Natural Vegetation Succession Under Climate Change and the Combined Effects on Net Primary Productivity

Climate change and the resulting natural vegetation succession can alter vegetation productivity. However, the mechanisms underlying future productivity changes under the two influences remain unclear. Here, we used the comprehensive sequence classification system to simulate changes in global potential natural vegetation under different climate scenarios (SSP1‐2.6, SSP2‐4.5, SSP3‐7.0, and SSP5‐8.5), and combined the Carnegie–Ames–Stanford Approach model with random forest to assess the response of net primary productivity (NPP) to climate change and vegetation succession from 2020 to 2100. Except for SSP126, terrestrial NPP in 2100 decreased by 0.86, 2.39, and 2.54 Pg C·a−1 versus 2020 under SSP2‐4.5, SSP5‐8.5, and SSP3‐7.0, respectively. Forest was the primary contributor to terrestrial NPP changes. The total forest area was projected to increase under all scenarios, with SSP2‐4.5 showing the largest increase (358.57 × 104 km2). However, expanding forest regions exhibited a relatively low mean NPP, while stable regions demonstrated a declining pattern. Consequently, forest NPP increased under SSP1‐2.6 but decreased by 4.03, 3.43, and 0.82 Pg C·a−1 in 2100 versus 2020 under SSP5‐8.5, SSP3‐7.0, and SSP2‐4.5, respectively. In comparison, grassland and desert exerted minor influence on terrestrial NPP changes, their total NPP decreased only under the SSP1‐2.6 scenario. The grassland area decreased, but the mean NPP increased, whereas the desert area expanded, resulting in consistent changes in both total and mean NPP. Our results analyzed the effects of climate change and vegetation distribution under its influence on the change of NPP, which can deepen our understanding of their relationship.

inhibit it (Ge et al., 2021;Ma et al., 2022;Piao, Liu, et al., 2019).The distribution of global vegetation and NPP is substantially influenced by precipitation, solar radiation, temperature, and their combined effects (Lucht et al., 2006;Reichstein et al., 2013).Therefore, accurately predicting the potential consequences of climate change on terrestrial vegetation ecosystems is a critical component of developing a comprehensive response to climate change.Therefore, accurately predicting potential scenarios in terrestrial vegetation ecosystems is crucial for addressing climate change.
Potential natural vegetation (PNV) encompasses the vegetation that has developed through natural succession, devoid of human influence or intervention and serves as a reliable depiction of the actual relationship between climate and vegetation (Chiarucci et al., 2010;Somodi et al., 2012).Models for simulating and predicting natural vegetation, such as the Dynamic Global Vegetation Model, BIOME, and Holdridge Life Zone, have gained extensive utilization in the study of climate change impacts on the geographic and ecological distribution of vegetation (Sitch et al., 2003;Yue et al., 2011;Zhao et al., 2021), as well as to elucidate the effects of climate change to terrestrial vegetation at different spatial scales (Holdridge, 1947;Salzmann et al., 2008).However, owing to the differences and complexities of input data quality and features, and the emphasis on vegetation classification, the widespread application of these models remains challenging (Kickert et al., 1999).In contrast, the comprehensive sequential classification system (CSCS) categorizes natural vegetation according to specific hydrothermal conditions in combination with vegetation, soil, and bioclimatic characteristics (Ren et al., 2008).Several studies have verified the reliability of the CSCS in theory and practice (Li et al., 2021;Liang et al., 2012;Lin et al., 2013).In the pursuit of estimating vegetation NPP, researchers have developed numerous statistical, biological process, and light-use efficiency (LUE) models.Among these models, the Carnegie-Ames-Stanford Approach (CASA) of the LUE model has emerged as the most widely adopted and has consistently demonstrated satisfactory outcomes (Bao et al., 2016;Pei et al., 2013;Wu et al., 2022).In comparison to other models, the input parameters required by the CASA model are comparatively simple and more appropriate for studying the spatiotemporal variation in NPP at regional scales.However, owing to the limitations of the input parameters, CASA is unsuitable for directly estimating future NPP.In recent years, the random forest (RF) algorithm has gained considerable traction in the field of geoscience, emerging as a highly reliable method for simulating and estimating NPP (Baez-Villanueva et al., 2020;Hou et al., 2020;Ozigis et al., 2020;Tramontana et al., 2015).Therefore, by combining CSCS, CASA, and the RF algorithm, it is possible to overcome the limitations of the CASA model on its own and better reflect the impact of climate and vegetation types on NPP, thereby providing a new approach for simulating and estimating NPP.
As indicated in the Sixth Assessment Report of the Intergovernmental Panel on Climate Change, there is a projected ongoing global increase in greenhouse gas emissions, leading to the exacerbation of climate change worldwide, this escalation is anticipated to pose heightened risks to terrestrial ecosystems across all regions (Masson-Delmotte et al., 2021).Based on the representative concentration pathways, the Coupled Model Intercomparison Project Phase 6 (CMIP6) developed a fresh set of shared socioeconomic pathways (SSPs) that describe potential future societal developments in the case of no climate change policies being implemented.These four climate scenarios, SSP1, SSP2, SSP3, and SSP5, represent sustainable development, middle-of-the-road development, regional rivalry development, and fossil-fueled development, respectively (O'Neill et al., 2016;van Vuuren et al., 2012).CMIP6 can provide more reasonable predictions of regional and global climate change using the latest comprehensive assessment models and emissions data (Stouffer et al., 2017).However, the analysis of the mechanisms of changes in the global carbon sequestration capacity of terrestrial vegetation when the combined effects of climate change and resulting vegetation succession are considered simultaneously has been limited (Li et al., 2022;Wang et al., 2022;Yuan et al., 2021).Therefore, by considering the possible vegetation changes on a global scale under various climate scenarios, this aspect of research deficiency can be supplemented.
In this study, we employed the CSCS and integrated the CASA model with random forest algorithm to conduct a systematic analysis of the spatial pattern and NPP of global PNV.Our aim was to answer the following two questions:

Remote Sensing Data
The Normalized Difference Vegetation Index (NDVI) product obtained from the MODIS was utilized to compute the fraction of photosynthetically active radiation (FPAR) intercepted by the vegetation layer from 2011 to 2020 (Didan, 2021).The spatial and temporal resolution was 1 × 1 km and 1 month, respectively.MODIS NDVI supplements the Advanced Very High-Resolution Radiometer (AVHRR) NDVI product from the National Oceanic and Atmospheric Administration (NOAA), with better improved resolution and continuous time series.
The digital elevation model (DEM) data used as input for the RF algorithm were chosen from the global DEM data set provided by NOAA, which possesses a spatial resolution of 1 × 1 km (Hastings et al., 1999).The slope, elevation, aspect, longitude, and latitude were derived from the DEM using ArcGIS 10.6 (Esri, Redlands, CA, USA) and used as parameters for the RF model.

Meteorological Data
This study employed meteorological data from the Climatic Research Unit Time-Series version 4.06 (CRU_ TS_4.06), a gridded climate data set produced by the UK National Centre for Atmospheric Science.The time range was 2011-2020, including monthly average precipitation, as well as average minimum, average maximum, and average temperature.CRU_TS is one of the most widely used high-resolution monthly global gridded data sets, generated using the angular distance weighting method, and has a spatial resolution of 0.5° × 0.5°, covering all land areas of the world, except Antarctica (Harris et al., 2020).Solar radiation was acquired from the Global Land Surface Satellite Downward Shortwave Radiation data set (GLASS), which is accessible through the National Earth System Science Data Center, National Science & Technology Infrastructure of China (Zhang et al., 2014(Zhang et al., , 2019; X. T. Zhang et al., 2016), and the spatial and temporal resolution of the acquired data is 0.05° × 0.05° and 1 d, respectively.
Data on future meteorology were acquired from 16 Earth System Models (ESM) containing all the data we used, which simulated four different future climate scenarios (SSP1-2.6,SSP2-4.5,SSP3-7.0, and SSP5-8.5)as part of the CMIP6 (Riahi et al., 2017).Monthly shortwave radiation, precipitation, as well as the average minimum, average maximum, and average temperature from the r1i1p1f1 ("r" for realization, "i" for initialization, "p" for physics, and "f" for forcing) series of each model were included (Petrie et al., 2021).Table S1 in Supporting Information S1 lists the basic characteristics of all ESMs.
In this study, the bilinear interpolation was used to resample all data to a 0.5° × 0.5° grid.In addition, all climate prediction model data under the four climate scenarios were subjected to a multi-model ensemble mean (MME) process (Sun et al., 2023), and the arithmetic mean of all model simulation results, was used as future meteorological data.

Analysis of Future Changes in Meteorological Factor Trends
To analyze the effects of future climate change on the PNV, we calculated the annual rate of change for three meteorological factors from 2020 to 2100 using the "regress" function in MATLAB 2018 with a one-dimensional linear regression approach (Chatterjee & Hadi, 1986;Chen et al., 2019) and examined the inter-annual trends of these three meteorological factors.This analysis was conducted at the pixel scale and utilized the MME results for mean annual precipitation, mean annual temperature, and total annual solar radiation from 2020 to 2100 across the four climate scenarios.

Global Terrestrial PNV Classification by CSCS
To simulate the global PNV types, we used the CSCS, a classification system that groups or clusters units with similar thermal and humidity characteristics (Ren et al., 2008).The CSCS comprises three basic classification fundamental tiers: class, subclass, and type.Among them, the class is primarily determined by bioclimatic conditions, which is the elementary unit of the CSCS and the core of the system.With class used as the basic unit, classification is primarily based on the zonal characteristics of the bioclimates, which are determined by the quantitative bioclimatic index that combines the humidity index (K) and annual accumulated temperature above 0°C (Ʃθ).
where 0.1 is an empirical parameter and MAP represents the mean annual precipitation.GDD0 is the growing degree days above 0°C; MATmin i and MATmax i represent the mean minimum and maximum temperatures, respectively, during the month denoted by i; T a is 0°C; MD i represents the total number of days in month i; and fun(x) evaluates a judgment condition: if x is not less than 0, fun(x) = x; otherwise, fun(x) = 0.
Based on the GDD0 and K, the CSCS classified seven thermal zones and six humidity zones, and 42 PNV types could be classified by combining the two (Figure S1 and Table S2 in Supporting Information S1).Based on the similarity of ecological characteristics, these types were classified into 10 broader vegetation categories (Table 1), which reflected the global large-scale spatial distribution of PNV more explicitly (Liang et al., 2012;Ren et al., 2008).In this study, the PNV was classified based on these 10 broad categories.The MAP was statistically derived based on the monthly precipitation of the MME under the four scenarios and of CRU_TS_4.06,respectively.The global terrestrial PNV distributions for 2020-2100 under the four scenarios and for 2011-2020 were simulated using the MAP as well as the monthly average minimum and maximum temperatures as input terms, respectively.In addition, we calculated the distribution area of each type of terrestrial PNV under each scenario and the proportion of the total land area based on latitudinal and longitudinal differences and in terms of pixel units, and the changes were analyzed.

NPP Simulation of Global Terrestrial PNV by RF Combined With CASA
The RF machine learning algorithm was used to simulate the NPP of the global terrestrial PNV under four scenarios in this study.RF is a typical integrated learning method in machine learning that improves regression trees using a large number of decision trees for classification and regression (Breiman, 2001).By constructing and combining multiple decision trees and averaging their results, the RF regression model converges the generalization errors of the decision trees to produce better prediction results.RF is simple to apply and can be optimized by adjusting the number of regression trees and predictive variables at each node, thereby improving prediction accuracy.We used the "RandomForestRegressor" function in the Python 3.9 (Python, Fredericksburg, VA, USA) "scikit-learn" package to build, optimize, and predict the RF regression model.For each of the 10 PNV categories, an individual RF regression model was developed in this study.In training and validating the model, we used the elevation, slope, slope direction, longitude, and latitude data extracted from the DEM from NOAA as the static independent variables and five climate drivers, namely, solar radiation from GLASS as well as annual mean temperature, maximum temperature, cumulative temperature above 0°C and precipitation from CRU_TS_4.06 for the years 2011-2020, were considered as the dynamic independent variables.In addition, we  The NPP estimation results from CASA were used as the dependent variable input for the RF model.
The CASA model calculates NPP using the following equation, which relies on the values of FPAR and LUE (Field et al., 1995;Potter et al., 1993): where SOL represents the total solar radiation that reaches the surface of the Earth, ε max is the maximum LUE under optimal growth conditions, W ε (x, t) is the extent to which moisture conditions limit LUE, T 1 (x, t) and T 2 (x, t) are represent the limiting factors of LUE at extreme temperature and deviation from optimal temperature, respectively.Potter et al. (1993) postulated that the maximum LUE of global vegetation is 0.389 g C•MJ −1 .Within a certain range, FPAR is dependent on vegetation type and cover and can be quantified by the NDVI simple ratio (SR): where SRmin represents the SR value for bare soil and SRmax represents the SR value when vegetation intercepts all incoming SOL that varies by vegetation type.These values, ranging from 4.14 to 6.17, according to Sellers et al. (1996) can correspond to the vegetation types classified by CSCS (Table S3 in Supporting Information S1).
In the original CASA model, the determination of Wε involved the utilization of the actual-to-potential evapotranspiration ratio.M. Zhang et al. (2016) excluded several soil parameters, used Ʃθ and K to calculate the water stress coefficient, combined CSCS and CASA model, and obtained good results for NPP estimation.This improved water stress coefficient was used in this study.
To evaluate the accuracy of the RF model simulation, the data were divided randomly into a training data set comprising 80% of the total data, and a validation data set consisting of the remaining 20%.The best RF model was determined through parameter adjustment.Subsequently, we simulated the NPP of the global terrestrial PNV for 2020-2100 under the four scenarios by replacing the dynamic independent variables in each model with the corresponding data from the MME.In addition, when calculating each terrestrial PNV as well as the global mean NPP and total NPP, we combined the area of each pixel to weight the calculations separately to minimize the error.

Changes in Climate Factors From 2020 to 2100 Under Different Scenarios
Different scenarios resulted in varying trends for average temperature, precipitation, and total solar radiation (Figure 1).The average temperature of PNV growth regions showed an increasing trend, with the warming rate increasing from south to north under all four scenarios.The fastest temperature increase and the most significant temperature amplitude were observed under SSP5-8.5.The southern region of South America showed a slight temperature increase.In contrast, the northern hemisphere's high-latitude tundra and alpine steppe showed noticeable upward trends, with the highest rate exceeding 0.17°C a −1 under SSP5-8.5; the highest rate of upward also exceeded 0.05°C a −1 under SSP1-2.6.A majority of the global regions showed an upward trend in precipitation, whereas precipitation trends decreased in Australia, the northern section of South America, and the Mediterranean coast.The changing trend of precipitation under SSP5-8.5 was highly significant, regardless of whether it increased or decreased, with maximum rates exceeding 24 mm a −1 and 7 mm a −1 for increase and decrease, respectively.Unlike for average temperature and precipitation, the trends of total solar radiation displayed significant differences among the four climate scenarios.Solar radiation demonstrated an overall upward trend in the majority of regions under SSP1-2.6 and SSP2-4.5.Apart from some areas in southeastern Asia, southern Africa, western Europe, and northern South America, solar radiation decreased in most regions under SSP3-7.0 and SSP5-8.5.

Change in Area Proportion of Terrestrial PNV From 2020 to 2100
Forests are the most widely distributed type of terrestrial vegetation, covering greater than half of the total land surface (Figure 2; Figure S2 in Supporting Information S1).Our results indicate that the overall forest area is set to expand from 2020 to 2100.Forest acreage had the largest increase under SSP2-4.5 (358.57× 10 4 km 2 ), while the proportion of forests reached its maximum in the 2050s and then declined gradually under SSP5-8.5, which showed the lowest increasing trend among the scenarios (138.02× 10 4 km 2 ) (Figure S3a in Supporting Information S1).Temperate forest was the most widely distributed and showed the greatest increase in area among the forest types.Temperate forest area showed the greatest increase under SSP2-4.5 (298.32 × 10 4 km 2 ), while under SSP5-8.5, it showed the smallest increase (0.22 × 10 4 km 2 ) due to a decreasing trend observed between 2060 and 2100.At the same time, tropical forest acreage also had an upward trend, but the increase was relatively small compared to that of temperate forest.Tropical forest acreage had the largest increase (152.97 × 10 4 km 2 ) under SSP5-8.5,whereas the smallest increase occurred under SSP1-2.6 (45.17 × 10 4 km 2 ).In contrast, subtropical forest acreage revealed a decreasing trend.The subtropical forest acreage decreased by 50.86 × 10 4 , 75.10 × 10 4 , 89.22 × 10 4 , and 15.18 × 10 4 km 2 under SSP1-2.6,SSP2-4.5, SSP3-7.0, and SSP5-8.5, respectively.The trend of change in subtropical forest area was comparable between the four scenarios, displaying a consistent pattern of first declining and then increasing.
Grasslands were the second largest type of PNV and are expected to shrink from 2020 to 2100.The highest decrease in grassland area (380.08 × 10 4 km 2 ) occurred under SSP2-4.5, and the lowest decrease under SSP5-8.5 (244.38 × 10 4 km 2 ) (Figure S3b in Supporting Information S1).Climate warming and humidification trends are expected to cause a substantial decrease in tundra and alpine steppe areas, with decreases of 380.52 × 10 4 and 1068.03× 10 4 km 2 under SSP1-2.6 and SSP5-8.5, respectively.The temperate humid grassland acreage also decreased.As the most widely distributed type of grassland, savanna area is expected to show significant expansion during the 21st century.The largest increase in area was observed under SSP5-8.5 (920.03 × 10 4 km 2 ), and the smallest increase occurred under SSP1-2.6 (166.81 × 10 4 km 2 ).Steppe was the grassland type with the least change in total area, with an increase of 13.57 × 10 4 and 31.18× 10 4 km 2 under SSP2-4.5 and SSP5-8.5, and a reduction of 18.45 × 10 4 and 12.18 × 10 4 km 2 under SSP1-2.6 and SSP3-7.0,respectively.
Deserts had the lowest distribution globally, and the area of deserts varied under the different scenarios from 2020 to 2100.By 2100, the desert area is expected to increase by 81.12 × 10 4 and 106.36 × 10 4 km 2 under SSP3-7.0 and SSP5-8.5 and reduce by 6.47 × 10 4 and 13.60 × 10 4 km 2 under SSP1-2.6 and SSP2-4.5, respectively (Figure S3c in Supporting Information S1).However, desert area showed an upward tendency generally under all scenarios from 2020 to 2100.Of three types of deserts, the proportion of frigid desert was the smallest and showed a decreasing trend in all four scenarios, with the greatest decrease (48.95 × 10 4 km 2 ) occurring under SSP5-8.5.Semi-desert area also decreased, with a trend consistent with that of frigid desert.Warm desert was the most widespread desert type, covering about 80% of the total desert area.Under SSP1-2.6,SSP2-4.5, SSP3-7.0, and SSP5-8.5, warm desert area showed an increasing trend, and the increases were 84.63 × 10 4 , 90.12 × 10 4 , 217.75 × 10 4 , and 308.58 × 10 4 km 2 , respectively.Overall, the expanding warm desert areas will offset the shrinking semi-desert and frigid desert regions, resulting in a general increasing trend in total desert area during the 21st century.

Validation Results for Each RF Model
The mean NPP simulated by the RF models showed a strong positive correlation with the test set (P < 0.001) (Figure 3).The overall accuracy for the mean simulated NPP was highest for the forest PNV.Although the RF model's simulation accuracy for mean desert NPP was lower than that for forests and grasslands, a strong positive correlation was still observed (R 2 > 0.68).These results indicate that the RF models had high precision and satisfactory performance to simulate the PNV NPP.

Changes in Mean and Total NPP of Terrestrial PNV From 2020 to 2100
The changes in mean NPP varied between various vegetation categories under the four scenarios from 2020 to 2100 (Figure 4).The mean NPP of forests decreased under all four scenarios, with the lowest decrease amount (18.49g C•m −2 •a −1 ) occurring under SSP1-2.6 and the highest (67.64 g C•m −2 •a −1 ) under SSP3-7.0(Figure S4a   S4 in Supporting Information S1.

10.1029/2023EF003903
10 of 19 all four scenarios, so the overall trend was not obvious.Savanna showed the highest mean NPP among the four grasslands, which increased under all four scenarios, with the highest increase amount (18.78 g C•m −2 •a −1 ) under SSP5-8.5.In contrast, under SSP2-4.5, the mean NPP decreased in 2100 compared to that in 2020 due to a decline after 2090.As for deserts, the mean NPP is projected to decrease only under SSP1-2.6 and increase under the other three scenarios (Figure S4c in Supporting Information S1).The mean NPP of frigid desert was the lowest among the three deserts, and it decreased under SSP5-8.5 and SSP1-2.6.The mean NPP of semi-desert decreased only under SSP1-2.6,with minor fluctuations, and increased under the other scenarios, with SSP3-7.0 showing the maximal increase amount (21.44 g C•m −2 •a −1 ).Warm desert exhibited the highest mean NPP among three desert types, with upward trends observed under SSP1-2.6,SSP2-4.5, SSP3-7.0, and SSP5-8.5, with increase amounts of 0.79, 1.71, 8.47, and 10.05 g C•m −2 •a −1 , respectively.
From the perspective of the total amount of NPP, under SSP1-2.6,total terrestrial NPP is expected to show an increase from 2020 to 2100 (Figure 5), but the increase amount would be relatively small (0.43 Pg C•a −1 ).However, under the other three scenarios, the terrestrial NPP exhibited a decreasing trend, the lowest decrease amount of 0.86 Pg C•a −1 occurred under SSP2-4.5, while the highest occurred under SSP3-7.0(2.54 Pg C•a −1 ) from 2020 to 2100.
The total NPP of forests, grasslands, and deserts exhibited varying trends from 2020 to 2100.Forests contributed to over 70% of the total terrestrial NPP, and except for a slight increase under SSP1-2.6, it decreased under the other three climate scenarios (Figure S5a in Supporting Information S1).Tropical forest contributed the most to forest NPP, and the trend of NPP was similar to overall forests, with more pronounced decrease amounts of 1.25 and 1.11 Pg C•a −1 under SSP3-7.0 and SSP5-8.5, respectively.Temperate forest ranked second in total NPP and is anticipated to rise under SSP2-4.5 and SSP1-2.6 but decrease under SSP5-8.5 and SSP3-7.0, with the largest decrease amount (1.32 Pg C•a −1 ) under SSP5-8.5.Under all four scenarios, the NPP of subtropical forest exhibited a decrease, with the greatest decrease amount under SSP3-7.0.Although the grasslands NPP under SSP1-2.6 is expected to increase after 2020, it showed an overall downward trend, whereas an upward trend was observed under three additional scenarios (Figure S5b in Supporting Information S1).Among them, the total savanna NPP comprised the largest proportion of grassland NPP and was predicted to increase by 0.89, 1.07, 2.39, and 3.27 Pg C•a −1 in 2100 compared to 2020 under SSP1-2.6,SSP2-4.5, SSP3-7.0, and SSP5-8.5.The total steppe NPP presented a fluctuating and decreasing trend under SSP3-7.0 and SSP1-2.6,whereas both SSP5-8.5 and SSP2-4.5 presented a decrease followed by an increase and an overall increasing trend.The largest decrease in total grasslands NPP occurred in the tundra and alpine steppe, with decrease amounts of 1.66, 1.51, 1.03, and 0.55 Pg C•a −1 under SSP5-8.5,SSP3-7.0,SSP2-4.5, and SSP1-2.6,respectively.The total temperate humid grassland NPP decreased only slightly.The NPP of deserts, which represents the smallest proportion of the total terrestrial NPP, exhibited similar changes as those of grasslands under all four scenarios (Figure S5c in Supporting Information S1).Cold desert NPP decreased under all four scenarios, with the largest and smallest decrease amounts occurring under SSP5-8.5 and SSP1-2.6,respectively.The semi-desert NPP decreased under three of the scenarios, while showing an overall upward trend under SSP3-7.0due to an increase after 2060.In contrast, the warm desert NPP was predicted to increase by 0.04, 0.05, 0.22, and 0.29 Pg C•a −1 in 2100 compared to 2020 under SSP1-2.6,SSP2-4.5, SSP3-7.0, and SSP5-8.5, respectively.

Response of Terrestrial PNV Distribution to Changes in Climatic Factors
Temperature, precipitation, solar radiation, and their spatial dynamics are important drivers of terrestrial PNV distribution and variations in NPP.Based on the average results of the multi-model ensemble, under SSP1-2.6,SSP2-4.5, SSP3-7.0 (Figures S6-S8 in Supporting Information S1) and SSP5-8.5 (Figure 6), the extent of vegetation in high-latitude and high-altitude areas, for example, tundra and alpine steppe, and cold desert, will continuously shrink and migrate to higher latitudes and altitudes.This result is in agreement with the conclusions drawn in earlier research (Gang et al., 2017;Lucht et al., 2006;Wookey, 2008).Predictions of terrestrial PNV changes at high latitudes in the Northern Hemisphere under a sharper temperature increase scenario are consistent with recent predicted vegetation changes in North America (Flanagan et al., 2016) and Europe (Gang et al., 2017;Kicklighter et al., 2014;Shiyatov et al., 2005).Bonannella et al. (2023) revealed a trend in which the polar/alpine biome trends to shift toward temperate-boreal forest biome, as indicted by integrated machine learning model simulations, this finding is consistent with the results of our study.By the conclusion of the 21st century, temperate forest is forecast to replace some of the original tundra and alpine steppe, together with temperate humid grassland, on a large scale (Figure S2 in Supporting Information S1).This trend is likely attributed to the ongoing increase in temperature throughout this century, resulting in milder winters and extended growing seasons, which are expected to weaken the distribution limitations of temperate forests in boreal regions (Moore, 1987).Simultaneously, increases in temperature, precipitation, and solar radiation are expected to promote a northward expansion of tropical forests and savannas, which are mainly distributed in tropical regions, replacing some of the original subtropical forest.Therefore, subtropical forest shows a shrinking trend before the 2060s.With the changing climate, subtropical forest vegetation is expected to replace some temperate forest vegetation in certain regions.In tropical forest regions, the projected sharp temperature increases have the potential to intensify droughts and increase the risk of wildfires (Gloor et al., 2013;Schwalm et al., 2017), for example, in South America, certain parts of tropical forest are expected to be replaced by savanna under the pressure of temperature increase, prolonged drought, and fire, among others.Predictions of tropical forest and savanna through integrated machine learning (Bonannella et al., 2023) and generalized linear (Anadon et al., 2014) models show a similar transition from tropical forests to Savanna.As the climate has warmed, decreased annual precipitation and increased solar radiation along the Mediterranean coast have led to increase in area of warm desert vegetation in the northern part of the Sahara and western coast of the Mediterranean, replacing the original savanna (Boko et al., 2007).The gradual replacement of some cold desert in Central Asia with warm deserts may also be the result of continuous warming.In southern Australia, the expansion of warm desert may result from the projected increases in prolonged droughts under future climate scenarios, which will promote the growth of desert vegetation (Risbey, 2011), thereby leading the replacement of unsuitable Savanna vegetation.

Discussion of the Mechanism of Total NPP Changes in Different Terrestrial PNV
Previous studies have shown that global terrestrial ecosystem NPP exhibits an increase in the 21st century under different scenarios, with increasing CO 2 concentration as the primary driver (X.Hu et al., 2021;Tharammal et al., 2019;Zhu et al., 2018), and that NPP has a high sensitivity to changes in CO 2 concentration (Qiu et al., 2023).However, climate change and land use/land cover (LULC) change have similarly important implications in ecosystem carbon sequestration (Krause et al., 2019).Pan et al. (2014) quantified the 21st century NPP using the Dynamic Land Ecosystem Model and found that the terrestrial NPP in 2090 decreased by 2.51 Pg C compared to 2020 when only climate change was considered under the high emissions scenario.X. Hu et al. (2021) used the Integrated Biosphere Simulator (IBIS) model to estimate the global terrestrial ecosystem NPP changes in the 21st century under three different climate scenarios, and found that the global NPP decreased under the influence of climate change and LULC.Similarly, Hou et al. (2022) found that global primary productivity (GPP) decreased under climate change, LULC, and both together when simulating GPP in the 21st century using the Common Land Model.In this study, we observed that except for SSP1-2.6, the total global terrestrial PNV NPP exhibited a decreasing trend from 2020 to 2100, indicating that future climate and vegetation succession may adversely affect the productivity of global vegetation.Rising temperatures and solar radiation could lead to more frequent droughts, which could severely inhibit vegetation growth (Cao et al., 2022).Elevated temperatures may exceed the optimal thresholds for plant growth and prolonged warming can lead to a reduction in vegetation productivity (Penuelas et al., 2017;Zhang et al., 2022).Moreover, the decrease in total NPP is primarily observed in vegetation with larger areas and biomass, such as tropical forest, whereas the increase is primarily observed in vegetation with small areas and biomass, like tundra and alpine steppe and savanna, this observation may explain the global NPP trend (Qian et al., 2019;Zhu et al., 2018).The dominant vegetation type driving variations in global NPP varied across the various scenarios.Generally, forests dominate the changes in global total terrestrial NPP (Figure 7).However, the mechanisms of change in total NPP for each PNV in response to the combination of area and mean NPP are still unclearly analyzed.
Regarding the tundra and alpine steppe, the mean NPP exhibited a consistent upward trend under all four scenarios from 2020 to 2100.This could be attributed to the warming temperature, which promote permafrost degradation and glacial melting, thereby replenishing soil moisture and alleviating drought (G.J. Hu et al., 2021).Additionally, plant photosynthesis can increase with reasonable temperature increases (Zhang et al., 2022).However, the substantial reduction in the tundra and alpine steppe areas, resulting from the large-scale replacement of temperate forest and other vegetation, ultimately led to a decrease in total NPP from 2020 to 2100.The mechanisms of change in total NPP for steppe under SSP1-2.6,frigid desert and semi-desert under SSP2-4.5,frigid desert under SSP3-7.0,and semi-desert and temperate humid grassland under SSP5-8.5 were the same as those for the tundra and alpine steppe.All of them were under the situation of mean NPP rising, and the decrease in area resulted in a decrease in total NPP.Only Semi-desert had an increase in total NPP under the SSP3-7.0scenario with a rise in mean NPP and a relatively small decrease in area.Although an increase in temperature can partly alleviates the inhibitory effect of low temperatures on NPP, in some areas (such as in the Amazon tropical forest), increases in temperature and solar radiation may aggravate the onset of droughts (Doughty et al., 2023;Lu et al., 2019;Xu et al., 2019;Yang et al., 2018), which inhibit vegetation growth and cause a decline in mean NPP.Our results are consistent with the predictions made by Pan et al. (2014) regarding the reduction of NPP at low latitudes.Moreover, although some subtropical forest vegetations in Asia and North America were replaced by tropical forest vegetations, the annual mean NPP in the regions where succession occurred remained lower than that of the original tropical forest (Figure 8, Figure S9-S11 in Supporting Information S1).Therefore, the mean NPP of tropical forest decreased during the expansion, ultimately resulting in a decline in the total NPP of tropical forests, except for SSP1-2.6, for the period of 2020-2100.The cold desert under SSP1-2.6,temperate forest under SSP3-7.0 and SSP5-8.5, and subtropical forest after 2060 under the four scenarios exhibited a similar mechanism of change in total NPP.In contrast, mean NPP decreased in the temperate forest and tropical forest under SSP1-2.6 and in the steppe and temperate forest under SSP2-4.5,however, total NPP increased owing to the continued increase in area.For savanna, the expansion primarily involved the replacement of forests vegetation (Figure S2 in Supporting Information S1).Therefore, as the area of savanna expanded, the mean NPP also exhibited an increasing trend, ultimately resulting in an overall increase in total savanna NPP from 2020 to 2100.Similarly, during the gradual replacement of Savanna, the total NPP in warm desert followed a similar change mechanism, as both vegetation area and mean NPP positively contributed to the change in total NPP.In addition, the decrease in mean NPP in subtropical forests from 2020 to 2100 is consistent with the predictions made by Pan et al. (2014) and may be attributed to enhanced autotrophic respiration and increased water stress due to rising temperatures (Cramer et al., 1999).Moreover, the subtropical forest continued to experience a reduction in area until 2060, and the combined adverse effects of both area and NPP contributed to a decline in total subtropical forest NPP from 2020 to 2060.Similarly, the total NPP during 2020-2100 for semi-desert, temperate humid grassland under SSP1-2.6,temperate humid grassland under SSP2-4.5,steppe under SSP3-7.0,temperate humid grassland, and frigid desert under SSP5-8.5 exhibited a similar mechanism of change in total NPP, all of which declining in response to a combined decrease in mean NPP and area.

Uncertainties
The classification of PNV types and simulation of future NPP in this research are subject to uncertainties.The CSCS was used to simulate PNV over time, and 42 PNV types worldwide were classified into 10 major categories to reduce simulation errors.However, this method did not consider the effects of groundwater and glacial snowmelt on precipitation supply and topography, which may reduce the accuracy of PNV simulations for some regions, such as mountainous and high-latitude high-altitude regions (Gang et al., 2017).In addition, owing to the absence of global measurements of NPP, the CASA model was applied to estimate historical NPP worldwide.Among NPP factors, indicators such as maximum LUE and maximum vegetation index differ for different vegetation types, and these parameters greatly impact the estimation of NPP, thus potentially increasing uncertainty in the simulation (Zhu et al., 2006).Additionally, there may also uncertainties in the input meteorological data owing to differences in the structure of the CMIP6 model and inconsistent spatial resolution, and because we directly used the MME results from 16 models.Some studies have shown that, in many regions, CMIP6 has a high degree of uncertainty in predicting precipitation (Collins et al., 2014), and the simulation of related climate events (such as drought) can be inconsistent with the observed data (Ukkola et al., 2020), which inevitably leads to a decline in the accuracy of the NPP simulation.To reduce the potential error caused by a single model, we used a multi-model ensemble average for the meteorological data; moreover, to simulate future NPP, we used the RF model combined with meteorological factors and DEM data.However, the use of machine learning in model training for annual variability remains questionable, particularly for tropical areas (Zhao & Running, 2010).
Future research should aim to address all of the aforementioned issues.In contrast, our method is understandable and reasonable for determining the allocation of future terrestrial PNV and conducting simulation with NPP, thus our results can unveil the reaction of PNV under future climate change and serve as reference for coping with future climate change.

Conclusion
We quantitatively assessed the dynamic response of terrestrial PNV distribution and NPP to climate change under various CMIP6 scenarios throughout the 21st century and analyzed the mechanisms by which the NPP changes under the combined effects of climate change and vegetation succession.Based on our analysis, we arrived at the following conclusions: (a) The total forest area increased with climate warming across all climate scenarios, while grasslands exhibited the opposite trend.By the end of the century, although different results were obtained under different scenarios, there was an overall increasing trend in the total area covered by deserts.(b) The global terrestrial NPP showed only a slight increase under SSP1-2.6,but decreased under SSP2-4.5,SSP3-7.0, and SSP5-8.5.Forests had the greatest impact on the variation of global NPP, while grasslands and desert displayed NPP trends opposite to those of the global PNV.(c) The NPP of terrestrial PNV changed under the combined effects of climate change and the resulting vegetation succession.For example, while climate change increased the mean NPP in tundra and alpine steppe, and semi-desert regions, the continual contraction of their distribution and replacement by more suitable PNV types resulted in a decrease in total NPP.Furthermore, for different PNV types, the reasons for changes in their NPP under the combined effects of climate change and vegetation succession were varied.

•
In the four climate scenarios, vegetation underwent mutual succession, yet all exhibited forest and desert expanded and grassland shrank • From 2020 to 2100, global terrestrial net primary productivity mainly declines when affected by climate change and vegetation succession • Net primary productivity of various vegetations has varying mechanisms of change in response to climate change and vegetation succession Supporting Information: Supporting Information may be found in the online version of this article.
(a) Assuming that future societies follow different paths, how will the global climate as well as the area, and distribution of terrestrial PNV change and; (b) If the NPP of future terrestrial PNVs is affected by the combination of climate change and its resultant vegetation succession, what are the changes and what are the mechanisms underlying this change?This study offers scientific evidence to substantiate the precise evaluation and forecasting of climate change's impact on terrestrial ecosystems.Moreover, it facilitates the informed development of climate adaptation and mitigation measures, thereby contributing to evidence-based decision-making in addressing the challenges posed by climate change.
of terrestrial PNVs using the CASA based on the 2011-2020 PNV types modeled by CSCS.

Figure 7 .
Figure 7. Annual rates of change in total NPP of global terrestrial vegetation, total NPP of all PNV types, and mean NPP and area of various PNV types over the period 2020-2100 under SSP1-2.6,SSP2-4.5, SSP3-7.0, and SSP5-8.5.From the innermost to the outermost layer, the first layer represents the annual rate of change in total NPP of terrestrial vegetation globally, second layer represents the annual rate of change in total NPP of all types of PNV, and third layer represents the annual rate of change in mean NPP and area of all types of PNV (the annual rates of change in all three layers are based on the years 2020-2100).(a) Tundra and alpine steppe, (b) Frigid desert, (c) Semi-desert, (d) Steppe, (e) Temperate humid grassland, (f) Temperate forest, (g) Sub-tropical forest, (h) Tropical forest, (i) Warm desert, (j) Savanna.

Table 1
Relationships Between Potential Vegetation Classes and Groups Delineated by the CSCS Model