Potential impacts of climate change on agriculture and fisheries production in 72 tropical coastal communities
Sampling of coastal communitiesHere, we integrated data from five different projects that had surveyed coastal communities across five countries47,48,49,50. Between 2009 and 2015, we conducted socioeconomic surveys in 72 sites from Indonesia (n = 25), Madagascar (n = 6), Papua New Guinea (n = 10), the Philippines (n = 25), and Tanzania (Zanzibar) (n = 6). Site selection was for broadly similar purposes- to evaluate the effects of various coastal resource management initiatives (collaborative management, integrated conservation and development projects, recreational fishing projects) on people’s livelihoods in rural and peri-urban villages. Within each project, sites were purposively selected to be representative of the broad range of socioeconomic conditions (e.g., population size, levels of development, integration to markets) experienced within the region. We did not survey strictly urban locations (i.e., major cities). Because our sampling was not strictly random, care should be taken when attempting to make inferences beyond our specific study sites.We surveyed between 13 and 150 households per site, depending on the population of the communities and the available time to conduct interviews per site. All projects employed a comparable sampling design: households were either systematically (e.g., every third house), randomly sampled, or in the case of three villages, every household was surveyed (a census) (see Supplementary Data file). Respondents were generally the household head, but could have been other household members if the household head was not available during the study period (i.e. was away). In the Philippines, sampling protocol meant that each village had an even number of male and female respondents. Respondents gave verbal consent to be interviewed.The following standard methodology was employed to assess material style of life, a metric of material assets-based wealth48,51. Interviewers recorded the presence or absence of 16 material items in the household (e.g., electricity, type of walls, type of ceiling, type of floor). We used a Principal Component Analysis on these items and kept the first axis (which explained 34.2% of the variance) as a material wealth score. Thus, each community received a mean material style of life score, based on the degree to which surveyed households had these material items, which we then scaled from 0 to 1. We also conducted an exploratory analysis of how material style of life has changed in two sites in Papua New Guinea (Muluk and Ahus villages) over fifteen and sixteen-year time span across four and five-time periods (2001, 2009, 2012, 2016, and 2002, 2009, 2012, 2016, 2018), respectively, that have been surveyed since 2001/200252. These surveys were semi-panel data (i.e. the community was surveyed repeatedly, but we did not track individuals over each sampling interval) and sometimes occurred in different seasons. For illustrative purposes, we plotted how these villages changed over time along the first two principal components.SensitivityWe asked each respondent to list all livelihood activities that bring in food or income to the household and rank them in order of importance. Occupations were grouped into the following categories: farming, cash crop, fishing, mariculture, gleaning, fish trading, salaried employment, informal, tourism, and other. We considered fishing, mariculture, gleaning, fish trading together as the ‘fisheries’ sector, farming and cash crop as the ‘agriculture’ sector and all other categories into an ‘off-sector’.We then developed three distinct metrics of sensitivity based on the level of dependence on agriculture, fisheries, and both sectors together. Each metric incorporates the proportion of households engaged in a given sector (e.g., fisheries), whether these households also engage in occupations outside of this sector (agriculture and salaried/formal employment; referred to as ‘linkages’ between sectors), and the directionality of these linkages (e.g., whether respondents ranked fisheries as more important than other agriculture and salaried/formal employment) (Eqs. 1–3)$${{{{{{rm{S}}}}}}}_{{{{{{rm{A}}}}}}}=,frac{{{{{{rm{A}}}}}}}{{{{{{rm{A}}}}}}+{{{{{rm{NA}}}}}}},times ,frac{{{{{{rm{N}}}}}}}{{{{{{rm{A}}}}}}+{{{{{rm{NA}}}}}}},times ,frac{left(frac{{{{{{{rm{r}}}}}}}_{{{{{{rm{a}}}}}}}}{2}right),+,1}{{{{{{{rm{r}}}}}}}_{{{{{{rm{a}}}}}}}+,{{{{{{rm{r}}}}}}}_{{{{{{rm{na}}}}}}}+1}$$
(1)
$${{{{{{rm{S}}}}}}}_{{{{{{rm{F}}}}}}}=,frac{{{{{{rm{F}}}}}}}{{{{{{rm{F}}}}}}+{{{{{rm{NF}}}}}}},times ,frac{{{{{{rm{N}}}}}}}{{{{{{rm{F}}}}}}+{{{{{rm{NF}}}}}}},times ,frac{left(frac{{{{{{{rm{r}}}}}}}_{{{{{{rm{f}}}}}}}}{2}right),+,1}{{{{{{{rm{r}}}}}}}_{{{{{{rm{f}}}}}}}+,{{{{{{rm{r}}}}}}}_{{{{{{rm{nf}}}}}}}+1}$$
(2)
$${{{{{{rm{S}}}}}}}_{{{{{{rm{AF}}}}}}}=,frac{{{{{{rm{AF}}}}}}}{{{{{{rm{AF}}}}}}+{{{{{rm{NAF}}}}}}},times ,frac{{{{{{rm{N}}}}}}}{{{{{{rm{AF}}}}}}+{{{{{rm{NAF}}}}}}},times ,frac{left(frac{{{{{{{rm{r}}}}}}}_{{{{{{rm{af}}}}}}}}{2}right),+,1}{{{{{{{rm{r}}}}}}}_{{{{{{rm{af}}}}}}}+,{{{{{{rm{r}}}}}}}_{{{{{{rm{naf}}}}}}}+1}$$
(3)
where ({{{{{{rm{S}}}}}}}_{{{{{{rm{A}}}}}}}), ({{{{{{rm{S}}}}}}}_{{{{{{rm{F}}}}}}}) and ({{{{{{rm{S}}}}}}}_{{{{{{rm{AF}}}}}}}) are a community’s sensitivity in the context of agriculture, fisheries and both sectors, respectively. A, F and AF are the number of households relying on agriculture-related occupations within that community, fishery-related and agriculture- and fisheries-related occupations within the community, respectively. NA, NF and NAF are the number of households relying on non-agriculture-related, non-fisheries-related, and non-agriculture-or-fisheries-related occupations within the community, respectively. N is the number of households within the community. ({{{{{{rm{r}}}}}}}_{{{{{{rm{a}}}}}}}), ({{{{{{rm{r}}}}}}}_{{{{{{rm{f}}}}}}}) and ({{{{{{rm{r}}}}}}}_{{{{{{rm{af}}}}}}}) are the number of times agriculture-related, fisheries-related and agriculture-and-fisheries-related occupations were ranked higher than their counterpart, respectively. ({{{{{{rm{r}}}}}}}_{{{{{{rm{na}}}}}}}), ({{{{{{rm{r}}}}}}}_{{{{{{rm{nf}}}}}}}) and ({{{{{{rm{r}}}}}}}_{{{{{{rm{naf}}}}}}}) are the number of times non-agriculture, non-fisheries, and non-agriculture-and-fisheries-related occupations were ranked higher than their counterparts. As with the material style of life, we also conducted an exploratory analysis of how joint agriculture-fisheries sensitivity has changed over time in a subset of sites (Muluk and Ahus villages in Papua New Guinea) that have been sampled since 2001/200252. Although our survey methodology has the potential for bias (e.g. people might provide different rankings based on the season, or there might be gendered differences in how people rank the importance of different occupations53), our time-series analysis suggest that seasonal and potential respondent variation do not dramatically alter our community-scale sensitivity metric.ExposureTo evaluate the exposure of communities to the impact of future climates on their agriculture and fisheries sectors, we used projections of production potential from the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP) Fast Track phase 3 experiment dataset of global simulations. Production potential of agriculture and fisheries for each of the 72 community sites and 4746 randomly selected sites from our study countries with coastal populations >25 people/km2 were projected to the mid-century (2046–2056) under two emission scenarios (SSP1-2.6, and SSP5-8.5) and compared with values from a reference historical period (1983–2013).For fisheries exposure (EF), we considered relative change in simulated total consumer biomass (all modelled vertebrates and invertebrates with a trophic level >1). For each site, the twenty nearest ocean grid cells were determined using the Haversine formula (Supplementary Fig. 5). We selected twenty grid cells after a sensitivity analysis to determine changes in model agreement based on different numbers of cells used (1, 3, 5, 10, 20, 50, 100; Supplementary Figs. 6–7), which we balanced off with the degree to which larger numbers of cells would reduce the inter-site variability (Supplementary Fig. 8). We also report 25th and 75th percentiles for the change in marine animal biomass across the model ensemble. Projections of the change in total consumer biomass for the 72 sites were extracted from simulations conducted by the Fisheries and marine ecosystem Model Intercomparison Project (FishMIP3,54). FishMIP simulations were conducted under historical, SSP1-2.6 (low emissions) and SSP5-8.5 (high emissions) scenarios forced by two Earth System Models from the most recent generation of the Coupled Model Intercomparison project (CMIP6);55 GFDL-ESM456 and IPSL-CM6A-LR57. The historical scenario spanned 1950–2014, and the SSP scenarios spanned 2015–2100. Nine FishMIP models provided simulations: APECOSM58,59, BOATS60,61, DBEM2,62, DBPM63, EcoOcean64,65, EcoTroph66,67, FEISTY68, Macroecological69, and ZooMSS11. Simulations using only IPSL-CM6A-LR were available for APECOSM and DBPM, while the remaining 7 FishMIP models used both Earth System Model forcings. This resulted in 16 potential model runs for our examination of model agreement, albeit with some of these runs being the same model forced with two different ESMs. Thus, the range of model agreement could range from 8 (half model runs indicating one direction of change, and half indicating the other) to 16 (all models agree in direction of change). Model outputs were saved with a standardised 1° spatial grid, at either a monthly or annual temporal resolution.For agriculture exposure (EA), we used crop model projections from the Global Gridded Crop model Intercomparison Project (GGCMI) Phase 314, which also represents the agriculture sector in ISIMIP. We used a window of 11×11 cells centred on the site and removed non-land cells (Supplementary Fig. 5). The crop models use climate inputs from 5 CMIP6 ESMs (GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL), downscaled and bias-adjusted by ISIMIP and use the same simulation time periods. We considered relative yield change in three rain-fed and locally relevant crops: rice, maize, and cassava, using outputs from 4 global crop models (EPIC-IIASA, LPJmL, pDSSAT, and PEPIC), run at 0.5° resolution. These 4 models with 5 forcings generate 20 potential model runs for our examination of model agreement. Yield simulations for cassava were only available from the LPJmL crop model. All crop model simulations assumed no adaptation in growing season and fertilizer input remained at current levels. Details on model inputs, climate data, and simulation protocol are provided in ref. 14. At each site, and for each crop, we calculated the average change (%) between projected vs. historical yield within 11×11 cell window. We then averaged changes in rice, maize and cassava to obtain a single metric of agriculture exposure (EA).We also obtained a composite metric of exposure (EAF) by calculating each community’s average change in both agriculture and fisheries:$${{{{{{rm{E}}}}}}}_{{{{{{rm{AF}}}}}}}=,frac{{{{{{{rm{E}}}}}}}_{{{{{{rm{A}}}}}}}+,{{{{{{rm{E}}}}}}}_{{{{{{rm{F}}}}}}}}{2}$$
(4)
Potential ImpactWe calculated relative potential impact as the Euclidian distance from the origin (0) of sensitivity and exposure.Sensitivity testTo determine whether our sites displayed a particular exposure bias, we compared the distributions of our sites and 4746 sites that were randomly selected from 47,460 grid cells within 1 km of the coast of the 5 countries we studied which had population densities >25 people/km2, based on the SEDAC gridded populating density of the world dataset (https://sedac.ciesin.columbia.edu/data/set/gpw-v4-population-density-rev11/data-download).We used Cohen’s D to determine the size of the difference between our sites and the randomly selected sites.Validating ensemble modelsWe attempted a two-stage validation of the ensemble model projections. First, we reviewed the literature on downscaling of ensemble models to examine whether downscaling validation had been done for the ecoregions containing our study sites.While no fisheries ensemble model downscaling had been done specific to our study regions, most of the models of the ensemble have been independently evaluated against separate datasets aggregated at scales down to Large Marine Ecosystems (LMEs) or Exclusive Economic Zones (EEZs) (see11). For example, the DBEM was created with the objective of understanding the effects of climate change on exploited marine fish and invertebrate species2,70. This model roughly predicts species’ habitat suitability; and simulates spatial population dynamics of fish stocks to output biomass and maximum catch potential (MCP), a proxy of maximum sustainable yield2,62,71. Compared with spatially-explicit catch data from the Sea Around Us Project (SAUP; www.seaaroundus.org)70 there were strong similarities in the responses to warming extremes for several EEZs in our current paper (Indonesia and Philippines) and weaker for the EEZs of Madagascar, Papua New Guinea, and Tanzania. At the LME level, DBEM MCP simulations explained about 79% of the variation in the SAUP catch data across LMEs72. The four LMEs analyzed in this paper (Agulhas Current; Bay of Bengal; Indonesian Sea; and Sulu-Celebes Sea) fall within the 95% confidence interval of the linear regression relationship62. Another example, BOATS, is a dynamic biomass size-spectrum model parameterised to reproduce historical peak catch at the LME scale and observed catch to biomass ratios estimated from the RAM legacy stock assessment database (in 8 LMEs with sufficient data). It explained about 59% of the variability of SAUP peak catch observation at the LME level with the Agulhas Current, Bay of Bengal, and Indonesian Sea catches reproduced within +/-50% of observations61. The EcoOcean model validation found that all four LMEs included in this study fit very close to the 1:1 line for overserved and predicted catches in 200064,65. DBPM, FEISTY, and APECOSM have also been independently validated by comparing observed and predicted catches. While the models of this ensemble have used different climate forcings when evaluated independently, when taken together the ensemble multi-model mean reproduces global historical trends in relative biomass, that are consistent with the long term trends and year-on-year variation in relative biomass change (R2 of 0.96) and maximum yield estimated from stock assessment models (R2 of 0.44) with and without fishing respectively11.Crop yield estimates simulated by GGCMI crop models have been evaluated against FAOSTAT national yield statistics14,73,74. These studies show that the models, and especially the multi-model mean, capture large parts of the observed inter-annual yield variability across most main producer countries, even though some important management factors that affect observed yield variability (e.g., changes in planting dates, harvest dates, cultivar choices, etc.) are not considered in the models. While GCM-based crop model results are difficult to validate against observations, Jägermeyr et al14. show that the CMIP6-based crop model ensemble reproduces the variability of observed yield anomalies much better than CMIP5-based GGCMI simulations. In an earlier crop model ensemble of GGCMI, Müller et al.74 show that most crop models and the ensemble mean are capable of reproducing the weather-induced yield variability in countries with intensely managed agriculture. In countries where management introduces strong variability to observed data, which cannot be considered by models for lack of management data time series, the weather-induced signal is often low75, but crop models can reproduce large shares of the weather-induced variability, building trust in their capacity to project climate change impacts74.We then attempted to validate the models in our study regions. For the crop models, we examined production-weighted agricultural projections weighted by current yields/production area (Supplementary Fig. 1). We used an observational yield map (SPAM2005) and multiplied it with fractional yield time series simulated by the models to calculate changes in crop production over time, which integrates results in line with observational spatial patterns. The weighted estimates were not significantly different to the unweighted ones (t = 0.17, df = 5, p = 0.87). For the fisheries models, our study regions were data-poor and lacked adequate stock assessment data to extend the observed global agreement of the sensitivity of fish biomass to climate during our reference period (1983-2013). Instead, we provide the degree of model run agreement about the direction of change in the ensemble models to ensure transparency about the uncertainty in this downscaled application.AnalysesTo account for the fact that communities were from five different countries we used linear mixed-effects models (with country as a random effect) for all analyses. All averages reported (i.e. exposure, sensitivity, and model agreement) are estimates from these models. In both our comparison of fisheries and agriculture exposure and test of differences between production-weighted and unweighted agriculture exposure we wanted to maintain the paired nature of the data while also accounting for country. To accomplish this we used the differences between the exposure metrics as the response variable (e.g. fisheries exposure minus agriculture exposure), testing whether these differences are different from zero. We also used linear mixed-effects models to quantify relationships between the material style of life and potential impacts under different mitigation scenarios (SSP1-2.6 and 8.5), estimating standard errors from 1000 bootstrap replications. To further explore whether these relationships between the material style of life and potential impacts were driven by exposure or sensitivity, we conducted an additional analysis to quantify relationships between the material style of life and: 1) joint fisheries and agricultural sensitivity; 2) joint fisheries and agricultural exposure under different mitigation scenarios. We present both the conditional R2 (i.e., variance explained by both fixed and random effects) and the marginal R2 (i.e., variance explained by only the fixed effects) to help readers compare among the material style of life relationships.Reporting summaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article. More
