MaxEnt and Marxan modeling to predict the potential habitat and priority planting areas of Coffea arabica in Yunnan, China under climate change scenario Authors: Xia Li, Zihao Wang, Shaoqiang Wang, Zhaohui Qian Published: 2024 ABSTRACT Introduction Coffea arabica (Arabica coffee) is an important cash crop in Yunnan, China. Ongoing climate change has made coffee production more difficult to sustain, posing challenges for the region’s coffee industry. Predictions of the distribution of potentially suitable habitats for Arabica coffee in Yunnan could provide a theoretical basis for the cultivation and rational management of this species.MethodsIn this study, the MaxEnt model was used to predict the potential distribution of suitable habitat for Arabica coffee in Yunnan under current and future (2021-2100) climate scenarios (SSP2-4.5, SSP3-7.0, and SSP5-8.5) using 56 distributional records and 17 environmental variables and to analyze the important environmental factors. Marxan model was used to plan the priority planting areas for this species at last.ResultsThe predicted suitable and sub-suitable areas were about 4.21×104 km2 and 13.87×104 km2, respectively, accounting for 47.15% of the total area of the province. The suitable areas were mainly concentrated in western and southern Yunnan. The minimum temperature of the coldest month, altitude, mean temperature of the wettest quarter, slope, and aluminum saturation were the main environmental variables affecting the distribution of Arabica coffee in Yunnan Province. Changes in habitat suitability for Arabica coffee were most significant and contracted under the SSP3-7.0 climate scenario, while expansion was highest under the SSP5-8.5 climate scenario. Priority areas for Arabica coffee cultivation in Yunnan Province under the 30% and 50% targets were Pu’er, Xishuangbanna, Honghe, Dehong, and Kunming.DiscussionClimate, soil, and topography combine to influence the potential geographic distribution of Arabica coffee. Future changes in suitable habitat areas under different climate scenarios should lead to the delineation of coffee-growing areas based on appropriate environmental conditions and active policy measures to address climate change. Introduction Coffee as one of the world’s three major beverage crops (Wu et al., 2016), ranks second only to petroleum as the world’s second-largest traded commodity (Eriyagama et al., 2013). Coffee plants are currently cultivated in over 80 countries globally, with the coffee industry significantly shaping the agricultural economies of major coffee-producing nations (ICO, 2022), providing livelihoods for millions of people directly or indirectly dependent on coffee. In Asia, China ranks as the fourth largest coffee exporter after Vietnam, Indonesia, and India (ICO, 2022), accompanied by an annual foreign exchange income of $532 million (Zhang et al., 2021b). Yunnan Province has a coffee cultivation history of over 100 years, with coffee planting area, production, and agricultural output value accounting for over 98% of the national total (Ma et al., 2022). In 2021, the planting area and green bean production in Yunnan Province accounted for 0.82% and 1.08% of the global total (Yunnan Provincial Department of Agriculture and Rural Affairs, 2022), making it a unique and dominant industry in China’s Yunnan Province. Coffea arabica (Arabica coffee) is the predominant variety in the region and globally (Rigal et al., 2020a), primarily used for specialty coffee, with high economic value (Fain et al., 2018), serving as the primary cash crop for local growers. In 2021, the proportion of specialty coffee in Yunnan Province was only 8%, but it increased to 22.7% by 2023, indicating a shift towards specialty coffee as the direction of development in Yunnan Province’s coffee industry (Yunnan Provincial Department of Agriculture and Rural Affairs, 2023). However, coffee cultivation will transform forests, and the future of Yunnan’s forests will likewise be highly influenced by the intensity of management (Kebebew and Ozanne, 2022). Despite commitments from the private sector, including coffee traders, aimed at eliminating deforestation from operations or supply chains, zero deforestation policies may not be sufficient to have a wider impact on their own, due to leakage, lack of transparency and traceability, selective adoption, and the marginalization of smallholder farmers (Lambin et al., 2018). Therefore, the promotion of coffee cultivation likewise needs to be integrated with public policies put in place by the government, especially spatial planning policies, to harmonize forest conservation with community interests. Climate is a primary abiotic factor influencing species distribution (Bezeng et al., 2017). Climate change affects plant growth, geographical distribution, and population size (Alan Pounds et al., 2006), posing challenges to global agriculture and forestry, including the coffee industry. Arabica coffee and Robusta coffee (Coffea canephora) are the most cultivated species, with relatively narrow ranges of bioclimatic parameters (Teketay, 1999). Among them, Arabica coffee is particularly sensitive to climate (Davis et al., 2012), and coffee planted under marginal bioclimatic conditions is more susceptible to environmental stress (Fain et al., 2018). Continued climate change may exacerbate existing issues and create new challenges for the coffee industry, making production more difficult to sustain (Reyer et al., 2017). Climate change due to strong emissions will reduce the global area suitable for coffee cultivation by about half, and higher temperatures may reduce Arabica coffee yields. In Asia, some forested areas may instead become more suitable (Bunn et al., 2015). Previous studies have shown the negative impacts of climate change on coffee production in major coffee-producing countries such as Brazil, Ethiopia, and India (Koh et al., 2020; Adane and Bewket, 2021; Byrareddy et al., 2024). Understanding the effects of climate change on the biological distribution of Arabica coffee in Yunnan Province, China, and how Arabica coffee adapts to changing climatic conditions is of significant importance for the future ecological sustainability of the coffee industry in this region (Wang et al., 2019). Species Distribution Models (SDMs) use environmental data for sites of occurrence (presence) of a species to predict a response variable to predict whether the environmental conditions are suitable for the species to persist meaning that the species is expected to occur (Araújo and Peterson, 2012). Since BioClim as the first tool to be widely used, SDMs have been widely applied to model biological responses to climate change (Booth et al., 2014). The Maximum Entropy (MaxEnt) model is a widely used SDM technique employed to estimate species distribution probabilities based on the MaxEnt principle. It relies on a statistical procedure using real observation data of species presence or abundance to infer potential suitability based on environmental characteristics, thus representing spatial suitability for species occurrence based on the variables used (Phillips et al., 2004; Phillips and Dudík, 2008). The MaxEnt model assumes that the relationship between species distribution and environmental variables maximizes entropy, and even if the information on species distribution data and the environmental variables in the distribution area is incomplete, it can still make relatively accurate predictions of the potential distribution areas of the species with good stability, and the predictions are basically in line with the actual distribution of the species, especially in the preliminary stage of the study, which provides valuable insights into the potential patterns of the species distribution (Xu et al., 2024). Compared to models such as BIOCLIM and DOMAIN, the MaxEnt model has a significant advantage of higher prediction accuracy, smaller confidence intervals, and is more stable and less affected by the random variable (Hernandez et al., 2006; Giovanelli et al., 2010; Hijmans, 2012; Duan et al., 2014). It is also capable of handling multiple types of environmental data (e.g., topography, soils, etc.) in addition to climate variables, comparing with the Climate Change Adaptation Modeler of TerrSet, to predict different climate scenarios, which provides a flexible tool to respond to diverse challenges (Ngoy and Shebitz, 2020; Lessmann et al., 2014; Han et al., 2024). Currently, the MaxEnt model has been widely used in plant protection-oriented species distribution studies (He et al., 2021). In the case of cash crops, several studies have applied the model to the suitability of corn, goji berries, citrus, coffee, etc (Davis et al., 2012; Fitzgibbon et al., 2022; Li et al., 2024; Lin et al., 2022; Zhang et al., 2021b; Cassamo et al., 2023). In China, although Zhang et al. (2021a) compared the performance of the AHP-GIS method with that of the MaxEnt model using default parameters and concluded that the latter is a more suitable tool for simulating the potential distribution of Arabica coffee in Yunnan, however, MaxEnt has been criticized as giving a biased representation of suitable climates if the parameters are not chosen carefully (Bunn et al., 2015). Therefore, the parameters need to be debugged and optimized to apply to a particular species when using the model. Systematic conservation planning is a method frequently employed internationally for biodiversity conservation and determining priority conservation areas. It involves systematically protecting the biodiversity features of an entire area, and it is a purposeful, effective investment (Segan et al., 2011). Based on the complementarity principle, the Marxan model uses simulated annealing algorithms to identify the optimal set of planning units through iterative operations and repeated selections (Lehtomäki and Moilanen, 2013). It allows for the comparison of different potential conservation areas based on user-defined goals and costs and determines the regional set that most effectively achieves its objectives (Cudlín et al., 2020), addressing the minimum set problem in conservation planning (Smith et al., 2010). While the MaxEnt model emphasizes the suitability of species distribution based on natural environmental variables such as temperature, precipitation, and soil, the Marxan model identifies priority areas that reduce conflicts between humans and ecosystems and minimizes human impacts on species, thus considering more socio-economic factors than the MaxEnt model (Yu et al., 2023). By combining the MaxEnt and Marxan models, relevant management costs can be minimized. Simulating the spatial distribution of coffee habitat as a feature to support a systematic conservation planning model can lead to spatial planning that maximizes benefits and minimizes land costs. This study utilized multi-source occurrence record geographic location data, current climate variables under different Shared Socioeconomic Pathways (SSPs) from CMIP6, as well as environmental data such as terrain and soil, to model the potential suitable habitat distribution of Arabica coffee in Yunnan Province, China using the MaxEnt model after parameter optimization with the Kuenm R package, including the geographical distribution characteristics under current climate conditions and the main environmental variables influencing it. Future potential geographical distribution, spatial changes, and centroid migration trajectories under different climate scenarios were also predicted. Finally, using the distribution probability values of Arabica coffee as the conservation feature and land use/land cover and natural protected areas data as costs, the Marxan model was employed to identify priority planting areas for Arabica coffee in Yunnan Province. The results of the study provide a sustainable and implementable reference for spatial planning and agricultural policies for local government decision-makers and smallholder farmers to promote the cultivation of Arabica coffee. Materials and methods Collection and processing of sample distribution points The distribution location data of Coffea arabica in Yunnan Province used in this study were obtained from the Global Biodiversity Information Facility (GBIF, https://www.gbif.org), Naturalist (https://www.inaturalist.org), Biological Plant Specimen Museum of the Chinese Academy of Sciences (https://pe.ibcas.ac.cn), Chinese Virtual Herbarium (CVH, https://www.cvh.ac.cn) and relevant papers on Yunnan Province Coffee (Ge et al., 2023; Hao et al., 2022; Li et al., 2021, 2023; Lu et al., 2022; Rigal et al., 2020a, 2020b; Wan et al., 2024; Yang et al., 2022) collected from the Web of Science (WOS, https://www.webofscience.com), totaling 331 sample points. All occurrence records especially those lacking precise geographic coordinates were revised with specific latitude and longitude information using Google Earth. To minimize spatial autocorrelation between sample points and reduce errors caused by model overfitting due to sampling bias, duplicate records were removed using ENMTools 1.4.4, ensuring that only one point was retained per grid cell, and each remaining point was moved to the center of its grid cell (Ma et al., 2024; Warren et al., 2021). After the aforementioned processing, a total of 56 valid species sample distribution records were adopted ( Figure 1 ). Sample distribution of Coffea arabica in the study area. Selection of environmental variables To simulate the current and future suitable habitat for coffee cultivation in Yunnan Province, a total of 32 environmental variables potentially influencing the distribution of Arabica coffee was selected based on the previous study (Zhang et al., 2021b). These factors include climate, terrain, and soil variables ( Table 1 ). For the current and future (from 2021 to 2100) climate data, we retrieved raster layers for future climate projections under three Shared Socio-economic Pathways (SSPs), namely SSP2–4.5(Upgrade of scenario RCP4.5 from SSP2 for the medium forcing scenario), SSP3-7.0 (Additional scenario RCP7.0 emission pathways based on SSP3 for a rocky road) and SSP5–8.5 (Upgrade of scenario RCP8.5 from SSP5 for high forcing scenario) for the bioclimatic variables (Lenoir et al., 2008; Jia et al., 2024). The layers were extracted from the sixth version of Model for Interdisciplinary Research on Climate (MIROC6) which was retrieved from the sixth phase of the Coupled Model Intercomparison Project (CMIP6) in WorldClim (https://www.worldclim.org/data/cmip6/cmip6_clim30s.html). The model has shown good performance in climate characterization (Tian et al., 2021), particularly in southern China, and has been widely used in the simulation of species distribution models in this region as well as in other similar climatic zones (Li et al., 2022; Coulibaly et al., 2023; Oduor et al., 2023). Terrain data were obtained from the NASA Digital Elevation Model (NASA EARTHDATA, https://search.earthdata.nasa.gov) with a spatial resolution of 30 meters. Elevation, slope, and aspect were extracted for the study area using ArcGIS 10.8. Soil data were sourced from the Harmonized World Soils Database version 2.0 (HWSD 2.0, https://gaez.fao.org/pages/hwsd), providing information on the morphological, chemical, and physical properties of soils at approximately 1 km resolution. The soil characteristics of the topsoil layer (D1) of the soil mapping units with the highest share proportion were utilized. All environmental variables were resampled to a spatial resolution of 30 arc-seconds (ca. 1 km2 at the equator) and converted into the appropriate format using ArcGIS 10.8. Table 1Environmental variables used to model the potential distribution of Arabica coffee.CategoryVariableDescriptionUnitBioclimatic variablesbio1Annual mean temperature°Cbio2Mean diurnal range°Cbio3Isothermality (bio2/bio7) (×100)–bio4Temperature seasonality (standard deviation ×100)–bio5Max temperature of the warmest month°Cbio6Min temperature of the coldest month°Cbio7Temperature annual range (Bio5-Bio6)°Cbio8The mean temperature of the wettest quarter°Cbio9The mean temperature of the driest quarter°Cbio10The mean temperature of the warmest quarter°Cbio11The mean temperature of the coldest quarter°Cbio12Annual precipitationmmbio13Precipitation of the wettest monthmmbio14Precipitation of the driest monthmmbio15Precipitation seasonality (coefficient of variation)mmbio16Precipitation of the wettest quartermmbio17Precipitation of the driest quartermmbio18Precipitation of the warmest quartermmbio19Precipitation of the coldest quartermmTerrain variablesaltitudeAltitudemslopeSlope°aspectAspect°Soil variablessandTopsoil sand fraction% wtsiltTopsoil silt fraction% wtclayTopsoil clay fraction% wtorg_carbonOrganic carbon contentg/kgph_waterpH Value (H2O)-log(H+)total_nTotal nitrogen contentg/kgcec_soilCation exchange capacity soilcmol/kgalum_satAluminum saturation% ECECtcarbon_eqCalcium carbonate% weightgypsumGypsum content% weight Environmental variables used to model the potential distribution of Arabica coffee. Species distribution models (SDMs) may be influenced directly or indirectly by the predictive variables used, due to factors such as approximation, causality, or correlation among them (Zhang et al., 2020). To mitigate issues such as multicollinearity, autocorrelation, and redundancy of environmental variables that may lead to model overfitting (Burgos et al., 2020), we first conducted a Pearson correlation coefficient analysis to retain factors within each environmental variable type with correlation coefficients |r|< 0.85 as illustrated in Figure 2 . Result of Pearson’s correlation coefficient. Subsequently, all environmental factors were inputted into a preliminary MaxEnt model with default parameters, and factors with low or zero contribution rates in the results were removed. Finally, 17 key factors (bio3, bio4, bio6, bio7, bio8, bio12, bio14, bio15, bio17, bio18, altitude, slope, aspect, alum_sat, clay, pH_water, tcarbon_eq) were identified as the environmental variables for predicting the distribution of Coffea Arabica in Yunnan Province. The results of the preliminary experiment are presented in Supplementary Table S1 . Result of Pearson’s correlation coefficient. Establishment, optimization, and evaluation of MaxEnt model The principle of maximum entropy refers to the full consideration of known information when inferring the unknown probability distribution, mainly based on Shannon’s information entropy theory, which decreases when the total amount of information increases (Allahverdyan et al., 2021). Under the premise of ensuring the inclusion of existing information, when the entropy is maximized, it can be determined that the least amount of unknown information is included, thus reducing the uncertainty brought by unknown information. Assuming a point exists in a multidimensional natural environmental space, associated with multiple environmental factor parameters, it becomes possible to extract the environmental factor parameters linked to all distribution points of a species within this space. By applying specific algorithms, the ecological requirements of the species can be inferred, and the results can then be projected onto different temporal and spatial geographic regions to predict the potential distribution range of the species in a particular area (Elith et al., 2011). A formal explanation of the MaxEnt model based on this principle can be found in Supplementary Material . Phillips et al. (2004) developed the MaxEnt software for predicting species’ potential distributions on the Java platform. This software establishes constraints based on the environmental variable characteristics of actual geographical distribution points of the species and subsequently explores possible distributions under the principle of maximum entropy. The probability distribution of species presence at maximum entropy is anticipated to closely align with the species’ actual distribution. The process for utilizing the MaxEnt model involves several key steps: first, the collected occurrence data is inputted in CSV format into the “Samples” module, organized by species name, longitude, and latitude. Next, relevant environmental factor variables are standardized in terms of boundaries, coordinate systems, and raster sizes using GIS, and then converted into ASC layer format for input into the “Environmental layers” module, where environmental variables can be designated as either discrete or continuous. Following this, the user sets the number of iterations (typically 10) or the proportion of validation data (20%-30%) and determines whether to generate response curves for evaluating model accuracy, as well as whether to employ the jackknife method for selecting dominant environmental factors (Phillips, 2017). Finally, the output results are interpreted to derive insights from the model’s predictions. Species distribution models (SDMs) utilize options and algorithm settings that influence model complexity to describe the relationship between occurrence data and environmental variables. However, default settings often lead to overly complex models and overfitting of training data, resulting in distorted estimates of suitability and poor transferability to other locations or times (Soley-Guardia et al., 2024). Therefore, before using the MaxEnt model to predict the suitable habitat for Arabica coffee in Yunnan Province, we optimized it using the R package kuenm. The analysis began with input data, including thoroughly filtered and sparse species occurrence records to calibrate a complete event set, randomly partitioned into training event subsets and a 30% subset of events for testing candidate models (Cobos et al., 2019), along with environmental variables for creating candidate models. Additionally, a set of Arabica coffee occurrence records from Yunnan Province, collected from papers in WOS that were not used in the calibration process, was used as fully independent occurrence data to test the final model. This dataset had a different source from the other datasets and had no spatial autocorrelation with the calibration data. A range of 0.1 to 4, with an interval of 0.1, for the regularization multiplier (RM), and feature combinations (FC) including linear (L), quadratic (Q), product (P), threshold (T), and hinge (H), were cross-combined to explore parameter interactions. A total of 1240 candidate models were evaluated, reflecting all combinations of 40 regularization multiplier settings, 31 feature class combinations, and 1 distinct set of environmental variables. Model performance was assessed based on statistical significance (Partial ROC), omission rates (OR), and the Akaike information criterion corrected for small sample sizes (AICc). Finally, significant models with omission rates ≤5% and Delta_AICc=0 were selected, indicating the best model mobility from known distribution areas to predicted areas and effectively avoiding overfitting (Warren et al., 2014; Warren and Seifert, 2011). Species habitat suitability prediction The optimal parameter combination was inputted into MaxEnt for modeling and the average value was chosen as the model output result after 10 repetitions. The impact of environmental factors on the model was determined using the jackknife test method, and predictive charts were plotted to evaluate the importance of environmental variables. During the model-building process, the accuracy of the model was verified using the Area Under the Curve (AUC) value in the jackknife test. AUC represents the area under the Receiver Operating Characteristic (ROC) curve and theoretically ranges from 0.5 to 1. Higher values indicate higher predictive accuracy, with values between 0.9 and 1 indicating excellent predictive performance. The model generated logical output (LO) values ranging from 0 to 1, representing the suitable index for different regions. The selection of thresholds is crucial for predicting different levels of suitability areas, thus affecting the calculation of different suitable regions. We selected the Maximum Training Sensitivity plus Specificity (MTSS) and the balance of training omission, predicted area, and threshold value (TPT) as classification thresholds for suitable and less suitable areas, respectively, to reclassify LO. As a result, the potentially suitable habitat for Arabica coffee in Yunnan Province was classified into three levels: suitable, less suitable, and unsuitable habitats (Clark et al., 2014; Liu et al., 2013). Future potential habitat prediction under different climate scenarios By altering future bioclimatic variables, this study created potential habitats for Arabica coffee in Yunnan Province under three different Shared Socioeconomic Pathways (SSPs): SSP2-4.5, SSP3-7.0 and SSP5-8.5, for four future periods (2021-2040, 2040-2060, 2060-2080 and 2080-2100, i.e., 2030s, 2050s, 2070s and 2090s). Using the Distribution Changes Between Binary SDMs tool in SDM Toolbox v2.6 within ArcGIS 10.8, this study further simulated the spatial pattern changes including range expansion, unoccupied areas, no change, and range contraction to analyze the impact of climate change on the size of potential habitats for the species (Brown, 2014). Simultaneously, considering the suitable areas for specific scenarios and periods as a whole, they were simplified into vector particles, with changes in the central point position reflecting the trend of habitat shifts due to climate change. The Centroid Changes tool was used to calculate the coordinate changes and migration distances of current and future suitable habitats (Liang et al., 2021). Construction of the Marxan model The Marxan model is primarily designed to address the minimum set cover problem which aims to protect the minimum number of conservation targets that must be protected at the lowest cost and make it easier to establish protected areas without disrupting existing community continuity (Klein et al., 2009). The Marxan software uses a minimum coverage set model and incorporates a simulated annealing algorithm to complete the protected area siting analysis with the following operational objective function (Watts et al., 2009). Considering that the government controls land use through spatial planning, this study defines the township administrative area as the smallest planning unit which reflects the human-land relationship in coffee cultivation for convenient management. After removing administrative regions with excessively small areas, the study area is divided into a total of 1392 planning units. Using the zoning statistics tool in ArcGIS 10.8, the existence probability of the target species in each planning unit under current climatic conditions is calculated to construct a species distribution feature matrix. ∑PUsCost is a cost factor that represents the total cost of the selected planning units in the program. We established its values for the ratio of the area of land use/land cover (LULC) types where coffee cultivation is prohibited and the area of nature reserves to the total area of the planning unit. The LULC data is derived from the 30-meter annual China Land Cover Dataset (CLCD) (Yang and Huang, 2021), selecting LULC types such as cropland, water, snow/ice, barren, impervious, and wetland for the year 2019. Natural protected area data are obtained from Key Biodiversity Areas (KBA, https://www.keybiodiversityareas.org), The World Database on Protected Areas (WDPA, https://www.protectedplanet.net/), and the China Natural Protected Area Specimen Resource Sharing Platform (PAPC, https://www.papc.cn). BLM is the boundary length modifier, ∑PUsBoundary is the boundary factor, characterizes the total boundary length of the selected planning units in the scheme, and together they can regulate the compactness of the selected planning units; ∑ConvalueSPF×Penalty is a penalty factor for species conservation, SPF is the penalty for a single planning unit that does not accomplish the protection goal, and Penalty is the penalty coefficient. By adjusting the size of the SPF value, the size of the objective function can be controlled when the protection goal is not achieved, so that the selected planning units can accomplish the protection goal as much as possible. Based on the requirements of the Yunnan Province Provincial Government’s Action Plan for Agricultural Modernization, the conservation target is set at 30% and the forward target is set at 50%. CostThresholdPenalty(t) is a penalty factor when the upper limit of the cost threshold is exceeded. The Species Penalty Factor (SPF) and Boundary Length Modifier (BLM) are used in the model to balance the conservation of species in planning and the length of protected area boundaries. Sensitivity analysis is conducted by plotting sensitivity curves to find the most suitable values for BLM and SPF. Finally, BLM = 1.6 and SPF = 1.6 are determined, and the model is iterated 100 times to obtain the optimal solution for planning units. The specific roles of the parameters and the sensitivity analysis process for determining them are detailed in the Supplementary Material . The conceptual and methodological flowchart of this study is shown in Figure 3 . Conceptual and methodological flowchart. Results Model optimization and accuracy assessment Based on 56 species distribution records and 17 environmental variables, the potential suitable habitat distribution of Arabica coffee in Yunnan Province was predicted. Among the 1240 model results, combinations that were statistically significant and met the criteria for model omission rates and AICc were selected: RM set to 1.5 and FC combination set to QTH. The model performance is shown in Supplementary Figure S4 . The model was run 10 times using this combination setting, resulting in an average training AUC for the replicate runs of 0.935 ± 0.013 ( Supplementary Figure S5 ). Among these runs, the model with the highest test gain value both exceeding 0.9 had an AUC of 0.938 for the training data and 0.968 for the test data ( Supplementary Figure S6 ). This indicates excellent predictive accuracy of the model, with high confidence in the predicted results. Main environmental variables The importance of environmental variables is mainly assessed by the percent of contribution (PC) and permutation importance (PI) indices where higher values indicate greater importance (Zhang et al., 2021a). Based on the MaxEnt simulation results for Arabica coffee, 7 environmental variables had very small contributions or permutation importance values (<1%), hence only 10 dominant environmental variables are provided ( Table 2 . Dominant environmental variables for the potential distribution of Arabica coffeeDominant environmental variables for the potential distribution of Arabica coffee). The top 5 factors by percent of contribution are the minimum temperature of the coldest month (bio6, 20.8%), altitude (16.8%), mean temperature of the wettest quarter (bio8, 14.7%), slope (14.6%), and alum_sat (12.2%), accounting for a cumulative contribution of 79.1%. The top 5 factors by permutation importance are altitude (32.2%), mean temperature of the wettest quarter (bio8, 19%), pH water (11.5%), temperature seasonality (bio4, 7.9%) and precipitation of driest quarter (bio17, 5.5%), accounting for a cumulative permutation importance of 76.1%. In the Jackknife plot ( Supplementary Figure S7 ), when analyzing with only one variable, the altitude had the best-regularized training gain value, followed by the mean temperature of the wettest quarter (bio8), precipitation of the warmest quarter (bio18) and minimum temperature of the coldest month (bio6), all with values exceeding 0.9 which significantly higher than other factors, indicating their substantial influence on the distribution of Arabica coffee in Yunnan Province. Therefore, the minimum temperature of the coldest month (bio6), mean temperature of the wettest quarter (bio8), and altitude are the dominant factors influencing the potential distribution of Arabica coffee in Yunnan Province, while precipitation of the warmest quarter (bio18), slope and aluminum saturation (alum_sat) also play important roles. Table 2Dominant environmental variables for the potential distribution of Arabica coffee.VariablePercent of contribution(%)Permutation importance(%)bio620.80.7altitude16.832.2bio814.719slope14.64.3alum_sat12.20.7bio1251.9bio43.57.9bio173.25.5bio182.73.8ph_water2.211.5 Dominant environmental variables for the potential distribution of Arabica coffee. The response curves of environmental factors reflect the dependency of suitability on variables (Zhang et al., 2021a). Based on the above results, single-factor modeling was conducted and single-factor response curves were plotted ( Supplementary Figure S8 ) to study the relationship between distribution probability and major environmental variables. It is generally believed that when the distribution probability is greater than 0.5, the corresponding environmental factor is suitable for plant growth (Wang et al., 2023). The range which is suitable for the survival of Arabica coffee of the lowest temperature of the coldest month (bio6) and the mean temperature of the wettest quarter (bio8) is approximately between 5 to 10°C and 23 to 25°C. Additionally, the temperature seasonality (bio4, standard deviation ×100) should be less than 450. For precipitation, annual precipitation (bio12) above approximately 1200mm is suitable for the growth. Arabica coffee in Yunnan Province is most suitable at altitudes around 1200m with the range approximately between 680m to 1750m while the slope should be less than 9°. The distribution probability of Arabica coffee decreases with increasing aluminum saturation (alum_sat) among the soil variables, indicating that soil with high aluminum content may adversely affect coffee growth. Current and future potential suitable habitat distribution Currently, the potential suitable habitat distribution area for Arabica coffee in Yunnan Province is illustrated in Figure 4 . Both suitable and sub-suitable areas are primarily concentrated in the southern and western regions of Yunnan Province, particularly in the Hengduan Mountains region where the elevation is below 1500 meters. The predominant climatic types in these areas are the South Asian tropical and North tropical regions. The distribution areas of suitable habitats and their proportion to the total study area are presented in Table 3 . Potential distribution areas under current climate conditions. The suitable and sub-suitable areas cover approximately 42,124.44 km2 and 138,732.46 km2, accounting for 47.15% of the total provincial area. Among them, the suitable habitat area accounts for approximately 11% of the total area of Yunnan Province. The suitable habitat areas are mainly distributed in cities (or prefectures) such as Baoshan, Dehong, Lincang, Pu’er, Xishuangbanna, Honghe, and Yuxi with their combined area accounting for about 10% of the total area of Yunnan Province. A smaller portion is distributed at the border between Dali and Lijiang, the northern part of Chuxiong, as well as Kunming and Qujing. In Pu’er and Xishuangbanna, the suitable habitat areas account for over 30% of the total city area while in Honghe and Yuxi the suitable habitat areas are mainly distributed along the Ailao Mountain range. The entire area of Wenshan has a suitable/sub-suitable habitat coverage of 75.56%, but the majority is a sub-suitable habitat with suitable habitat occupying only 4.08% of the total area. Predicted distribution of Arabica coffee in Yunnan Province under current climate condition. Table 3Potential distribution areas under current climate conditions.CityUnsuitable habitat(km2)Sub-suitable habitat(km2)Suitable habitat(km2)Percentage of suitable areas in the city(%)Percentage of suitable areas in Yunnan Province(%)