendency for SSMI drought to develop through the winter half year period.In contrast, the lower persistence means SPI drought develops less frequently in the north-east of the South Island (Figure S8 in Supporting Information S1).These regional differences in the ability of the SSMI and SPI to capture drought conditions highlight the complicated dynamics and differences in land surface interactions, and the propagation of drought through the hydrological cycle.Investigating these seesaw event occurrences requires further exploration, particularly relating to an exploration of the temporal delay to capture seasonal cycles (He & Sheffield, 2020).
The rapid transition from dry to wet during seesaw events implies substantial and/or persistent precipitation events.Atmospheric rivers-narrow bands of intense water vapor transport (Newell et al., 1992)-are becoming increasingly associated with extreme precipitation and flooding across New Zealand (Prince et al., 2021;Shu et al., 2021).Furthermore, Reid et al. (2021) identified that eight (Christchurch and New Plymouth) and nine (Dunedin) of the top 10 rainfall events were associated with an atmospheric riverall locations identified herein as having a high occurrence of seesaw events (Figure 9).Further, Reid et al. (2021) identified a strong seasonal cycle in atmospheric river occurrence, with over 60% of events occurring during the warm period (January-April), with high seesaw event occurrence during the summer half year phase also revealed in the present work (Figure 9).The presence of strong seesaw event occurrence in similar regions to those that experience frequent atmospheric rivers (Prince et al., 2021;Reid et al., 2021) suggests the possibility of "drought buster" behavior associated with atmospheric rivers (Dettinger, 2013).Isolating the impacts of atmospheric rivers on the resultant land surface remains embedded within the uncertainty in drought quantification as it propagates through the hydrological cycle and its representation via differing accumulation periods and variable metrics (i.e., SPI vs. SSMI).While the present study indicates preliminary findings of seesaw event behavior for New Zealand, a more focused investigation is needed, including understanding the role atmospheric rivers play during this transitional phase.

Conclusion
For regions with physically diverse landscapes such as New Zealand, the increased resolution of current generation reanalysis data sets makes them an increasingly attractive option for climatological and hydrological analysis.The ability of the reanalysis data sets here to capture the seasonal cycle and residual anomalies highlights the strong utility reanalysis soil moisture products have, particularly considering the real value in soil moisture data are their temporal variability rather than their representation of absolute magnitudes.With existing soil moisture data across New Zealand often employed as an offline proxy metric, the ability of the current generation products to capture the soil moisture cycles and coupling regimes is a key benefit.The results here indicate good agreement in the representation of soil moisture in the three investigated reanalysis data sets for the period 1990-2018 (ERA5 Land, BARRA and ERA5; correlation range of 0.02).While trends in soil moisture are unable to be adequately captured by reanalysis products (mean of 0.08 mm yr −1 compared to 0.56 mm yr −1 in observations), the performance must be considered relative to the difficulties of comparing point based and grid cell data, while the agreement in seasonal cycle (correlations of 0.97-0.98)and ability to capture anomalies (correlations of 0.78-0.80) of the reanalysis data sets are promising.
Land-atmosphere coupling in a New Zealand context is poorly understood, with land variation often assumed to be driven by precipitation interactions.While clearly playing a significant role, the interaction of SM-T correlations reveals key areas of the country where soil moisture responds strongly to temperature variation, helping our understanding of compounding events in particular.Apparent shifts in land states is present across energy limited climates during growing seasons, with hypersensitive thresholds in particular across the upper North Island and east coast of both islands.Further work should be directed toward a more detailed investigation involving heat and energy fluxes to unravel the role soil moisture plays on temperature in a New Zealand context.Examining changes in drought (via soil moisture) behavior under a changing climate using these coupled products would be insightful, particularly when compared to the soil moisture proxy metrics traditionally employed in a New Zealand context.
For the first time, compound and seesaw events are examined in a New Zealand context, reflecting the shift in focus of the international research community.With regards to compound events, the present study highlights large portions of the country where compounding hot and dry conditions occur (maximum occurrence of 10% across the time period 1990-2018), including key agricultural areas where traditional energy-limited regimes appear to reveal a shift to a dry or transitional, water limited state.Taken collectively with the previously revealed SM-T relationship, the historical increase in these hot and dry conditions has important implications for the understanding of land responses to atmospheric changes under a warming climate.The current work indicates strong SM-P relationships across much of the country, which combined with the variable regional SM-T relationships, indicate that an average of 17% of droughts are followed by pluvial activity .With regional correspondence between atmospheric rivers and strong seesaw activity, a worthy new direction for atmospheric river research in New Zealand has also been identified.Collectively, the present work has provided a preliminary look at compounding and seesaw event behavior across New Zealand, revealing both areas to be a promising avenue for future research. .Depiction of seesaw events as the percentage of droughts which are followed by pluvials at each grid cell, for each reanalysis data set, using a one-month delay (for the period January 1990 to December 2018).Periods have also been broken into a winter half year (April-September) and summer half year (October-March).The far right column shows precipitation-based definitions of droughts and pluvials, while stippling indicates significance at the 5% level.

Figure 1 .
Figure 1.Observational site locations and grid cell locations from each reanalysis data set (boundaries as represented by coloration) used for statistical analysis.Elevation is represented by gray scale.

Figure 2 .
Figure 2. Annual cycles of averaged monthly mean soil moisture at each of the 12 observational locations and corresponding reanalysis data sets, for the time periods specified in TableS1in Supporting Information S1.
The upper North Island displays positive correlation between soil moisture and temperature (significant in BARRA), represented across all reanalysis data sets, while this positive correlation extends into the middle reaches of the North Island within the BARRA data set.The strongest evaporative fractions are similarly found across these upper north regions, particularly as expressed by the ERA5 and ERA5 Land data sets.Strong changes in regression slopes, indicating an increase in sensitivity at the critical soil moisture threshold, are present across the upper North Island in all data sets, with BARRA revealing similar sensitivity increases on the east coast of both islands.Northern locations (Hamilton and Kaitāia) indicate transitions into wet phases (SM > critical) during the growing season within the ERA5 and ERA5 Land data sets, while BARRA identifies the same locations as having a transitional regime (Figure S5 in Supporting Information S1).

Figure 3 .
Figure 3.Time series decomposition (STL) of monthly mean soil moisture integrated across all 12 sites (observations and associated grid cells for reanalysis data sets; note the differences in time periods at each site (Table S1 in Supporting Information S1)).Showing (a) original time series, (b) trend component, (c) seasonal component and (d) residual component.Note the different axis range in each panel.The blacked dotted line in panel (b) signifies the linear trend in observational data, significant at the 1% level.

Figure 4 .
Figure 4. Aggregated monthly soil moisture correlation, represented by SM-P (a-c) and SM-T (d-f), showing ERA5-Land (a, d), BARRA (b, e) and ERA5 (c, f).Aggregated total precipitation and mean temperature are represented as VCSN data, aggregated to each data sets native resolution.Period shown is growing seasons (November-March) from 1990 to 2018, with seasonality removed.Stippling indicates significance at the 5% level within individual grid cells.

Figure 6 .
Figure 6.Dry months (as defined by bottom third of ranked monthly mean soil moisture, data set specific) SM-P (a-c) and SM-T (d-f) correlation across reanalysis data sets (ERA5-Land (a, d); BARRA (b, e); ERA5 (c-f)) for the period January 1990 to December 2018.Total precipitation and mean temperature are represented as VCSN data, aggregated to each data sets native resolution.All data have had seasonality removed.Stippling indicates significance at the 5% level within individual grid cells.S represents mean spatial correlation.

Figure 7 .
Figure 7. Wet months (as defined by top third of ranked monthly mean soil moisture, data set specific) SM-P (a-c) and SM-T (d-f) correlation across reanalysis data sets (ERA5-Land (a, d); BARRA (b, e); ERA5 (c-f)) for the period January 1990 to December 2018.Total precipitation and mean temperature are represented as VCSN data, aggregated to each data sets native resolution.All data have had seasonality removed.Stippling indicates significance at the 5% level within individual grid cells.S represents mean spatial correlation.

Figure 8 .
Figure 8. Co-occurrence of hot and dry months across New Zealand for the period January 1990 to December 2018, as represented by reanalysis data sets (ERA5-Land (a, d, g); BARRA (b, e, h); ERA5 (c, f, i)).The top row (a-c) indicates the number of co-occurring hot and dry months which occur during the growing season (November-March).The middle row (d-f) signifies the total number of months where hot (=>1 of the STI) and dry (≤−1 of the SSMI) events co-occur, while the bottom row (g-i) indicates the trend (Mann-Kendall) of co-occurrence on the total number of months.Stippling indicates significance at the 5% level.

Figure 9
Figure9.Depiction of seesaw events as the percentage of droughts which are followed by pluvials at each grid cell, for each reanalysis data set, using a one-month delay (for the period January 1990 to December 2018).Periods have also been broken into a winter half year (April-September) and summer half year (October-March).The far right column shows precipitation-based definitions of droughts and pluvials, while stippling indicates significance at the 5% level.

Table 3
Statistics of Seasonal Trend Decomposition (Performed Using STL) of the Reanalysis Data Sets SoilMoisture and  Observational Soil Moisture (See Figure 3)