Impacts of droughts and floods on agricultural productivity in New Zealand as measured from space

We estimate the impact of precipitation extremes on the productivity of agricultural land parcels in New Zealand using satellite data. This type of post-disaster damage assessment aims to allow for the quantification of disaster damage when an on-the-ground assessment of damage is too costly or too difficult to conduct. It can also serve as a retroactive data collection tool for disaster loss databases where data collection did not happen at the time. We use satellite-derived observations of terrestrial vegetation (the enhanced vegetation index (EVI)) over the growing season, with data at the land parcel level identifying five land use types (annual and perennial crops, and three types of pasture), and with precipitation records, which we use to identify both excessively dry (drought) and excessively wet (flood) episodes. Using regression analyses, we examine whether these precipitation extremes had an observable impact on agricultural productivity. We find statistically significant declines in agricultural productivity that are associated with both droughts and floods. The average impact of these events is usually less than 1%, but the impacts are quite heterogeneous across years and across regions, with some parcels experiencing a much more significant decline in the EVI. We also identify several impact patterns related to the varying drought and flood vulnerability of the analysed land use types.


Introduction
In the last few decades, satellite-based observations of terrestrial phenomena have been increasingly used in the field of disaster risk management. Emergency responders, disaster risk managers, and scientists can use these data to obtain reliable information that can assist disaster risk reduction efforts and inform public policy (Voigt et al 2016). This progress has been driven by technological advances in remote sensing that have led to a massive increase in temporal, spatial, and spectral resolutions of satellite imagery and by improvements in the processing methods and computing power that are required to effectively interpret it (Notti et al 2018).
Satellite data are currently used for post-disaster damage assessment of various types of disasters caused by natural hazards, such as tropical cyclones, floods, droughts, earthquakes, landslides, and tsunamis. These assessments vary in their use of satellite sources (varying sensor types, resolutions, data accessibility), focus of the assessment (e.g. flood mapping, vegetation or agricultural impacts, buildings and infrastructure) and processing methods (Joyce et al 2009). The satellite sensors employed for damage assessment typically fall within three basic categories: optical, synthetic aperture radar (SAR) and light detection and ranging (LiDAR). The use of passive optical sensors for damage assessments is partially limited due to its reliance on the absence of atmospheric interferences such as clouds (especially relevant for floods and tropical storms), haze or smoke (Gillespie et al 2007). Furthermore, high-resolution imagery typically has lower temporal resolution, is not publicly available and can be very costly to acquire (Fayne et al 2017). Active sensors such as SAR or LiDAR do not rely on good weather conditions as they can penetrate cloud cover, but can be limited with respect to pixel size, interpretation difficulty and classification accuracy (Sanyal andLu 2004, Ge et al 2020).
This study employs optical satellite data to investigate the impact of droughts and floods on the agricultural sector, with a specific focus on plant productivity of agricultural land. If severe enough, drought and flood events can lead to significant adverse impacts on plant growth and productivity, with potential consequent effects on agricultural production, food security and knock-on shocks in the economy (Devereux 2007, Ding et al 2011, Allaire 2018.
The effects of droughts and floods on plant productivity can be approximated by satellite-based vegetation indices which are based on light reflectance from plant canopies. These indices effectively measure vegetation 'greenness' and can provide a consistent spatial and temporal representation of vegetation conditions (Xue and Su 2017). They have been extensively used for vegetation condition and crop productivity assessments across the globe in the past two decades (Zeng et al 2022, Radočaj et al 2023 and lend themselves well for impact assessments related to natural hazards.
Many studies employ vegetation indices such as the normalised difference vegetation index (NDVI), the vegetation condition index, the enhanced vegetation index (EVI) or their modifications for drought or flood assessments of vegetation cover in an affected area (Vicente-Serrano 2007, Jinghua et al 2012, Wang and Jing 2012, Zhang et al 2012, Song et al 2019, Rousta et al 2020, Orimoloye et al 2021, Ashraf et al 2022, but also more specifically for assessments of impacts on agricultural land or various crop types (Zhang et al 2012, Yu et al 2013, Ahmed et al 2017, Di et al 2018, Zambrano et al 2018, Chen et al 2019. Numerous studies have demonstrated that NDVI values are significantly correlated with yields of crop such as wheat (Gupta et al 1993, Labus et al 2002, Das et al 2013, Doraiswamy and Cook 2014, Hochheim and Barber 2014, Lopresti et al 2015, Guan et al 2019, sorghum (Potdar 1993), corn (Prasad et al 2006, Hayes andDecker 2007), rice (Quarmby et al 1993, Nuarsa et al 2011, Guan et al 2019, soybean (Prasad et al 2006), barley (Weissteiner and Kühbauch 2005), millet (Groten 2007) and tomato (Koller and Upadhyaya 2005). The EVI has been used more recently, as it is more responsive to variation in vegetation canopy structures than the NDVI and reduces residual atmospheric contamination and variable soil background reflectance (Huete et al 2002). Previous research has shown a significant correlation between the EVI and the yield of corn (Bolton and Friedl 2013) and rice (Son et al 2014), and a study considering an ensemble of ten globally significant crops found that the EVI shows better correlation with crop yields than the NDVI (Johnson 2016). Therefore, we use the EVI to proxy agricultural productivity in this study.
We focus on New Zealand, which, in principle, has institutions in place that could collect ground data on disaster impacts, but this is rarely done in practice. Unlike in some other high-income countries, most agricultural production in New Zealand is not insured, so there is little reason for insurance companies (private or state-owned) to collect such data, and the state itself does not do so, either. However, even in other temperate high-income countries, it is rare for insurance companies to share their data with the research community, so assessments of disaster damage (especially those based on less publicly accessible high resolution satellite imagery) are uncommon.
Satellite-based post-disaster damage assessment is a useful tool that should be added to the toolkit of post-disaster assessments. This applies in lower-income countries, where it may be the only available option, but also in high-income countries, where the remote sensing data can be more easily paired with other spatial data (due to higher data availability) to improve the assessment (Teeuw et al 2013). This pairing is often useful or necessary because of various challenges in interpreting the remote-sensing imagery. Thus, combining satellite imagery with ancillary data such as satellite-derived digital elevation models, static land use maps, and other datasets is commonly used to increase the quality of the assessment and is also utilised in this study.
Our aim is to quantify the impact of precipitation extremes (both excess and absence of precipitation) on productivity of agricultural land parcels in New Zealand as measured by the EVI. For this purpose, we analyse the correlation of the EVI over the growing seasons with excessive rainfall events and dry spells. This quantification is important, since in a largely agricultural exporting country such as New Zealand, agricultural productivity changes may lead to significant economic effects. Past drought and flood events in the last two decades indeed caused substantial economic damages (Kamber et al 2013, Frame et al 2020, Nixon et al 2021. We first identify the growing season of each land parcel in the country. We then calculate the peak EVI for each growing season for each land parcel as a measurement of agricultural productivity. We identify the land use type in each agricultural land parcel using the land use carbon analysis (LUCAS) land use map. We pair these data with the precipitation weather records in the form of the standardised precipitation index (SPI), which we use to identify both excessively dry (drought) and excessively wet (flood) episodes during the growing season in each parcel. Using regression analysis, we then examine whether these episodes had any observable impact on agricultural productivity (as measured by the peak EVI value during the growing season).
Overall, our study contributes in several ways. It is the first study that attempts to implement these approaches in the New Zealand context, with its rich land use data at a very small (parcel level) scale. We also use the long weather time-series available for New Zealand, which enable better identification of precipitation extremes (both droughts and floods). The availability of long time-series of weather data also enables us to focus on a longer time period, rather than on single events like the vast majority of the papers cited above. Furthermore, the detailed land use data enable us to separately estimate the impact of droughts and floods on five different land use categories. This is important, as it is very likely, as we indeed show, that land use is important in determining the impact of droughts and flooding.

Data
The dataset used to estimate the impact of droughts and floods on agricultural land productivity is constructed at the parcel level and provides information on the land use type, vegetation condition (the EVI), growing seasons, and precipitation from 2001 to 2017.

Parcel boundaries
To determine parcel boundaries, we use New Zealand primary parcels polygons, which are publicly available at Land Information New Zealand (LINZ n.d.). LINZ defines a primary parcel as 'a portion of land that is intended to be: owned by the Crown, except moveable marginal strips; held in fee simple (predominately private ownership); Maori freehold land or Maori customary land; public foreshore and seabed; the bed of a lake or river; road or railway; vested in a local authority.' The layer has a nominal accuracy of 0.1-1 m in urban areas and 1-100 m in rural areas. For this study, we excluded parcels classified as 'public foreshore and seabed' , 'bed of a lake or river' , and 'road or railway' .

Land use
To determine land use within each parcel, we used information from the LUCAS land use map (v. 008) developed by the New Zealand Ministry for the Environment (MfE n.d.). An example of the LUCAS land use map is shown in figure 1. The LUCAS land use map uses a range of remote sensing, environmental and land use data sources to distinguish 12 land use classes in New Zealand, including three forest classes (Pre-1990 natural forest, Pre-1990 planted forest, and Post-1989 forest), three classes of grassland (high-and low-producing, and with woody biomass), and two classes of cropland (annual and perennial). For some of these classes, subclasses are also defined (e.g. for high-and low-producing grassland, five sub-classes are distinguished: unknown, winter forage, grazed-dairy, grazed-non-dairy, and un-grazed). We do not use these sub-classes when producing our empirical estimates.
We remove from our dataset all parcels with non-agricultural uses (wetland, settlements, and 'other'), as defined in the LUCAS land use maps, and forests. We keep grassland (further referred to as pasture) and cropland. Land use information is available for 1 January 1990, 1 January 2008, 31 December 2012, and 31 December 2016. We consider that for a given year, the land use type corresponds to 2008 land use for the years up to 2010, the 2012 land use for the period from 2011 to 2013, and the 2016 land use for the period 2014 and after. Each primary parcel is associated with a land use type. If a parcel has multiple land uses, the parcel is sub-divided into as many parts as there is land use classes.

Vegetation indices
We use the EVI and consider its maximum value within the growing season to be representative of crop productivity in that season. We use the images from Terra MODIS Vegetation Indices (MOD13Q1) Version 6, which are provided every 16 days at a 250 m resolution (USGS 2022a). Cloudy and low-quality pixels are masked. An example of a map illustrating the varying EVI values in primary parcels is shown in figure 2. The EVI is calculated using the near-infrared, red and blue spectral bands reflectance ρ as:

Growing season
To determine the growing season, we use the MODIS global land cover dynamics product (MCD12Q2) Version 6 (USGS 2022b). This product records annually from 2001 to 2019 at a 500 m resolution and covers vegetation phenology metrics such as greenness increase and peak, senescence, and dormancy, which  characterise vegetation growth cycles. An algorithm is used to determine the timing of phenometrics which are derived from time series of MODIS adjusted surface reflectance EVI (NBAR-EVI2). The start of the growing season is characterised by the green-up onset (when EVI2 first crossed 15% of the segment EVI2 amplitude), and the end by the onset of dormancy (when EVI2 last crossed 15% of the segment EVI2 amplitude). Up to two vegetation cycles are detected to account for multi-cropping. For this study, which spans from 2001 to 2017, we calculated the average growing season for each agricultural land parcel, and linked the average growing seasons also with the first year of the sample, for which MCD12Q2 data were not available.

Climate
To estimate the occurrence of droughts and floods, we first consider the SPI. This measure is commonly used to represent the occurrence of precipitation over regions characterised by multiple climatic zones as it represents a standardised departure from the mean of a long-term trend (Jones and Hulme 1996). The SPI is calculated by first fitting a gamma probability density function to the frequency distribution of rainfall over a reference period (of at least 30 years), which is then used to determine the cumulative probability of a particular precipitation level for a chosen time scale and finally transformed into a normal distribution ∼N(0,1) (McKee et al 1993). SPI values are therefore expressed in terms of standard deviations from the median. Negative values imply below normal precipitations, while positive values indicate above normal rainfall. Using the SPI, it is possible to identify periods of droughts and floods (McKee et al 1993, Seiler et al 2002. Most alternatives to the SPI, like the standardized precipitation evapotranspiration index (SPEI), include other variables, such as evaporation, and are therefore not suitable as an index that is used to identify both droughts and floods.
A moderate drought starts when the SPI falls below 0 and ends when the index returns to a positive value after reaching a value of-1. A moderate flood is calculated similarly, but with a threshold of +1. The thresholds for severe and extreme droughts or floods are identified with thresholds of ±1.5 and ±2, respectively. Time scales of 1-48 months can be used to calculate the SPI depending on the responsivity of the sector considered (meteorologic, agricultural, hydrologic, and socio-economic).
For instance, longer time scales are more suitable for water resources management (e.g. for reservoirs), while shorter scales are better suited for detecting drought events affecting agriculture, especially for areas that are not irrigated (McKee and Edwards 1997). Statistical analyses of the impact of droughts on crops have used scales from 3 to 12 months (Yamoah et al 2000, Blanc and Strobl 2013, Hoffman et al 2017, Feng et al 2018. In this study, the SPI is calculated on a 2, 3 and 6 month time scales, using 1972-2005 as a reference period. The precipitation data used to calculate the SPI are extracted from the virtual climate station network (VCSN) database available from the National Institute of Water and Atmospheric Research. Daily precipitation data are available at a 5 km resolution. Daily minimum and maximum diurnal temperatures are also obtained from the VCSN, and based on these, we calculate the mean daily temperature as: Following McKee et al (1993), we also calculate the magnitude of a drought (DM) and the magnitude of a flood (FM). These are calculated as: where m is the starting month of the drought/flood and x is either the end of the drought/flood or the end of the growing season (whichever occurs first within the growing season that the drought/flood started). To account for the effect of the growing season duration, the magnitude is taken as a ratio of the growing season length (GSL). An illustration of the drought and flood duration and magnitude is shown in figure 3. We construct three different types of drought/flood magnitude indicators: (1) moderate or stronger drought/flood magnitude; (2) severe or stronger drought/flood magnitude; and (3) extreme drought/flood magnitude. These indicators are constructed so that they account for the total magnitude of droughts/floods in a parcel during a growing season, so they may include multiple drought/flood events. By construction, as severe and extreme events are included in the 'moderate or stronger' category, a 'moderate or stronger' category always has the highest magnitude out of the three magnitude indicators, followed by the 'severe or stronger' indicator.

Flood return interval
The SPI measures precipitation rather than physical flooding. Therefore, information on soil physical attributes is beneficial to more precisely estimate the effect of excess precipitation on plant productivity. To this end, we use the flood return interval from the New Zealand Land Records Information System (LRIS n.d.), which represents the probability of flooding. This delineation of land at risk of flooding provides a proxy for physical attributes such as slope, catchment area, surface permeability, etc, which play a role in how excessive rainfall would affect plants. However, as a full hydrological modelling of flooding throughout New Zealand is beyond the scope of this project, the flood return intervals are characterised by six classes described in Webb and Wilson (1995) 4 .
The map of the flood return interval in New Zealand is provided in figure 4 and shows that most of the land in New Zealand is not regularly affected by floods. To match flooding risk at the parcel level, we attribute the flood return interval class the most common to each parcel.

Sample description and summary statistics
The short name and description of the variables used in the analysis are shown in table 1. Summary statistics for those variables are provided in table 2 for crops and table 3 for pasture.

Method
A regression model was applied to identify the relationships between drought, flood, and temperature indicators and agricultural productivity (approximated by the EVI). The base regression specification was formulated as follows: where EVI max pt is the maximum observed peak EVI during the growing season in parcel (p) and growing season year (t), T min pt and T max pt are the average minimum and maximum diurnal temperature over the growing season in the same parcel/year combination; and their squared terms are also included. D is the drought magnitude indicator (as described above), and F is the flood magnitude indicator. µ is a parcel fixed-effect to account for parcel-specific time-invariant differences in plant productivity, and ε is the error term.
Regarding the selection of independent variables, average minimum and maximum temperatures are included as temperature is an important factor influencing plant growth. We include the squared terms for the temperature variables to identify potential non-linear relationships. Precipitation is another crucial determinant for plant productivity, thus we include the drought and flood variables to assess the effect of these precipitation extremes.
The drought and flood indicators are calculated using different thresholds that represent either moderate, severe, or extreme events. To separate the effect of excess water on crops located in flood prone area (versus land that is not prone to periodic flooding), we also added to these specifications an interaction term  Zone with flood return interval of category 2, 3, 4, 5 and 6 rf1 Zone with flood return interval of category 1 (Nil) between flood variable and the rf2345 flood-zone binary variable (delineating land with a flood return interval <60 years). We also considered the effect of irrigation and the expected lessening of the effect of drought on productivity that irrigation may generate. However, the data on irrigation are only available at the parcel level for the years 2017 and 2020 (MfE 2017, 2021; respectively) and at the subregional level for 2002 and 2017 (MfE 2019a, 2019b; respectively). None of these irrigation data provided reasonable results, most likely due to the sparse time and spatial coverage of the data and the significant changes in irrigation systems in the last two decades. We also estimated specifications representing non-linear effects of temperature using a fractional polynomial, but the results were very similar to the reported quadratic specifications, so the more parsimonious quadratic approach was preferred. Figure 5 shows a workflow chart representing the methodology adopted in the study. Firstly, agricultural land parcels are identified based on parcel boundaries and land use data. Subsequently, climate and flood-return data are assigned to each parcel. Once this data preparation process is completed, regression analysis is run.

Results
In this section, we describe the impact of droughts and floods on the productivity of annual and perennial crops, and three types of pasture. For each, we present the regression results, box plots that describe the impacts of droughts and floods over time and in each geographical region, and a series of maps that allows us to summarise these findings spatially.
Regression results are presented in table 4 for annual crops, table 5 for perennial crops, and for pasture in tables 6-8 (for low-producing pasture, pasture with woody biomass, and high-producing pasture, respectively). The first two columns in each table present results for moderate or above droughts and floods (without and with control for whether the parcels are flood-prone, respectively). The next two columns provide the same for severe or stronger events, and the last two for extreme events. In all regressions reported in tables 4-8, minimum and maximum average temperatures are also included in their linear and quadratic forms.
In tables 4 and 5, the results suggest that both droughts and floods have a negative impact on the productivity of both annual and perennial crops, as measured by the seasonal maximum of the EVI. The coefficients for the drought and flood variables are always negative and in almost all cases statistically significant. While the impacts of droughts on annual and perennial crops appear to be of comparable magnitude, floods appear to disproportionately impact annual crops, as the flood variable coefficient for annual crops is approximately 3× larger than the coefficient for perennial crops (in the case of moderate or stronger events). This suggests that annual crops may be significantly more vulnerable to floods than perennial crops. Columns 2, 4, and 6 show the results with flood impacts separately for areas with a higher potential for flooding (i.e. flood return interval <60 years). With moderate and stronger events, the effect of excess precipitation is not necessarily larger in flood prone areas than in those with less flood risk. However, with severe and stronger events, excess precipitation appears to have a disproportionate impact on flood-prone areas growing perennial crops (see table 5

, column [4]).
The results for the temperature variables show, consistently for crops, that as minimum temperatures increase, crop productivity increases with a strengthening effect as the squared term is also positive and significant (though the linear term is non-significant in the case of perennial crops in table 5). For mean daily maximum temperatures, the effect is concave and always statistically significant, with the beneficial effect of higher temperatures tapering off as the average maximum temperature increases.
Overall, the main conclusions from tables 4 and 5 are that indeed both droughts and floods damage annual and perennial crop productivity, and that annual crops appear to be significantly more vulnerable to floods than perennial crops. While the impact of flood-proneness on perennial crops is observable with severe and stronger events, the results for other event types and for annual crops are not consistent.
The box plots for annual and perennial crops are presented in figures 6 and 7, respectively. The drought and flood impacts are demonstrated by the specific drought/flood event effect, estimated using the coefficients from column (1) in tables 4 and 5. These impacts are calculated as a proportion (in percentage terms) of the long-term average maximum EVI within a given parcel. The box plots present the range of the parcel level impacts of droughts and floods between the 25th and 75th percentiles across all regions by year in the top panel and across all years by region in the bottom graph. The lines inside the boxes represent the  Notes: robust standard errors in parentheses; * * * p < 0.01, * * p < 0.05, * p < 0.1. Notes: robust standard errors in parentheses; * * * p < 0.01, * * p < 0.05, * p < 0.1.   The results in figure 6 show that floods disproportionately impacted annual crops during some years (e.g. 2008, 2011, and 2017) and in some regions (e.g. Gisborne). These spatial and temporal patterns of flood impacts are repeated for perennial crops, but the impacts seem to have been smaller. For droughts, the more adverse years (as measured by declines in the EVI) seem to have been 2015 and 2016 (for both perennial and (1) (2) (3) (4) (5)  Notes: robust standard errors in parentheses; * * * p < 0.01, * * p < 0.05, * p < 0.1. (1)  Notes: robust standard errors in parentheses; * * * p < 0.01, * * p < 0.05, * p < 0.1.
annual crops). The variation across regions appears to be less pronounced for perennial crops than it is for annual crops. The maps in figure 8 summarise the impact of each drought/flood event type that is presented in figures 6 and 7, averaged over the time period covered (2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017), and aggregated to the regional level. The patterns are quite different for droughts when compared to floods, and relatively similar between the crop types. For droughts, it is the eastern part of the South Island, and Taranaki and Northland in the North   Island that had the steepest declines in the EVI. For floods, the most impacted regions are Tasman in the South Island, and the eastern part of the North Island.
The estimated impacts of droughts and floods on pastureland are presented in tables 6-8, using the same format as for crops. The results show that similarly to the effects on crops, droughts and excess precipitation are both detrimental to pasture growth. The results also suggest that the vulnerability of pasture to either of these events varies depending on pasture type.
Regarding droughts, high-producing pasture appears to be approximately half as vulnerable to droughts as low-producing pasture and pasture with woody biomass (possibly because a significant fraction of it is likely to be irrigated). However, the situation is reversed in the case of floods. While low-producing pastureland is the least impacted, the flood coefficients for pasture with woody biomass are approximately twice as high, and the coefficients for high-producing pasture are roughly 3-4 times as high. This suggests that high-producing pasture, which is the most common as it is found in about twice as many parcels as the other two pasture types combined, is the most vulnerable to floods out of the three pasture types (the magnitude of the effect here is comparable to the one observed in annual crops).
We also find interesting effects of flood-proneness. While with crops, we have not consistently observed a disproportionate effect of floods on crop productivity in flood-prone areas, the flood impacts on low-producing pasture appear to be significantly stronger in flood-prone areas, and a less pronounced effect of the same kind is also observable in pasture with woody biomass. However, high-producing pasture in flood-prone areas appear to be less impacted.
As was the case for crops, we find little evidence that the set of more intense events (severe and extreme) is more damaging to pasture (as measured by the EVI) than the moderate droughts and floods. It appears that most of the damage associated with these events is already incurred due to moderate events, so that only little incremental damage is incurred as a result of severe and extreme events.
The effect of average minimum and maximum temperature on pasture differs somewhat from the one observed for crops. Here, the temperature effects on all three types of pasture are concave both for minimum and maximum temperatures, with the linear term always positive and significant, and the squared term always negative and significant.
The box plots representing the spatial and temporal effects of extreme precipitation on low-producing pasture are presented in figure 9. Again, we see the expected variation over time, with some years experiencing a bigger adverse impact associated with droughts and floods (e.g. 2015 for droughts and 2017 for floods). The bottom panel of figure 9, and the two upper maps in figure 10 present the spatial distribution of the impact of droughts and floods on low-producing pasture. The spatial pattern is relatively similar to the one we described for crops, with the eastern part of the South Island most vulnerable to droughts, and the eastern part of the North Island most vulnerable to floods.
Apart from the differences between the pasture types discussed above, many of the results for pasture with woody biomass and high-producing pasture (tables 7 and 8, respectively) are qualitatively similar to the results for low-producing pasture. While there are some differences in the spatial distribution of impacts of droughts and floods on the different pasture types, these differences are not very pronounced, and the general pattern (e.g. more flood impact in the eastern part of the North Island) still holds very clearly.

Discussion
Overall, we find statistically significant declines in agricultural productivity that are associated with both floods and droughts, and identified in all the five types of land use we examine-annual and perennial crops, and three types of pasture. This, in and of itself, is a finding that is closely related to other papers that have used remote sensing data to look at the impact of disasters; overwhelmingly, and unsurprisingly, these papers find a negative impact (e.g. Wang et al 2003, Jain et al 2010. We quantify this impact. The average impact of these events, as measured by peak EVI reduction, is usually less than 1%-this is a surprisingly moderate, though statistically significant, impact. However, this average hides a temporal and spatial heterogeneity of impacts, with some parcels experiencing a much more significant decline in the EVI. These results suggest that there is a considerable variation in the vulnerability of the five analysed land use types to these precipitation extremes, and even spatially across Aotearoa New Zealand. With droughts, we observe a relatively lower vulnerability of high-producing pasture compared to the other land use types. This could be explained by a higher use of irrigation on this agricultural land type. Regrettably, the available irrigation data were not available with sufficient temporal detail to provide meaningful results or a verification of this claim. In terms of flood impacts, annual crops and high-producing pasture appear to be the most vulnerable, with the flood coefficients being approximately 2-3 times higher than the coefficients for the other land types.
These results imply a more nuanced need to consider the risk of both floods and droughts, based on specific land uses and locations. Indeed, it suggests that as climate change modifies the risk profile of different agricultural areas around the country-in some cases increasing drought risk, and in others reducing it; and similarly for flooding-there is an additional reason to consider land use changes, as these land uses affect the vulnerability of agricultural activity to weather extremes.
Considering the effect of flood-proneness on the vulnerability to excess precipitation has not yielded consistent results. With crops, a disproportionate effect of high precipitation on flood-prone areas has only been observed for perennial crops in the case of severe or stronger events, but not for the remaining two event categories or annual crops. Regarding pasture, flood-proneness was shown to considerably increase vulnerability only with low-producing pastureland, with the opposite effect being identified for high-producing pastures. It seems that flood-proneness can proxy for both the actual frequency of the hazard, and the prevention, preparation and adaptation actions that farmers pursue, thus leading to inconsistent patterns in our findings.
In terms of the economic consequences of agricultural productivity declines, the adverse effects of droughts and floods we identified may not necessarily imply financial losses to farm businesses. Pourzand et al (2020), for example, has shown that droughts have not caused a dramatic decrease in the profitability of dairy farms over the past two decades in New Zealand. This is likely the case because, as we document here, agricultural production in New Zealand seems to be quite resilient to these weather shocks. At least in the case of droughts, it may be the case that high-producing pastureland, which constitutes the majority of agricultural land parcels in New Zealand, is disproportionately irrigated. This could explain its lower vulnerability to droughts we identified. It may also likely be the case that the drought and flood events affect market prices in ways that ameliorate their impact on farm revenue and profitability (given that the negative supply shock will lead to price increase), which seems to have been the case following a severe drought that occurred in 2013 (Kamber et al 2013). As such, supply-driven price increases may be obscuring some of the adverse wellbeing impacts caused by these climatic events, as the purchasing power of the consumers of agricultural products is reduced, and the consumers are the ones bearing the burden of these events' impacts. Kamal and Noy (2023) also show that while farmers' profitability may not have been affected adversely as much, farm debt does increase following drought events.
It is therefore important to point out the main caveat to our analysis. While we identify the decline in vegetation that is associated with the occurrence of precipitation extremes, we do not quantify the economic materiality of these impacts. Quite clearly, agriculture is an economic activity, so it is important to evaluate the economic implications of the vegetation declines we observe. This is, at this point, still impossible, as the data we have about economic outcomes (profit, price, sales, etc) are not spatially and temporally detailed enough to allow us to pursue the matching exercise we conduct between the remote-sensed data, the weather data, and the land-use data. A detailed and geo-referenced agricultural census, including farm balance sheet data, would enable such an investigation.
Lastly, there is also the question of external (or spatial) validity. Are our results relevant to other jurisdictions, where the climate and soil conditions are different, and the types of crops grown are not the same. Our main finding was a heterogeneity of the impact of droughts and floods on plant productivity. There is every reason to assume that heterogeneity will be even larger across more different climate conditions, and a more diverse set of crops. The relationships between these crops and global, national, and local market prices are also different, and raise, again, doubts that our results are valid for elsewhere. If anything, our results suggest that similar investigations should be taken elsewhere.

Conclusion
In the analysis presented here, we assess the impacts of droughts and floods on agricultural productivity in New Zealand using satellite data, most importantly satellite-based vegetation indices (the EVI). This type of post-disaster damage assessment allows for quantification of disaster damage when an on-the-ground assessment is too costly or too difficult to conduct, or when an assessment needs to be conducted retroactively. We show that drought and flood events were indeed associated with agricultural productivity declines and that these effects may be quantified using remote sensing data. We also identify spatial and temporal patterns in the heterogeneity of the events' impacts, as well as variations in vulnerability of the analysed agricultural land types.
We contribute to the existing literature on post-disaster agricultural damage assessment based on satellite data by presenting a first study applying these methods in New Zealand. Our second contribution to the literature consists of using detailed ancillary data, especially data on land use, in conjunction with the remote sensing imagery. These are primarily high-resolution land use maps and detailed weather records including both temperature and precipitation data. Thirdly, we analyse the drought and flood impacts in the long-run, employing long time-series of weather data, rather than focusing on a single event. This enables us to identify potentially persistent impact and vulnerability patterns which could prove useful in future efforts to mitigate the adverse consequences of these events.
Overall, we argue that remote sensing assessment is a useful tool that should be added to the toolkit of post-disaster assessments and the construction and collection of disaster loss databases. Better insight into disaster losses may consequently serve to inform policies that aim to mitigate disaster risk.

Data availability statement
The data cannot be made publicly available upon publication because they are owned by a third party and the terms of use prevent public distribution. The data that support the findings of this study are available upon reasonable request from the authors.