Modelling Climate Change Impacts on Location Suitability and Spatial Footprint of Apple and Kiwifruit

Abstract

The threats and opportunities faced by primary production industries from future climate changes can be adequately prepared for only with the guidance of model projections that can assist the development of robust policy and climate adaptation plans by governments and industries. We used continuous suitability models capable of reflecting incremental changes to project the suitability of locations across New Zealand for cultivating apple and kiwifruit in the mid- and late-century. These projections used future weather data from climate model simulations for two contrasting greenhouse gas (GHG) pathways: stringent GHG mitigation and unabated GHG emissions. To improve the suitability of the modelled temperature data, specifically for use with biologically driven, crop suitability models, we developed new bias-variance adjustments that preserved climate change signals within the data. Preliminary projections of land use across a range of alternative primary industries were obtained from a multinomial logit model incorporating continuous suitability scores as predictors. We refined the preliminary land-use projections by providing them as inputs into a simulation model of land use incorporating other drivers and constraints. This methodology provides a means for projecting future land use and the spatial footprints of primary industries, based on biological and econometric considerations, under different modelled climate change scenarios.

1. Introduction

New Zealand, whose export revenue is strongly contributed to by horticultural production, is expected to be significantly impacted by climate change [1]. Climate change will affect all horticultural crops in New Zealand, with the major factor being temperature changes affecting the development rates of key phenology stages and causing direct damage to plant tissues [2,3]. Changes in rainfall patterns [4] and incidences of pests and diseases [5,6,7] are also anticipated. Horticulture is distributed across niche locations around New Zealand, which has a diverse range of climatic systems that will be differentially affected by climate change [8]. Climate change could threaten some horticultural industries in some regions, while offering opportunities in others [2,9,10].

A number of studies exist in the broader area of future climate change and address a range of climate-change-related issues and crops. Recent work includes projection of how expansion of cropping and urban areas worldwide would affect land-use competition and impact on the climate mitigation potential of natural climate solutions [11]. Xu and Yao [12] combined linear programming to optimise areas used for different land use with a cellular automata algorithm for the spatial arrangement of land use patches. This approach, which incorporated constraints related to climate change adaptation (suitability, resource availability and social adaptation) and climate change mitigation (carbon sequestration), identified that, due to trade-offs, climate change adaptation and mitigation must be considered jointly for sustainability under climate change.

Projections of climate change impacts under the A1B socio-economic (medium-emissions) scenario indicated changing suitability in a Mediterranean region of Spain would reduce yields considerably more for sunflowers than wheat, with soil type governing the size of the impacts [13]. Gardner et al. [14] used a mechanistic model of crop growth to project future suitability in terms of yield for a range of arable crops in the south west UK. Gebrehiwot et al. [15] used agent-based modelling to incorporate human and natural system interactions to identify specific land cover transitions to fulfil community food needs in areas with low income and food nutrition. Although the model was run with current climate data, it has the potential to be used with climate data under different scenarios.

A global-level study that projected future land use for energy crops under two different Representative Concentration Pathways (RCPs) found that significant areas of land suitable for energy crops would overlap with forested or protected areas, or with land suitable for food crops, and this highlighted potential hotspots for land competition [16]. Li et al. [17] presented a system, incorporating remote sensing and artificial neural networks (ANN), that coupled human and natural effects to simulate global-level conversions between a number of land uses under future climates. Truong et al. [18] used agent-based modelling to predict farmer decision making in the conversion of rice–shrimp systems to shrimp systems, taking into account land suitability and convertibility, neighbouring land use and the benefits of the system.

Projections of climate change impacts can stimulate adaptation responses, both to avoid damage and to exploit beneficial opportunities [19]. For horticulture, adaptation strategies have been categorised into a three-pronged approach: (i) tactical adaptations which involve modifying production practices within the current system; (ii) strategic adaptations which involve a substantive modification to the current production system; and (iii) transformational adaptations which change the spatial footprint of the production system [2]. Adaptation decisions, especially ones with long-term consequences, have impacts beyond the orchard, with socio-economic effects on surrounding communities influencing labour, industries and the environment [10]. Thus, a well-informed and planned policy by the government, regional authorities and horticultural industry bodies is needed to underpin climate change adaptation, which in turn requires evidential material, and this can be provided by robust models [20,21].

Tait et al. [22] performed a study on location suitability in New Zealand for ‘Hayward’ kiwifruit under climate change. However, that work was expressed in terms of mean seasonal temperatures and did not consider plant phenology, plant physiology or soil properties, and also did not attempt to project how land uses might change between primary industries. Thus, this paper has three objectives: (i) develop projections of the future suitability of locations across New Zealand for growing apples and kiwifruit in the mid-century and late-century, based on considerations of plant phenology, plant biology and soil and terrain requirements; (ii) use the suitability projections to obtain preliminary projections of land use across a range of competing primary industries including apples and kiwifruit; (iii) produce refined spatial footprint projections from the preliminary land-use projections using an econometric model that incorporates other drivers and constraints on land use to spatially allocate land-use changes across the country. The schema in Figure 1 describes the framework of this paper.

Objectives (i) and (ii) use models developed in a companion paper [23] and these are discussed in the Background/theory section. The econometric model used for Objective (iii) was the “Land Use in Rural New Zealand” (LURNZ) model [24]. Additionally, we describe corrections that we made to the climate model data [25] in order to improve their suitability for our purposes.

2. Methods

Projecting climate change impacts on crop location suitability requires climate model data that extend from a historic period into the future, to provide inputs to crop suitability models. For consistency, impacts must be gauged by comparing suitability scores that are calculated for past and future periods using data from the same climate simulation runs.

As part of good practice, crop models used for investigating climate change impacts should be evaluated for accuracy using historical observed data [26]. Additionally, for meaningful estimates of future change, model calculations using climate model data from the historic period must closely resemble calculations using historical observed data from the same past period. If this is not the case, then it is imperative to make corrections to the climate model data in order to achieve this [26]. If bias-corrected data are to be used to guide real-world adaptation decisions, there is an imperative that they should be plausible and defensible [27].

2.1. Models

2.1.1. Continuous Suitability Models

We used sliding-scale suitability models developed and calibrated by Vetharaniam et al. [23] to reproduce accurate maps of the suitability of locations across New Zealand for growing apple and kiwifruit. These models use land, soil and weather data to predict suitability on a continuous scale from 0 (totally unsuitable) to 1 (extremely suitable), based on phenological and physiological considerations, and allow a combined assessment of all growing criteria along a continuous scale. This provides a more nuanced and holistic consideration than discrete suitability approaches, whether binary (suitable/unsuitable e.g., [9]) or considering several levels of suitability, e.g., [28].

2.1.2. Multinomial Logit Model of Land Use

We used a multinomial logit model that was estimated using industry data on orchard locations, with location suitability being included as a predictor for apples and kiwifruit [23]. The general approach is to model the choice of land use as a function of location-depended independent variables, and estimation is performed by maximum likelihood methods [29]. Binomial, nested and multinomial logit approaches have been used extensively to investigate, among other examples, decisions associated with tropical deforestation, environmental policies and urban development [30,31,32].

2.1.3. Land Use in Rural New Zealand Model

The LURNZ econometric model is a partial-equilibrium, national-scale, spatial model designed to consider the implications of environmental policies on future land use, production and greenhouse gas emissions [24]. It includes all private rural land in New Zealand and produces annual land-use maps at a 25 ha resolution. The LURNZ model was developed to simulate spatial changes in dairy farming, sheep–beef farming, plantation forestry, horticulture and unproductive scrub in response to changes in economic incentives.

The foundation of the LURNZ model is provided by econometrically estimated models that establish the relationship between observed drivers of land use and land-use outcomes [29,33,34]. The revealed preference approach requires relatively few assumptions about farmers’ objectives and decision processes: the LURNZ model results are largely driven by how land use has responded to its main drivers in the past. The model’s underlying datasets and processes have been validated [35], and its results are consistent with data and trends at the national scale, including New Zealand’s Greenhouse Gas Inventory [36]. Underlying assumptions and processes involved in our application are documented in Appendix B of the report by the Parliamentary Commissioner for the Environment (PCE) [37]. The LURNZ model is informed by climate impacts on suitability via the logit model outputs.

2.2. Software

Suitability modelling was carried out in version 6.2.0 of the modelling environment GNU Octave (https://octave.org accessed on 2 March 2021). The LURNZ model [24] written in Matlab R2020b (The MathWorks, Inc., Natick, MA, USA) was used for land-use simulations.

2.3. Data

We used the land, soil and Virtual Climate station Network (VCSN) databases described by Vetharaniam et al. [23]. Other data, such as a forest-age map, were also used to constrain simulated harvest and deforestation decisions when considering competition in multiple land uses in the LURNZ model as described by Anastasiadis et al. [35] and Kerr et al. [24].

The New Zealand National Institute of Water and Atmospheric Research (NIWA) supplied modelled future climate data derived from the NIWA’s high resolution Regional Climate Model (RCM) run on a domain encompassing all of New Zealand, using boundary conditions obtained from six CMIP5 (Coupled Model Intercomparison Project Phase 5) global climate models (GCMs): BCC-CSM1.1, CESM1-CAM5, GFDL-CM3, GISS-EL-R, HadGEM2-ES and NorESM1-M. Each GCM had been run under different RCPs. We focussed on two extremes: RCP 2.6 corresponds to a high greenhouse gas (GHG) mitigation pathway and RCP 8.5 to unabated GHG emission, with respective total radiative forcings of 2.6 and 8.5 W m⁻² at 2100 relative to 1750 [25].

For each RCP, we used simulated daily data from 1971 to 2099 (RCP datasets). RCM simulations from 1971 to 2005 were considered historical simulations and are referred to as ‘RCP Past’. For each CMIP5 model, both RCPs shared the same RCP Past dataset. Simulations from 2006 onwards were considered future simulations (‘RCP Future’). The RCP datasets were downscaled to the 5 km × 5 km VCSN grid and bias corrected in order to reflect New Zealand climates, although the year-to-year variability would be different [25].

When the calculations required hourly temperature values, we assumed the temperature would vary sinusoidally over a 24 h period in order to obtain hourly profiles from the maximum and minimum temperature.

2.4. Correction to RCP Datasets

2.4.1. Biases

We found biases in the RCP-Past-based suitability scores compared with the corresponding VCSN-based scores, with poor resemblance between the corresponding maps. These biases worsened for 2006 to 2016 (referred to as the ‘Contemporary period’). A large contribution to these biases was due to differences between the VCSN and RCP Past datasets in their means and variances at different time scales (see Appendix A and Figure A1).

Adjusting for differences in the means and variances of temperatures is important for suitability criteria that relate to either temperature extremes (e.g., frost or sunburn risk) or to the mean temperatures of individual or consecutive months, or seasons. By having lower variances in monthly mean maximum and mean minimum temperatures, the RCP datasets reflect a climate with less difference between cold and warm years than the New Zealand climate reflected by the VCSN data. Differences in variances for daily temperatures within each month would also affect the frequency of temperature extremes.

The consequences of statistical differences between the VCSN and RCP Past datasets are highlighted by a case study on frost risk in Alexandra, a significant apple growing area in New Zealand. This analysis indicated that the RCP Past data under predicted the number of September frosts events below −2° by 70–80%, albeit varying with the GCM (see Supplementary Materials).

2.4.2. Correction to RCP Datasets to Align Monthly Temperature Statistics

Corrections were made to align the RCP datasets and VCSN datasets with the RCP Past period (1972 to 2005) and to propagate the correction adjustments to RCP future data in a consistent way. To distinguish the new climate-model datasets, we refer to them as “SLM RCP” datasets.

Correcting RCP Past Temperature Datasets

Corrections to the RCP datasets were applied separately for each month of the year based on the statistics presented in Figure A1. The procedure is outlined below, with Steps 1 and 2 further described in Figure 2.

• Adjust for differences between each RCP dataset and the VCSN dataset in the mean variance of daily maximum and minimum temperatures for the month:

  $$T_i\to (T_i-{\overline{T}}_i)·{\sigma }_{d,V }/{\sigma }_{d,R }+{\overline{T}}_i,\mathrm{for}\mathrm{each}\mathrm{year},i=1972,\dots ,2005$$

  (1)

  (a)

  ${\overline{T}}_i$ is the mean daily maximum or minimum temperature for the month in year $i;$

  (b)

  ${\sigma }_{d,R}$ is the standard deviation of the daily (suffix d) RCP (suffix R) maximum or minimum temperature for the month, calculated over the 1972–2005 period using the following formula, with $N$ being the number of years and $n_i$ being the number of days in the month for year $i.$

  $${\sigma }_{d,R}^2=\frac{1}{N}{{\displaystyle \sum }}_{i=1}^N{{\displaystyle \sum }}_{j=1}^{n_i}\frac{{\left(T_i^j-\overline{T_i} \right)}^2}{n_i-1}$$

  (2)

  (c)

  ${\sigma }_{d,V}$ is the standard deviation of the daily VCSN (suffix V) temperatures for the month over the same period, calculated by applying an equation analogous to Equation (2) to the VCSN data.

• Adjust for the difference between each RCP dataset and the VCSN dataset in the variation of the monthly mean maximum or minimum temperature around the corresponding monthly mean for the period:

  $$\begin{gathered} T_i\to T_i-\mathsf{\Delta }{T}_i \\ \mathsf{\Delta }{T}_i=\left({\overline{T}}_i-\overline{T_i^s}\right)\left(1-{\sigma }_{m,V}/{\sigma }_{m,R}\right) \end{gathered}$$

  (3)

  where

  (a)

  $\overline{T_i^s}$ is the value at year $i$ of the spline smooth of the entire 1972–2099 data series.

  (b)

  ${\sigma }_{m,R}$ is the standard deviation of the RCP monthly (suffix m) mean of individual years around ${\overline{T}}_i$, calculated as:

  $${\sigma }_{m,R}^2=\frac{1}{N-1}{\displaystyle \sum }_{i=1}^N{\left(\overline{T_i}-\overline{T}\right)}^2$$

  (4)

  where $\overline{T}$ is the mean of ${\overline{T}}_i$ for the 1972–2005 period.

  (c)

  ${\sigma }_{m,v}$ is the standard deviation of the VCSN mean (maximum or minimum) temperature for the month over individual years, calculated around the VCSN mean monthly (maximum or minimum) temperature over the 1972–2005 period using an analogous equation to Equation (2).

• Correct for bias with respect to monthly means for the RCP maximum and minimum temperature.

$$T_i\to T_i-\left(\overline{T}-\overline{V}\right)$$

(5)

where $\overline{T}$ and $\overline{V}$ are the RCP and VCSN mean (maximum or minimum) temperature for the month over the entire 1972–2005 period.

Additionally, we corrected a data anomaly where, for some locations on some days, the maximum temperature was lower than the minimum. We swapped values in such instances.

Correcting RCP Future Datasets

The corrections in Steps 1 to 3 above were applied to the RCP future temperature data using the same standard deviation ratios calculated for the Past period. Since these ratios are scaling factors, the correction preserves the relative increases or decreases in variation projected by the climate models. The use of spline-smoothed values in Step 2 ensures that the extension of the correction to the future is consistent with corrections to the RCP Past data.

2.5. Estimation of Uncertainty in Projections

While the corrections we performed improved the bias for the Contemporary period, a significant bias remained (see Section 3.1), and this is reflective of the natural variations in the VCSN data about long-term trajectories that can be caused by a number of factors [38]. We exploited this residual bias to estimate uncertainties in our future projections, since the variability between the six GCM-forced datasets was small within an RCP.

For each crop, the root mean square error (RMSE) across locations was calculated for SLM RCP-based 2006–2016 scores with respect to the corresponding VCSN-based scores. SLM RCP-based scores were first separated into histogram bins of width 0.1, and an RMSE was calculated for each bin. We chose a 30-year period average for future projections to reduce the effect of volatility in weather patterns. Accordingly, we estimated the margin of error by discounting the RMSE calculated for the 11-year Contemporary period by $\sqrt{1/3}$.

2.6. Future Location Suitability

Suitability scores for an early-mid-century (2028–2058) period and a mid-late-century period (2068–2098) were calculated from the SLM RCP datasets, averaging the 31-year scores. For each RCP, the scores were calculated separately for the six SLM RCP datasets and then averaged. For each suitability model, this gave one Past-period prediction and two future periods projections per RCP. The projected future suitability scores were mapped across New Zealand, as were changes from the Past Period suitability scores.

2.7. Land-Use Modelling

2.7.1. Integration of Logit Model Results with the LURNZ Model

The multinomial logit model (estimated from RCP Past location suitability maps [23]) was used with the future location suitability maps to project land-use change for the mid-century and late-century under RCP 2.6 and 8.5. Climate change impacts for other land uses were not directly modelled but land use changes in kiwifruit and apples were offset by changes in other land uses so that the total modelled land area did not change.

The projected land-use changes from the multinomial logit model and the predicted land-use probabilities were used as inputs into LURNZ model runs. The LURNZ model combines these inputs with other land-use changes representing the impact of economic drivers and applies various constraints on future land-use change. Therefore, the simulation outcomes from the LURNZ model will not exactly match the projections of the econometric model.

2.7.2. Climate-Driven Land-Use Change

The LURNZ model integrates information from current and future suitability layers during the execution of the desired climate pathway. In the LURNZ model, the RCP Past climate is assumed to be current and applicable until 2027. The model does not therefore project any climate-driven land-use change in this period. Mid-century climate is assumed to apply to the period 2028–2062. The LURNZ model determines the overall land-use change projected for this period and attempts to allocate it in a linear manner in annual steps over the period. The end-of-century climate is assumed to apply to the period 2063–2098. The LURNZ model determines the additional land-use change for the period and attempts to allocate it linearly over the period. The LURNZ model’s simulation horizon ends in 2075, so land use will not have fully adjusted to the end-of-century climate by the model’s final simulation year.

2.7.3. Economic and Policy Context

Economic and policy parameters affect simulation outcomes for dairy farming, sheep–beef farming, plantation forestry and scrub (potentially productive land not being used for production). The parameters applied to the LURNZ scenarios in this project were based on a requirement for the land sector to reach a 20% emissions reduction by 2050, and other sectors to meet net zero emissions (integrated Model Run 7 in the study by the PCE [37]). The carbon price applied to the land sector increased along a sigmoid function to NZD 48 (per tonne CO₂ equivalent) by 2075, the end of the simulation period. As a larger share of mitigation is performed by other sectors, there is a relatively small disruption to the land sector compared with the other policy scenarios considered. Therefore, this scenario is considered the most consistent with the assumption of no exogenously driven growth in horticulture, kiwifruit and apple land areas.

Consistent with assumptions in the reports of both the Productivity Commission [39] and the PCE [37], we constrained the expansion of dairy farming in the LURNZ simulations such that no new land could be converted to dairying beyond 2025. This constraint reflects an anticipation of regional councils setting water quality limits in their regions in compliance with the National Policy Statement for Freshwater Management [40].

2.7.4. Implementation

The algorithm governing the geographic allocation of simulated land-use change in the LURNZ model has three key features: (i) it deals with net annual changes only; (ii) it prioritises changes in higher value land uses; and (iii) if a land use is expanding (decreasing), cells with the highest (lowest) probability for that use will be converted first. Here, we amend the original allocation algorithm documented in detail in Anastasiadis et al. [35] to deal with changes in kiwifruit and apple land use, as described below.

In each simulation year, changes to kiwifruit are allocated first, and changes to apples are allocated next. Then, the algorithm moves on to other horticulture, dairy farming, sheep and beef farming, forestry and finally scrubs. The order reflects the relative value of the land uses. If the kiwifruit land area increases, cells in other land uses with the highest kiwifruit land-use probability are converted to kiwifruit use. If kiwifruit land decreases, kiwifruit cells with the lowest kiwifruit land-use probability are assumed to convert to other horticulture lands. The allocation of apples proceeds similarly, with the exception that kiwifruit land is not available to convert to apples. This constraint is applied to avoid infinite loops and serves to restrict conversion from a higher to lower value use.

Land-use conversions must adhere to some additional constraints. For example, the deforestation of a plantation forestry cell is not allowed unless the forest is of a harvestable age. Pre-1990 forests are not allowed to change land use if a positive carbon price applies.

The algorithm deals only with changes in land use each year, thereby minimising the number of cells that change use. Reshuffling across land uses does not happen. This feature, a reflection of both unobservable factors and costs associated with land conversions, leads to patterns of land-use change that match the observed short- and medium-term changes reasonably well [35].

We extended the LURNZ algorithm to improve its suitability for long-term simulations in the context of climate change, by introducing an additional step to simulate conversions out of kiwifruit and apples. For each of these two land uses, we assumed that the loss of land area within a region is proportional to the change in mean land-use probability of the land in the given use. That is, a 50% fall in mean kiwifruit land-use probability on current kiwifruit land would lead to a projected loss of half of the region’s kiwifruit land. The land area lost is then reallocated elsewhere by the standard algorithm in the LURNZ model. Essentially, this leads to a reallocation of kiwifruit and apple land areas away from regions with falling suitability to areas with the highest suitability.

Without this addition, the algorithm would be purely driven by land use, and would not allow cells to convert out of a crop in response to decreased suitability if the national area of that crop were increasing.

3. Results

3.1. RCP Data Corrections

The bias adjustments generally had very small impacts on climate change signals for locations with elevations below 500 m, which are more likely than more elevated locations to be used for horticulture (see Appendix A and

Table A1. Impact on climate change signals of the bias adjustments, expressed for each representative concentration pathway (RCP) as the difference in change signals between the ensemble of six adjusted “SLM RCP” datasets and the ensemble of six original RCP datasets. Change signals were calculated for 2031–2050 and 2080–2099, using a reference period of 1986–2005. Change signals for temperature were calculated as differences between period means, while change signals for temperature variances were calculated as the ratio of variances. Change signals were calculated for each location. Statistics are for all grid locations, mapped across the country. $P_1$ and $P_{99}$ are the first and 99th percentile values, and LE and UE are lower and upper extrema.

	Elevations < 500 m				All Elevations
Period	2031–2050		2080–2099		2031–2050				2080–2099
	$P_1$	$P_{99}$	$P_1$	$P_{99}$	$P_1$	$P_{99}$	LE	UE	$P_1$	$P_{99}$	LE	UE
Annual minimum temperature signal impact (°C)
RCP 2.6	−0.02	0.02	−0.02	0.01	−0.04	0.02	−0.17	0.03	−0.04	0.02	−0.20	0.02
RCP 4.5	−0.03	0.00	−0.03	0.01	−0.06	0.00	−0.19	0.01	−0.05	0.01	−0.17	0.01
RCP 6.0	−0.03	0.02	−0.02	0.02	−0.04	0.02	−0.19	0.02	−0.04	0.02	−0.21	0.02
RCP 8.5	−0.02	0.02	−0.02	0.01	−0.04	0.02	−0.20	0.03	−0.05	0.01	−0.21	0.01
Mean annual maximum temperature signal impact (°C)
RCP 2.6	−0.01	0.02	−0.01	0.02	−0.01	0.06	−0.02	0.22	−0.01	0.06	−0.03	0.19
RCP 4.5	−0.02	0.01	−0.01	0.01	−0.03	0.02	−0.06	0.13	−0.01	0.04	−0.03	0.16
RCP 6.0	−0.01	0.01	−0.01	0.02	−0.01	0.04	−0.02	0.18	−0.01	0.06	−0.03	0.21
RCP 8.5	−0.02	0.02	−0.01	0.02	−0.02	0.05	−0.03	0.20	−0.01	0.05	−0.03	0.18
Mean annual mean temperature signal impact (°C)
RCP 2.6	−0.01	0.01	−0.01	0.01	−0.01	0.02	−0.02	0.04	−0.01	0.02	−0.03	0.05
RCP 4.5	−0.02	0.00	−0.01	0.00	−0.03	0.00	−0.06	0.01	−0.02	0.01	−0.03	0.02
RCP 6.0	−0.02	0.01	−0.01	0.01	−0.02	0.01	−0.04	0.03	−0.01	0.02	−0.03	0.03
RCP 8.5	−0.01	0.01	−0.01	0.01	−0.01	0.02	−0.03	0.05	−0.01	0.01	−0.03	0.01
Annual minimum temperature variance ratio signal impact (1)
RCP 2.6	−0.02	0.00	−0.02	0.01	−0.02	0.00	−0.05	0.01	−0.02	0.01	−0.04	0.01
RCP 4.5	−0.02	0.01	−0.02	0.01	−0.02	0.01	−0.03	0.02	−0.02	0.01	−0.04	0.03
RCP 6.0	−0.02	0.01	−0.03	0.02	−0.02	0.01	−0.05	0.01	−0.03	0.02	−0.05	0.03
RCP8.5	−0.03	0.01	−0.04	0.04	−0.02	0.01	−0.05	0.02	−0.04	0.04	−0.07	0.06
Annual maximum temperature variance ratio signal impact (1)
RCP 2.6	−0.02	0.00	−0.03	0.00	−0.04	0.00	−0.08	0.00	−0.03	0.00	−0.07	0.01
RCP 4.5	−0.02	0.01	−0.03	0.01	−0.03	0.01	−0.08	0.02	−0.05	0.01	−0.09	0.02
RCP 6.0	−0.02	0.00	−0.04	0.01	−0.03	0.00	−0.08	0.01	−0.06	0.02	−0.12	0.03
RCP8.5	−0.02	0.01	−0.05	0.02	−0.04	0.01	−0.10	0.02	−0.07	0.03	−0.15	0.05

). Variance corrections could affect the change signals for variance in annual temperatures since adjustment factors vary with the month of the year, and climate change signals can differ between months or seasons. However, monthly change signals are unaffected. Since adjustments to means are additive, they will have no impact on the calculated change signals, apart from the difference between spline fit to mean trajectories and the window averaging used to change signals.

The new SLM RCP datasets had a better statistical agreement with the VCSN data (see Supplementary Materials). In general, the agreement between the RCM Past and VCSN suitability scores improved, as shown in the one-to-one graphs in Figure 3 and in the Supplementary Materials.

3.2. Suitability Projections

The sliding scale system can be categorised to facilitate discussion, and for convenience we (arbitrarily) refer to suitability scores in the 0.9 to 1.0 range as “excellent”, 0.8 to 0.9 as “very good”, 0.7 to 0.8 as “good” and 0.6 to 0.7 as “acceptable”.

3.2.1. Apple

Apple RCP 2.6

Under RCP 2.6, most changes to suitability are projected to occur by the mid-century, with the patterns of change continuing at a slower pace into the late-century. The biggest declines in location suitability will occur across northern parts and around the easternmost point of the North (Figure 4). However, these are not currently significant apple-growing areas. Smaller declines in suitability are projected to occur for other North Island areas that are outside the central North Island. The latter region is expected to experience slight increases in suitability, as is the South Island. Areas of land with acceptable or higher suitability are projected to have a modest increase through to the late-century, although the area of land in the excellent range will decrease compared with the mid-century (

Table A2. Apples: land area in the historic period falling into different cultivation suitability ranges, forecast change for the mid- and late-century under RCP 2.6 and best and worst cases.

Apple SLM RCP 2.6	Historic (km2) 1972–2004	Area Change from Historic (km2) 2028–2058			Area Change from Historic (km2) 2068–2098
Suitability Range		Forecast	Best Case	Worst Case	Forecast	Best Case	Worst Case
0–0.1	14,891	−1498	−3175	354	−1603	−3176	83
0.1–0.2	4286	−828	−1980	−673	−909	−2126	−648
0.2–0.3	4972	−1266	−617	−134	−1404	−711	−358
0.3–0.4	8321	−2494	−4201	−367	−2711	−4302	−479
0.4–0.5	12,457	−2057	−4464	393	−1947	−4767	582
0.5–0.6	22,444	−2387	−6174	855	−2340	−5949	1101
0.6–0.7	31,668	3815	916	4969	4474	1137	5157
0.7–0.8	32,230	948	3752	−2846	507	3689	−2872
0.8–0.9	31,947	4558	5270	2744	5380	6011	3525
0.9–1.0	20,509	1209	10,673	−5295	553	10,194	−6091

in Appendix B).

Projection uncertainty allows that, under RCP 2.6, land with very good or excellent suitability for apples could increase in area by 30% (best case) or decrease by 5% (worst case) and thus the outlook favours an increase in higher suitability land (see Table A2 in Appendix B).

Apple RCP 8.5

Under RCP 8.5, the projected changes in the location suitability for apples tend to be larger than under RCP 2.6, which is not unexpected since the latter is a pathway to mitigate climate change. The mid-century suitability maps under RCP 8.5 (Figure 5) are similar to the late-century maps for RCP 2.6 (Figure 4). Under RCP 8.5, location suitability for apples declined severely across northern parts and around the easternmost point of the North Island. Significant declines are expected south of these areas and in some eastern coastal areas (Figure 5) which could disadvantage a small number of established orchards. Smaller declines in suitability are projected to occur across much of the rest of the North Island, except for central and elevated areas which are projected to improve in suitability and generally have at least a good location suitability. By the late-century, low to moderate increases in suitability are projected for the South Island except for declines in a north-east corner and scattered areas about one third down the eastern coast (Figure 5). This pattern of change could shift the footprint of apples further south.

Under RCP 8.5, a very small increase in the area of land with excellent suitability is projected by mid-century, and significant increases in land with suitability scores in the acceptable to very good ranges. However, a loss of 50% of excellent suitability land (10,000 km²) is projected to occur by the late-century, although this will be accompanied by about a 14,000 km² increase in land falling into the slightly lower ‘very good’ range (see

Table A3. Apples: land area in the historic period falling into different cultivation suitability ranges, forecast change for the mid- and late-century under RCP 8.5 and best and worst cases.

Apple SLM RCP 8.5	Historic (km2) 1972–2004	Area Change from Historic (km2) 2028–2058			Area Change from Historic (km2) 2068–2098
Suitability Range		Forecast	Best Case	Worst Case	Forecast	Best Case	Worst Case
0–0.1	14,891	−2027	−3173	−778	−2803	−3182	−2319
0.1–0.2	4286	−1442	−2263	−1235	−3032	−3575	−2289
0.2–0.3	4972	−1606	−1245	−956	−609	−2420	3118
0.3–0.4	8321	−3184	−4552	−1291	2473	−1186	2765
0.4–0.5	12,457	−2104	−4868	312	−1299	−276	−576
0.5–0.6	22,444	−2675	−5170	481	−2841	−6048	−320
0.6–0.7	31,668	3690	887	4116	1518	420	2898
0.7–0.8	32,230	886	3221	−1587	2695	516	3576
0.8–0.9	31,947	8143	9460	5662	14,150	19,873	7711
0.9–1.0	20,509	319	7703	−4724	−10,252	−4122	−14,564

, in Appendix B). Within projection uncertainty, land with very good or excellent suitability for apples could increase in area by 30% (best case) or decrease by 13% (worst case), with more chance of an increase (see Table A3, in Appendix B).

3.2.2. Kiwifruit

Kiwifruit RCP 2.6

Under RCP 2.6, location suitability for kiwifruit is projected to have moderate declines in the upper north of the North Island which has a kiwifruit industry and thus kiwifruit orchards in these areas will be negatively affected, and spatial footprints may decline. However, a number of locations in this area would maintain good suitability (Figure 6). Smaller declines in suitability are expected south of this area and in some eastern coastal areas. However, most of the remaining North Island and the whole of the South Island is projected to experience a small increase in suitability (Figure 6). Thus, most kiwifruit orchards outside the northern areas are likely to experience a small positive impact. Notably, high suitability areas are projected around the mid-western point of the North Island, which is not currently a kiwifruit-growing area.

Under RCP 2.6, the area of land with suitability scores in the acceptable range or higher is expected to increase moderately. However, the area of land with excellent suitability will increase only modestly (see

Table A4. Kiwifruit: land area in the historic period falling into different cultivation suitability ranges, forecast change for the mid- and late-century under RCP 2.6 and best and worst cases.

Kiwifruit SLM RCP 2.6	Historic (km2) 1972–2004	Area Change from Historic (km2) 2028–2058			Area Change from Historic (km2) 2068–2098
Suitability Range		Forecast	Best Case	Worst Case	Forecast	Best Case	Worst Case
0–0.1	37,541	−6215	−13,048	−1090	−6716	−13,358	−1591
0.1–0.2	12,757	−1983	475	−3079	−2080	256	−3287
0.2–0.3	14,845	−2929	−2971	−1100	−3176	−3311	−1417
0.3–0.4	16,808	−1861	−2746	−1457	−1891	−2890	−1636
0.4–0.5	19,596	1515	−698	3121	1578	−677	3637
0.5–0.6	21,305	2873	3919	1477	3035	3968	1774
0.6–0.7	19,058	4377	4089	3680	4988	4894	4079
0.7–0.8	18,653	763	2778	−1107	784	2765	−956
0.8–0.9	16,787	2604	3936	586	2884	4414	698
0.9–1.0	6375	856	4266	−1031	594	3939	−1301

, in Appendix B). Within the projection uncertainty, the area with very good and excellent suitability land could increase by 36% (best case) or decrease by 3% (worst case), and thus the outlook favours an increase in the area of higher suitability land.

Kiwifruit RCP 8.5

Under RCP 8.5, the location suitability projections for the mid-century are similar to the late-century projections under RCP 2.6. However, by the late-century, location suitability is projected to decline further in the northern area of the North Island and in many areas outside the central North Island that are closer to the coast. Many locations in these areas would drop below acceptable suitability especially in the very north (Figure 7). Other areas of the North Island (especially central and mid-western) and all of the South Island are projected to have modest to moderate increases in suitability. Many of these are projected to have very good or excellent location suitability for kiwifruit but are not currently used for kiwifruit. There is the potential for significant changes in the spatial footprint of kiwifruit across the country.

A closer examination of the climatic reasons for the change in suitability would indicate whether adaptation strategies might be possible. For example, adopting cultivars that are more suitable and the use of acceptable bud-break enhancers might enable successful kiwifruit production in areas with decreased suitability. This would protect infrastructural investments such as packhouses and coolstores.

Strong increases in areas of land with high suitability are projected for the mid-century, although the gains for the excellent suitability category are expected to reverse to a small loss by the late-century. Within the uncertainty limits of the projection, the areas of land with suitability scores lying in the acceptable, good and very good categories are consistently projected to increase for both the mid-century and late-century (see

Table A5. Kiwifruit: land area in the historic period falling into different cultivation suitability ranges, forecast change for the mid- and late-century under RCP 8.5 and best and worst cases.

Kiwifruit SLM RCP 8.5	Historic (km2) 1972–2004	Area Change from Historic (km2) 2028–2058			Area Change from Historic (km2) 2068–2098
Suitability Range		Forecast	Best Case	Worst Case	Forecast	Best Case	Worst Case
0–0.1	37,541	−9686	−14,697	−4759	−17,463	−19,709	−15,442
0.1–0.2	12,757	−2002	−889	−4351	−5839	−6114	−5716
0.2–0.3	14,845	−4198	−3394	−3249	−5741	−5724	−5195
0.3–0.4	16,808	−2418	−3948	−1778	−5228	−5849	−3959
0.4–0.5	19,596	383	−613	2948	338	−1100	1939
0.5–0.6	21,305	4372	3953	3224	6462	4249	7256
0.6–0.7	19,058	6768	6813	5570	7760	8058	7735
0.7–0.8	18,653	1655	3830	706	10,464	10,312	9672
0.8–0.9	16,787	3650	5184	1842	9716	13,421	5911
0.9–1.0	6375	1476	3761	−153	−469	2456	−2201

, in Appendix B). Within projection uncertainty, the area of land with very good or excellent suitability could increase from between 16% and 69%.

3.3. Land Use Change

With climate change, the average kiwifruit suitability scores of lands across New Zealand are projected to rise by 9% under RCP 2.6 and by 24% under RCP 8.5 by the end of the century. Changes in the average apple suitability scores of the lands are smaller, being a 4% increase under RCP 2.6 and only a 2.5% increase under RCP 8.5 over the same period.

Figure 8 and Figure 9 illustrate suitability changes graphically by current land use type, focusing on land uses with relatively high baseline suitability for kiwifruit and apples. Kiwifruit suitability decreases significantly on current kiwifruit land but increases in current dairy and non-apple and non-kiwifruit horticulture production areas. Apple suitability also decreases significantly on current apple land. Apple suitability decreases on current dairy land as well, but it increases slightly on current horticulture land.

3.3.1. Multinomial Logit Model Projections

The logit model projections for changes in land use relative to historic baselines are presented in

Table 1. Projected land-use change (hectares) for kiwifruit and apples from the multinomial logit by climate pathway, relative to baseline land-use areas.

Land Use	RCP 2.6 Mid	RCP 2.6 Late	RCP 8.5 Mid	RCP 8.5 Late
Kiwifruit	1875	1825	2675	3800
Apple	150	125	100	−1050

RCP = Representative Concentration Pathway.

. In general, kiwifruit area is projected to expand with climate change, growing by 3800 hectares (about 25% of the current land area) by the late-century under severe climate change. Apple areas are projected to change only marginally under mild climate change (RCP 2.6) but will contract by 1050 hectares (about 10% of the current land area) by the end of the century under severe climate change (RCP 8.5). Mid-century and end-of-century outcomes are nearly identical under RCP 2.6.

3.3.2. LURNZ Simulations

Combining the effects of climate, economic and policy parameters,

Table 2. Simulated land use (hectares) with mild climate change (Representative Concentration Pathway (RCP) 2.6.

Land Use	Basemap	2030	2045	2060	2075
Kiwifruit	15,600	15,750	16,550	17,350	17,475
Apple	9425	9425	9500	9550	9575
Other horticulture	447,450	447,425	447,075	446,800	446,775
Dairy	2,098,200	2,291,375	2,290,925	2,290,500	2,290,450
Sheep–beef	8,152,825	7,488,125	7,243,025	6,916,975	6,685,075
Forestry	2,002,075	2,426,775	2,894,325	3,387,875	3,687,350
Scrub	1,743,000	1,789,700	1,567,175	1,399,525	1,331,875

and

Table 3. Simulated land use (hectares) with severe climate change Representative Concentration Pathway (RCP) 8.5.

Land Use	Basemap	2030	2045	2060	2075
Kiwifruit	15,600	15,825	16,975	18,375	18,675
Apple	9425	9425	9475	425	9125
Other horticulture	447,450	447,400	446,950	446,550	446,850
Dairy	2,098,200	2,291,325	2,290,750	2,290,050	2,290,025
Sheep–beef	8,152,825	7,488,025	7,242,575	6,823,850	6,683,325
Forestry	2,002,075	2,426,775	2,894,300	3,508,800	3,687,300
Scrub	1,743,000	1,789,800	1,567,550	1,371,525	1,333,275

report simulated land-use outcomes over time at the national level for RCPs 2.6 and 8.5, respectively. As economic and policy parameters are fixed across the runs, differences across the tables reflect the effect of changing kiwifruit and apple suitability.

In

Table 4. Simulated land use (hectares) by region for kiwifruit under mild (Representative Concentration Pathway (RCP) 2.6) and severe (RCP 8.5) climate change in New Zealand.

			RCP 2.6			RCP 8.5
Region	Basemap	Loss	Gain	2075	Loss	Gain	2075
Auckland	375	50	0	325	200	0	175
Bay of Plenty	11,625	650	75	11,050	4275	75	7425
Canterbury	0	0	0	0	0	1625	1625
Gisborne	600	25	25	600	125	150	625
Hawkes Bay	350	0	325	675	0	2325	2675
Manawatu-Wanganui	100	0	0	100	0	0	100
Marlborough	0	0	0	0	0	950	950
Nelson	0	0	0	0	0	50	50
Northland	1000	225	0	775	600	0	400
Otago	0	0	0	0	0	400	400
Southland	0	0	0	0	0	0	0
Taranaki	0	0	2425	2425	0	2275	2275
Tasman	700	0	0	700	0	625	1325
Waikato	825	25	0	800	225	0	600
Wellington	0	0	0	0	0	25	25
West Coast	0	0	0	0	0	0	0
Outside RC boundaries	25	0	0	25	0	0	25
Total	15,600	975	2850	17,475	5425	8500	18,675

and

Table 5. Simulated land use (hectares) by region for apples under mild (Representative Concentration Pathway (RCP) 2.6) and severe (RCP 8.5) climate change in New Zealand.

			RCP 2.6			RCP 8.5
Region	Basemap	Loss	Gain	2075	Loss	Gain	2075
Auckland	75	0	0	75	50	0	25
Bay of Plenty	0	0	0	0	0	0	0
Canterbury	50	0	50	100	0	875	925
Gisborne	150	0	25	175	50	25	125
Hawkes Bay	5900	175	125	5850	1600	25	4325
Manawatu-Wanganui	0	0	0	0	0	0	0
Marlborough	0	0	0	0	0	0	0
Nelson	0	0	0	0	0	0	0
Northland	0	0	0	0	0	0	0
Otago	725	0	0	725	0	525	1250
Southland	0	0	0	0	0	0	0
Taranaki	0	0	125	125	0	75	75
Tasman	2250	0	0	2250	50	0	2200
Waikato	150	0	0	150	50	0	100
Wellington	100	0	0	100	25	0	75
West Coast	0	0	0	0	0	0	0
Outside RC boundaries	25	0	0	25	0	0	25
Total	9425	175	325	9575	1825	1525	9125

, a more detailed regional analysis of the simulated land-use changes for kiwifruit and apples is presented. Under each climate pathway, regional losses and gains in land area are shown separately. For example, under the severe climate change pathway represented by RCP 8.5, kiwifruit area across New Zealand is projected to grow to 18,675 hectares by 2075. This is an increase of 3075 hectares. This increase is the net effect of a loss of 5425 hectares of current kiwifruit land and a growth of 8500 hectares elsewhere.

For kiwifruit, the area losses tended to be relatively large in the Bay of Plenty and Northland regions. At the same time, conversions to kiwifruit are projected to happen in some regions that are not traditionally associated with kiwifruit production. Hawke’s Bay and Taranaki experienced relatively large increases under both pathways, while conversions to kiwifruit happened only with severe climate change in Canterbury.

Several previous studies [8,41,42] suggest that a lack of winter chilling due to climate change may make Actinidia chinensis var. deliciosa ‘Hayward’ kiwifruit production uneconomic in the Northland. These studies also indicate a large potential decline in production in the Bay of Plenty. However our suitability scores assess suitability across all cultivars, not just ‘Hayward’; while rising temperatures may result in ‘Hayward’ (a hexaploid cultivar [43]) not receiving sufficient winter chill, diploid and tetraploid cultivars (which are lower chill [44]) might be viable. Thus, while kiwifruit suitability falls significantly in the Northland under RCP 8.5, it does not fall to zero, meaning more suitable cultivars could be available or developed. Consequently, we have not projected a complete loss of kiwifruit land in the Northland.

For apples, the impacts under mild climate change are negligible. Under severe climate change, a large decrease in land use area is projected for the Hawke’s Bay, which is the main apple-growing area of New Zealand. This is partially offset by increases in the South Island regions of Canterbury and Otago. The net effect is a projected decline of 300 hectares in apple land area by 2075.

4. Discussion

4.1. Bias and Variance Corrections to Climate Model Data

The adjustments to the RCP datasets to produce the SLM-RCP datasets were important to improve the baselining of the RCP Past suitability maps to the VCSN past suitability maps. The improvements from these adjustments show that modelled climate data that are bias corrected only for annual means may not be adequate for use with models that are sensitive to monthly or seasonal means, or for models that include considerations of crop damage from extreme temperatures. Bias correction at annual means cannot account for variance in temperatures at scales smaller than a year, such as temperature differences between months or between days in a month. It also cannot account for differences from year to year. Such variances are important for plant physiology and phenology, and thus for biologically motivated models in which these effects are non-linear.

The improvements we obtained from the SLM-RCP data were due to the improved agreement in weather patterns since the corrections did not change the trajectories of the annual means (apart from the slight smoothing introduced by the spline fits).

However, the biases in the RCP-based scores for the Contemporary period were caused not only by a mismatch in monthly variances and means, but also by biases in the predicted annual temperatures. The latter is expected, and is caused by variability in solar radiation, volcanic eruptions and in natural and human emissions of greenhouse gases and aerosols, which are reflected in the observed annual temperatures but not in climate model data for future periods, although there is agreement in the long-term trends [38]. For example, the mean temperature change calculated from an ensemble of climate models was found to track above the observed change for the period 2007 to 2015 [38].

4.2. Suitability Projections

Under RCP 2.6, for the mitigation pathways for kiwifruit and apple crops, there is very little difference between the mid-century suitability maps and late-century maps, with most changes having occurred by the mid-century. This is not surprising, since under RCP 2.6, greenhouse gas emissions are assumed to have peaked in 2020. For both crops, the area of land with excellent suitability (0.9 to 1.0) is projected to increase modestly by mid-century compared with the historic period.

Mid-century suitability maps projected under RCP 8.5 show similarities to the late-century maps projected under RCP 2.6. In comparison with the historic period, the total area of land with excellent overall location suitability is projected to halve for apples by late-century but remain more or less constant for kiwifruit.

Location suitability for both apples and kiwifruit tended to decrease the most in areas of New Zealand with warmer winters, such as the northern and north-eastern locations of the North Island. The decreases in suitability were associated with decreased winter chill, so to enable existing orchards to continue in these regions, cultivars that require less winter chill would need to be planted. By contrast, colder areas of NZ with very good winter chill tended to increase in suitability, which is related to increased GDD accumulation and decreased frost risk with warming climates. In these areas, there would be no pressure for new cultivars since existing cultivars would be a very good fit.

A number of previous studies have concluded that a lack of winter chilling due to large rises in temperature would make kiwifruit cultivation with the ‘Hayward’ cultivar uneconomic in the north of the North Island [8,41,42,45] and in New Zealand’s dominant kiwifruit growing region of Te Puke, in the Bay of Plenty [22,45]. Our study is consistent with these findings; however, it has a more positive outlook because our modelling assesses suitability for all cultivars and not just ‘Hayward’. A lower chill diploid or tetraploid could be used to sustain the industry in locations that have insufficient chill for the current cultivars. Thus, how scores are interpreted is very important. For example, within our suitability score system, a lower chill suitability score could suggest that while ‘Hayward’ would not receive adequate chill, lower chill cultivars might.

For apples, Austin and Hall [46] concluded from studies using the CLIMPACTS system of models that the New Zealand apple industry is unlikely to observe major changes in apple production due to global climate change, whereas we reached a different conclusion which can be explained by our choice of suitability criteria. The CLIMPACTS study differed from ours in its considerations and examined the effect of temperature on (i) the date of bloom; (ii) the date of fruit maturity; and (iii) the rate of fruit growth between these two dates [46], but with no representation of winter chill requirements. Where studies have considered winter chill for apple in climate changes studies, such as the Himachel Pradesh region of India [47] or Japan [48], warming climates have been associated with a decreasing suitability of warmer locations and an increasing suitability of colder locations, which is consistent with our findings.

Changes in location suitability do not necessarily translate to changes in land use, as can be seen in Figure 8 and Figure 9, which show that some cultivations of kiwifruit and apples took place in locations with lower suitability scores, and some locations with high suitability were used for other land uses. A number of factors underlie land use decisions, and thus econometric modelling provides a useful tool to gain more insight.

4.3. Land-Use Econometric Projections

We have incorporated data on the location of kiwifruit and apple blocks as well as on present and future location suitability for kiwifruit and apples in the LURNZ model to analyse the future spatial footprint of these sectors. The continuous nature of our suitability scores provided more nuance to the econometric analysis than if they had been discrete categories.

Nationally, mean kiwifruit suitability rises with climate change. Mean apple suitability also rises, but it rises less under severe climate change than it does under mild climate change. Despite the increasing average suitability across both sectors, suitability for kiwifruit decreases in some current kiwifruit areas, and suitability for apples decreases in some current apple areas.

Our land-use simulations suggest that mild climate change is unlikely to disrupt either sector significantly. However, reflecting changing suitability, we see potentially large shifts in the spatial footprints of kiwifruit and apples under severe climate change. Both sectors experience relatively large losses of land area in currently important production regions. For kiwifruit, the losses are more than offset by increases in land area elsewhere. For apples, this is not the case as total apple area is projected to decline.

4.4. Policy Implications

Policy and planning can be informed by these projections. This study has identified potential reductions in land suitability in major growing regions of kiwifruit and apple that would translate into a reduced presence of these crops in those regions. This could have a flow-on effect on local economies and the local government through impacts on supporting industries, the labour market and increased social pressures. Industries would need to develop mitigation and adaptation strategies in order to retain a significant presence in those regions. Strategies could include importing or supporting the breeding of low chill varieties, or developing alternatives to current agrichemical practices used to promote the uniform bud break and flowering of kiwifruit. Local governments might relax regulatory policies in order to support these strategies, or alternatively may seek diversification and support the establishment of alternative industries.

We have also identified regions that have potential for the new or expanding production of apple and/or kiwifruit. In such regions, existing industries could be put under pressure through increased competition for resources or could be displaced by land-use change with flow-on effects. Local governments would need to consider planning in advance, for example, preparing for future water use requirements, effects on the community (labour requirements, possible population expansion), infrastructure requirements and land rezoning for the development of supporting industries.

4.5. Limitations and Context with Respect to Other Studies

The results discussed reflect direct climate impacts only. They do not consider other important drivers of land-use change such as the global market prices for these crops. Changes in the future suitability of other land uses were also not modelled. Environmental policies, climate variability and the extent and prevalence of extreme weather events, the potential development of new fruit cultivars and the adoption of new technologies will also all affect future land-use decisions, sometimes in unforeseeable ways. Nonetheless, we believe that the simulations presented in this paper for kiwifruit and apples are of interest because they help identify and characterise pressures and opportunities from climate change in these sectors.

Although previous suitability projections for ‘Hayward’ kiwifruit carried out by Tait, Paul, Sood and Mowat [22] were expressed in categories of “poor”, “marginal” and “good” rather than on a continuous scale, it was visually evident that they do not align with our suitability projections for the hindcast and future periods of the climate model data. This is not surprising since, as we noted in the introduction, those authors expressed suitability only in terms of mean temperatures, which discounts soil limitations and extreme temperature effects such as frost. Although that study and ours used data from the same RCM runs, we performed additional bias and variance corrections that will influence suitability predictions. Additionally, we had modelled across a range of kiwifruit cultivars, which will give higher suitability for some areas since some cultivars can be grown where ‘Hayward’ cannot.

Our study is similar to work by Cronin, Zabel, Dessens and Anandarajah [16], who also used continuous suitability scores; however, we combined suitability scores for different criteria by using geometric averaging to get an overall score, whereas those authors used the approach of Zabel et al. [49] by taking the minimum value across criteria. Their study asked different questions, and used spatial land-use projections from the IMAGE model [50] to identify potential conflicts between energy crops and other land uses. The IMAGE model includes a number of process-orientated drivers to project land-use change between a number of arable crops, livestock and forestry in order to meet projected demands [50], while our approach was more crop specific and less-process oriented, with changes in land use driven by suitability rather than demand. Artificial intelligence or ANN approaches (e.g., [17]) could alleviate the need to represent the mechanisms underlying suitability, but this may be at the expense of understanding causative factors.

Our modelling considers the influence of a number of socio-economic factors such as farmers’ decision making and social attitudes in an implicit manner, through parameter estimations from the data. This does not allow changes in these factors to be modelled, which is a limitation compared with recent agent-based approaches (e.g., [15,18]).

An additional limitation of our approach is we have not sought to model yield, as have other recent approaches (e.g., [13,14]). Doing so would require data on production, weather patterns and management strategies at the orchard level to calibrate models, but would be useful for modelling the effectiveness of adaptation strategies.

The horticulture sector in the LURNZ modelling included all horticultural endeavours other than apples and kiwifruit, and these would be differentially affected by climate change. For example, warm-climate crops such as avocados do not require winter chill. Areas that become less suitable for deciduous crops may increase in suitability for avocados, presenting a natural land-use change. Representing these other crops and other competing land uses with separate location suitability scores in the LURNZ model would allow a more complex dynamic to be modelled.

The 5 × 5 km resolution of the climate data is relatively coarse. There could be microclimates within the area represented by each grid cell but this could not be modelled. Furthermore, the VCSN data used for obtaining calibration maps are modelled, being spatially interpolated from observation data from weather stations, and thus bring unknown uncertainties induced from the smoothing splines. A key source of uncertainty lies in the reliability of the climate change signals generated from the climate models and how that translates to suitability predictions. While we obtained uncertainty estimates by comparing the deviation in suitability predictions between the VCSN and RCM data for 2006 to 2016, how well this will translate for future period projections is unclear. We consider that these cannot be ameliorated uncertainty estimates based on the variance arising from the RCM forcing by the different GCMs. Such estimates indicate the level of disagreement between the GCMs, but not the reliability of the average trajectory. Another important source of uncertainty is the accuracy of the land and terrain databases, and the limitations of the gridded data in capturing the spatial variability within a grid similarly applied here as with gridded climate data.

5. Conclusions

This paper has addressed the threats and opportunities faced by the apple and kiwifruit industries under future climate change by developing projections of location suitability using biologically driven suitability models, and incorporating these suitability projections into econometric models to project land-use change and the spatial footprints of these industries. This provides valuable insight to guide robust policy and climate adaptation plans by governments and industries. The methodology has sufficient generality in that it can be extended to a range of other crops by incorporating appropriate suitability models and it could be applied in other geographical areas.

Key future areas where our suitability modelling can be improved include (i) further refinement of the bias and variance corrections to the predicted climate data; (ii) extending the suitability modelling to predict yield under changing climates, including the impact of management strategies; and (iii) modelling the effectiveness of climate change adaptations. Incorporating yield and/or profits under climate change with different adaptations could provide additional power to the econometric modelling of land use, as would the explicit incorporation of the attitudes and priorities of growers, society and governments.