Study of the Mechanisms Driving Land Use/Land Cover Change and Water Yield in the Ganjiang River Basin Based on the InVEST-PLUS Model

Abstract

Water yield is a critical component of hydrological ecosystem services, influenced by both natural environments and human activities. Changes in land use and land cover (LULC) are particularly pivotal in causing water yield variations at the basin level, particularly for the ecologically fragile Ganjiang River Basin (GRB) in southern Jiangxi province, China. Over the last 33 years, the GRB has undergone substantial LULC changes that have significantly affected its water yield. Initially, this study assessed water yield from 1990 to 2022 using the InVEST model, then predicted future LULC scenarios using the PLUS model, including natural development (ND), cropland protection (CP), ecological protection (EP), and urban development (UD). The Geodetector model was then employed to analyze the influence of various factors on water yield changes. Key findings include the following: (1) Significant landscape changes were observed, including increases in impervious surfaces, cropland, and water areas, accompanied by substantial reductions in forest and other natural lands. The most pronounced decline occurred in forested regions. (2) The total water yield decreased by 0.44 × 10¹⁰ m³ over the study period, exhibiting fluctuations until 2016 and stabilizing afterward. Water yield was generally higher in the northeast and lower in the southwest, primarily influenced by actual evapotranspiration, LULC, and precipitation. (3) The impact of LULC changes on water yield varied by scenario, with the scenarios ranked from most to least impactful as follows: UD, ND, CP, EP. This variation is mainly due to the different rates of evapotranspiration and infiltration associated with land cover. These insights are crucial for guiding policymakers in developing effective LULC strategies that promote ecological restoration and sustainable water management in the basin.

1. Introduction

Ecosystem services represent a broad spectrum of benefits that ecosystems provide [1,2], including provisioning services [3,4], regulating services [5], and cultural services [6]. As human society continues to develop, the demand for natural resources and ecosystem services has steadily increased [7]. However, due to destructive human activities, more than 60% of ecosystem services worldwide are facing degradation [8]. These destructive behaviors alter regional environments, reshape surface forms, and lead to various land use and land cover (LULC) patterns [9]. LULC change reflects the capacity of human activities to shape the Earth’s landscape and is a primary driver of changes in ecosystem services [7]. For instance, the development of agricultural land and landscapes enhances regulating and cultural services [5,6]; the reduction of forested areas leads to a decrease in carbon storage services [10]; and the increase in forest land results in a reduction in surface water yield and surface sediment production [2].

The study of hydrological ecosystem services is currently a prominent research topic. The 2019 report by the Intergovernmental Platform on Biodiversity and Ecosystem Services (IPBES) underscored the degradation of water resources as a significant threat to both biodiversity and ecosystem services [11,12]. The dynamics of water yield, which is influenced by an amalgam of natural and socio-economic factors [13,14], are fundamental to developing effective water resource protection strategies, instituting ecological compensation mechanisms, and promoting sustainable development [15,16]. LULC changes are influenced by a combination of natural and socio-economic factors and are key determinants of water yield services [3,17]. For example, Wamucii et al. studied the impact of LULC and climate change on water yield in East African forested water towers [18]; Zhou et al. investigated the hydrological responses to climate and LULC changes [17]; and Wang et al. examined hydrological ecosystem services under multiple scenario simulations in the Qilian Mountains [19]. Models for predicting long-term LULC changes include the cellular automata–Markov (CA-Markov) [20], the future land use simulation (FLUS) [21,22], and the conversion of land use and its effects at small regional extent (CLUE-S) [23,24]. The CA-Markov model does not account for social, economic, or political factors influencing LULC changes and ignores the historical state of LULC in making predictions [25]. The FLUS model builds on the traditional CA model by employing Artificial Neural Networks (ANNs) to fit multiple driving factors for various LULC types, thus predicting the spatial patterns of LULC suitability [26]. However, it struggles to identify the underlying forces driving LULC changes and to dynamically capture the evolution of multiple types of LULC patches [15]. To overcome these limitations in identifying the key driving factors behind LULC changes, the patch-generating land use simulation (PLUS) model was developed, integrating advanced simulation capabilities [27,28].

Hydrological modeling has become a critical tool in water resource management and assessing ecosystem service functions. Models like the soil and water assessment tool (SWAT) [29], BEPS-TerrainLab [30], and the integrated valuation of ecosystem services and trade-offs (InVEST) are widely used to simulate water yield distributions. The InVEST model, in particular, is favored for ecosystem assessments due to its simplicity, low data requirements, quick computation, and global applicability [31]. Redhead et al. validated the InVEST model on a national scale [32], while Yin et al. used it to estimate water yield in northern China [33]. Currently, numerous scholars utilize PLUS models combined with the InVEST model for research purposes [34,35]. However, research on the impact of regional LULC changes on water yield and their driving mechanisms remains incomplete.

The Ganjiang River, the largest river in Jiangxi and the region’s lifeblood, serves as a vital ecological barrier in the red soil hilly area of southern China. Positioned within the monsoon region, the Ganjiang River Basin (GRB) experiences significant climatic variability and is characterized by numerous mountains and hills, complicating water resource storage [36]. Climatic fluctuations and anthropogenic alterations to surface characteristics have led to variable water yields in the GRB, with decreases causing water shortages and sharp increases leading to flooding disasters [37]. These environmental challenges have escalated soil erosion, landslides, and other geohazards in the area. In response, the Ministry of Water Resources’ Water Allocation Plan for GRB, released in July 2022, mandates the rational allocation of water resources to sustain a healthy ecological system. A thorough understanding of regional water yield is essential for refined water resource management efforts. The GRB has undergone significant LULC changes, with varying impacts on water yield services [35]. In addition to natural factors, government land policies are key drivers of LULC changes. Estimating water yield under multiple future LULC scenarios influenced by policy has become a research hotspot. Currently, there is a lack of long-term series analysis of water yield drivers in the GRB [36], limiting support for policies related to watershed ecological protection and water resource management. The Geodetector can effectively explore the intrinsic factors driving geographical phenomena [38]. Currently, many scholars have used the InVEST, PLUS, and Geodetector models separately to explore water yield. However, the existing literature lacks in-depth discussions on coupling these three models to examine water yield in specific regions and explore its intrinsic driving factors. The approach of coupling these three models can also provide valuable guidance for related research. Therefore, this study integrates the InVEST, PLUS, and Geodetector models to simulate long-term water yield in the GRB, explore its driving factors, and examine changes under different future LULC scenarios, providing support for policymakers.

This study aims to achieve the following objectives: (1) identify the spatial and temporal distribution patterns of water yield in the GRB from 1990 to 2022; (2) determine the primary factors driving changes in water yield; and (3) predict future changes in water yield under various LULC scenarios. The findings of this study will provide strategic insights that enhance water resource protection and land management efforts in the GRB.

2. Materials and Methods

2.1. Study Area

The Ganjiang River, the principal river of the Poyang Lake system, originates from Wuyi Mountain and flows northward to Poyang Lake. Situated in southern Jiangxi Province, China, the GRB spans coordinates 113°6′–116°63′ E and 24°50′–28°70′ N, covering a main channel of 823 km in length [39] and a catchment area of 83,500 km², which accounts for 51% of the province’s territory (Figure 1). The GRB experiences a subtropical humid monsoon climate that is characterized by mild temperatures, with an average annual temperature of 18.9 °C, and substantial rainfall, averaging 1573 mm annually, primarily from March to June [40]. In terms of LULC, forests dominate, followed by cropland, while urban and impervious areas are gradually expanding. The basin’s topography mainly consists of undulating hills and mountains surrounding a central valley.

2.2. Data Sources and Processing

This study utilized a comprehensive suite of data, including meteorological, soil, remote-sensing, and socioeconomic datasets, as outlined in

Table 1. The dataset sources used in this study.

Type	Data Name	Spatial Resolution	Temporal Resolution	Source
Climate data	Precipitation (PRE) Potential	1 km	Yearly	National Earth System Science Data Center (https://www.geodata.cn, accessed on 19 December 2023)
	Evapotranspiration (PET)
	Temperature (TEM)
Soil data	Soil depth	100 m	–	https://doi.org/10.6084/m9.figshare.11358929, accessed on 19 December 2023 [44]
	Soil type	1 km	–	Data Center for Resources and Environmental Sciences of the Chinese Academy of Sciences (https://www.resdc.cn, accessed on 19 December 2023)
	AWC	250 m	–	ISRIC (https://www.isric.org, accessed on 19 December 2023)
Remote sensing data	LULC	30 m	Yearly	https://zenodo.org/record/8176941, accessed on 19 December 2023
	DEM	30 m	–	Geospatial Data Cloud (http://www.gscloud.cn, accessed on 19 December 2023)
	NDVI	1 km	Yearly	http://dx.doi.org/10.5067/MODIS/MOD13A3.006, accessed on 19 December 2023
Socioeco-nomic data	Population (POP)	1 km	5-year	Data Center for Resources and Environmental Sciences of the Chinese Academy of Sciences (https://www.resdc.cn, accessed on 19 December 2023)
	Gross domestic product (GDP)
	Road network	–	–	Open Street Map (https://www.openstreetmap.org, accessed on 19 December 2023)

. Annual potential evapotranspiration (PET) was aggregated from China’s monthly PET dataset [41,42]. Estimations of plant available water capacity (PAWC) were based on assessments of available soil water capacity (AWC). The LULC dataset [43] categorizes the landscape into the following seven distinct types: cropland, forest, shrub, grassland, water, barren, and impervious areas. Topographical slope measurements were extracted from the Digital Elevation Model (DEM). The Normalized Difference Vegetation Index (NDVI) was sourced from NASA Earthdata (https://www.earthdata.nasa.gov, accessed on 19 December 2023) utilizing the MOD13A3 monthly NDVI dataset at a 1 km resolution and applying a maximum value composite method to compile annual data. Road network data, crucial for spatial analysis in this study, were analyzed using ArcMap 10.8 to create buffer zones and calculate distances to primary, secondary, and tertiary roads.

2.3. Methodology

The framework of this study is delineated into three primary components (Figure 2). Initially, a simulation of water yield and its spatiotemporal dynamics within the GRB from 1990–2022 was conducted using the InVEST model. Subsequently, the Geodetector tool was utilized to ascertain the primary factors affecting water yield, drawing on both natural environmental and socioeconomic datasets. Lastly, the PLUS model was deployed to predict the 2030 LULC of GRB under four distinct policy scenarios, thereby projecting the subsequent characteristics of water yield as influenced by anticipated land management policies.

2.3.1. PLUS Model

The PLUS model, predicated on cellular automata (CA), is engineered to simulate dynamic shifts LULC and comprises two primary modules [27]. The initial module, the Land Expansion Analysis Strategy (LEAS), leverages bi-temporal LULC data to identify and sample the expansion segments of each LULC type. This module employs a random forest algorithm to train each LULC category and to formulate transformation rules for their respective expansion patterns. The second module, Cellular Automata based on Random Seeding (CARS), implements a threshold descent-based, multi-type stochastic patch seeding mechanism to emulate the patch evolution of multiple LULC types under developmental probability constraints. Additionally, the PLUS model integrates a Markov Chain module to project future LULC demands and simulate scenarios based on varied settings.

In this study, aligned with the natural environment, socioeconomic development of the research area, and data availability, 13 factors driving LULC change were selected, encompassing 5 natural elements (TMP, PRE, soil type, DEM, and slope), 2 socioeconomic factors (GDP and population), and 6 accessibility factors (distance to primary, secondary, and tertiary roads; railways; motorways; and open water). To validate the PLUS model, LULC for 2020 was simulated using data from 2000 and 2010, yielding a Kappa coefficient of 0.82 and an overall accuracy of 0.91, thus demonstrating the model’s effectiveness in predicting future LULC patterns. Using 2010 and 2020 LULC data, this study simulated 2030 LULC under diverse scenarios through the CARS module. These scenarios conform to the Land Spatial Planning of Jiangxi Province (2021–2035) and include natural development (ND), cropland protection (CP), ecological protection (EP), and urban development (UD) scenarios. To ensure the simulations of different scenarios are more realistic, a LULC transition cost matrix for the four scenarios was set up. This matrix is based on the actual conditions of the study area (

Table 2. LULC cost matrices under four scenarios.

	ND							CP							EP							UD
	a	b	c	d	e	f	g	a	b	c	d	e	f	g	a	b	c	d	e	f	g	a	b	c	d	e	f	g
a	1	1	1	1	1	1	1	1	0	0	0	0	0	0	1	1	1	1	1	1	1	1	0	0	0	1	0	1
b	1	1	1	1	1	1	1	1	1	1	1	1	1	1	0	1	0	0	0	0	0	1	1	1	0	1	1	1
c	1	1	1	1	1	1	1	1	1	1	1	1	1	1	0	1	1	0	0	0	0	1	1	1	1	1	1	1
d	1	1	1	1	1	1	1	1	1	1	1	1	1	1	0	1	1	1	1	0	0	1	1	1	1	1	1	1
e	1	1	1	1	1	1	1	1	0	1	1	1	0	0	0	1	1	1	1	0	0	1	0	1	1	1	1	1
f	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1
g	0	0	0	0	0	0	1	0	0	0	0	0	0	1	0	0	0	0	0	0	1	0	0	0	0	0	0	1

Notes: The annotations “a”, “b”, “c”, “d”, “e”, “f” and “g” refer to the LULC types of cropland, forest, shrub, grassland, water, barren, and impervious, respectively. Within the LULC transition matrix, a value of 0 signifies that conversion between specified LULC types is restricted, whereas a value of 1 indicates that conversion is allowed. This matrix serves as a key tool for modeling potential changes in land use based on predefined and permitted transitions.

). The neighborhood weighting parameter represents the expansion intensity of each land type and ranges from 0 to 1 [45]. The total area change of each land type better reflects the expansion intensity. The calculation formula is referenced from Wang et al. (

Table 3. Neighborhood weighting parameters for each LULC in the GRB from 2010 to 2020.

	Cropland	Forest	Shrub	Grassland	Water	Barren	Impervious
Neighborhood weighting	1	0	0.57	0.56	0.57	0.57	0.73

) [46]. The ND scenario continues the LULC expansion rates from 2000 to 2010 without modifications. The CP scenario aims to curtail cropland reduction, reducing the conversion probability to impervious surfaces by 60%. The EP scenario is designed to enhance the protection of ecological land. So that the probability of converting the three LULC types—forest, grassland, and shrub—to impervious surfaces will be reduced by 50%, the probability of converting cropland to impervious surfaces will be reduced by 30%, and the probability of converting cropland and grassland to forest will be increased by 30%. Conversely, the UD scenario, aligned with the province’s urbanization strategy, intensifies urban land expansion by elevating the conversion probability of natural and agricultural lands to urban areas (increases by 50%), while decreasing the reversal from impervious to other LULC types except for cropland (decreases by 20%).

2.3.2. InVEST Water Yield Model

The InVEST water yield model employs the Budyko curve [47] and the water balance principle to estimate the annual water yield for each pixel, as calculated by:

$$\begin{array}{c}WY\left(x\right)=\left(1-{\displaystyle \frac{AET\left(x\right)}{PRE\left(x\right)}}\right)·PRE\left(x\right)\end{array}$$

(1)

where WY(x) is the annual water yield for pixel x; AET(x) is the annual actual evapotranspiration (AET); and PRE(x) is the annual precipitation. For vegetated LULC, it is calculated based on Fu et al. [48] and Zhang et al. [49],

$${\displaystyle \frac{AET\left(x\right)}{PRE\left(x\right)}}=1+{\displaystyle \frac{PET\left(x\right)}{PRE\left(x\right)}}-{\left[1+{\left({\displaystyle \frac{PET\left(x\right)}{PRE\left(x\right)}}\right)}^{\omega }\right]}^{{\displaystyle \frac{1}{\omega }}}$$

(2)

where PET(x) represents the potential evapotranspiration, and ω(x) is an empirical, non-physical parameter that characterizes natural climate and soil properties. ω(x) is calculated using [50]

$$\omega \left(x\right)=Z{\displaystyle \frac{AWC\left(x\right)}{PRE\left(x\right)}}+1.25$$

(3)

where AWC(x) represents the available soil water capacity, which is determined by the soil’s capacity to retain and release water, calculated from the plant-available water content. The Z parameter is a seasonal factor reflecting local PRE patterns and other hydrogeological characteristics, and it is positively correlated with the annual number of rainfall events. For other LULC types, AET is directly calculated from the reference evapotranspiration.

The biophysical table for the InVEST water yield model utilizes a ‘LULC_veg’ to categorize LULC types. In this coding, a value of 1 is assigned to all vegetation types except wetlands, and a value of 0 is given to non-vegetative types. The ‘root depth’ parameter, which indicates the depth where 95% of a vegetation type’s root biomass is found, is actively used only for vegetation types. The vegetation types are based on the values of the same types in similar regions [36,51]. For non-vegetation types, it is set to −1, indicating that the parameter does not apply. The ‘Kc’ value for each LULC type represents the plant evapotranspiration coefficient, established according to the Food and Agriculture Organization (FAO) reference values for vegetation types (http://www.fao.org/3/X0490E/x0490e0b.htm, accessed on 19 December 2023) and adjusted for non-vegetation types following the InVEST user guide. The configuration of the biophysical table is based on directives from the InVEST user guide and corroborated by studies from neighboring regions [51], as detailed in

Table 4. Biophysical parameters for the InVEST model.

Lucode	LULC	LULC_Veg	Root_Depth (mm)	Kc
1	Cropland	1	2000	0.9
2	Forest	1	3500	1
3	Shrub	1	3000	0.3
4	Grassland	1	2400	0.65
5	Water	0	−1	1
6	Barren	0	−1	0.5
7	Impervious	0	−1	0.1

.

The Z parameter, an empirical constant used to describe the hydrogeological characteristics and seasonal distribution of PRE in the study area, ranges from 1 to 30. Various methods are used to determine the Z parameter. This study adjusted the Z value based on the total water resources of GRB from 2010 to 2022, as reported in the Water Resources Bulletin of Jiangxi Province. By modifying the Z value to two decimal places, we selected the Z value corresponding to the annual water yield that matches the actual total water resources. The average of all matching Z values was calculated to be 1.573, consistent with the findings of Wang et al. in GRB [36].

2.3.3. Geodetector Model

The Geodetector model is designed to explore the spatial heterogeneity of geographic phenomena and their underlying drivers [38]. It consists of four sub-models: the factor detector, interaction detector, risk detector, and ecological detector. In this study, we specifically utilized the factor detector and the interaction detector.

The factor detector is used to analyze the spatial variability of geographic phenomena, employing the q-statistic to measure the degree to which a particular factor influences this variability. The q-statistic ranges from 0 to 1, where a higher value indicates that factor X has a stronger explanatory power over the dependent variable Y, and a lower value suggests a weaker influence. The formula for the q-statistic is structured as follows:

$$\begin{array}{c}q=1-{\displaystyle \frac{\sum _{h=1}^L{N}_h{\sigma }_h^2}{N{\sigma }^2}}=1-{\displaystyle \frac{SSW}{SST}}\end{array}$$

(4)

$$SSW=\sum _{h=1}^L{N}_n{\sigma }_h^2$$

(5)

$$SST=N{\sigma }_h^2$$

(6)

where h = 1, …, L is the stratification; $N_h$ and N are the number of cells in stratum h and the whole region, respectively; ${\sigma }_h^2$ and ${\sigma }^2$ are the variance of the dependent variable Y in stratum h and the whole region, respectively; SSW stands for the sum of the variances within the stratum, and SST stands for the total variance in the whole region.

Interaction detectors are employed to assess the synergies or antagonisms between two distinct factors, examining whether their combined influence on the dependent variable Y is greater or lesser than the sum of their individual effects. This tool also ascertains whether each factor independently contributes to the explanation of Y.

In this study, a comprehensive analysis of 12 factors spanning natural environmental and socioeconomic dimensions was conducted to identify the drivers influencing water yield changes, utilizing the Geodetector model. These factors included PET, LULC, PRE, DEM, slope, aspect, TEM, soil type, AET, NDVI, GDP, and POP. Due to data limitations, the analysis of GDP and POP was restricted to data points from the years 2000, 2005, 2010, 2015, and 2019.

3. Results

3.1. LULC Changes and Their Driving Mechanisms from 1990 to 2022

Figure 3 illustrates the dynamics of LULC changes within the GRB over a span of 33 years. The primary LULC categories identified are forest, cropland, and impervious surfaces, with forests occupying the most extensive area. In 2022, the LULC classes were hierarchically arranged from the most to the least expansive as follows: forest, cropland, impervious surfaces, water bodies, grassland, shrubland, and barren areas, as detailed in

Table 5. LULC area from 1990 to 2022 (km²).

	Cropland	Forest	Shrub	Grassland	Water	Barren	Impervious
1990	22,089.94	56,109.34	14.91	421.60	1012.58	16.78	1262.26
2000	21,157.02	56,828.96	21.11	131.47	1147.64	14.60	1626.62
2010	20,689.63	56,719.86	20.62	109.33	1191.32	9.20	2187.45
2022	22,956.09	53,796.18	12.15	26.25	1140.07	1.73	3012.06
1990–2022	866.15	−2313.16	−2.76	−395.34	127.49	−15.04	1749.80

. The period from 1990 to 2022 witnessed substantial LULC transitions, notably between forested areas and croplands: a conversion of 2497.42 km² of cropland was reforested, whereas 4732.01 km² of forested land was transformed into cropland, exhibiting conversion rates of 11.30% and 8.43%, respectively (

Table 6. Percentage of LULC Type Transitions from 1990 to 2022 (%).

	Cropland	Forest	Shrub	Grassland	Water	Barren	Impervious
Cropland	81.26	11.30	0.00	0.02	1.05	0.00	6.36
Forest	8.43	90.94	0.01	0.02	0.02	0.00	0.57
Shrub	2.97	62.56	25.87	8.29	0.01	0.02	0.28
Grassland	29.10	54.38	0.14	2.05	6.43	0.09	7.80
Water	13.22	1.77	0.00	0.12	80.42	0.02	4.45
Barren	9.10	0.30	0.00	3.99	62.05	2.10	22.46
Impervious	1.09	0.04	0.00	0.01	3.63	0.00	95.23

). Grasslands predominantly shifted towards cropland and forested areas, contributing to areas of 122.65 km² and 229.19 km², respectively. Urban expansion was remarkably evident, with impervious surfaces significantly expanding through the transformation of both cropland and forested regions, amounting to 1405.46 km² and 322.71 km², respectively. The majority of barren lands underwent hydrological conversion to water bodies, with a notable proportion also transitioning into impervious and cropland areas.

From 1990 to 2022, the forested regions within the GRB exhibited the most profound changes, with an aggregate reduction of 2313.16 km². Initially, there was a marginal increment in forest cover from 1990 to 2010, subsequently overshadowed by a pronounced decrease from 2010 to 2022, which amounted to 2923.68 km². The area classified as impervious surfaces displayed a significant uptrend, expanding from 1262.26 km² in 1990 to 3012.06 km² in 2022, a cumulative increase of 1749.80 km² (Table 5). Cropland depicted an overall growth of 866.15 km², despite experiencing a gradual decline from 1990 to 2010, followed by a robust escalation from 2010 to 2022. Aquatic areas also increased, rising from 1012.58 km² in 1990 to 1140.07 km² in 2022, an augmentation of 127.49 km². In stark contrast, grassland areas precipitously diminished to a mere 26.25 km² in 2022, a decrement of 395.34 km². Changes in shrub and barren landscapes remained minimal. Collectively, the LULC dynamics within the GRB have witnessed extensive changes over the past 33 years, with the interval from 2010 to 2022 characterized by considerable shifts, predominantly the conversion of forested areas to impervious and agricultural lands.

The influence of various driving factors on LULC changes was quantitatively analyzed using the PLUS model (Figure 4). Excluding the impervious category, the DEM emerged as the predominant factor influencing six different land types: cropland, forest, shrub, grassland, water, and barren. The respective contributions of DEM to these categories were 13.71%, 14.82%, 31.86%, 30.10%, 26.37%, and 26.27%. Beyond DEM, cropland is also significantly affected by slope, contributing 10.10%, and GDP, which accounts for 9.62% of its variability. Additionally, TEM and PRE have substantial impacts on the dynamics of forests and shrublands. Slope also plays a critical role in shaping the distribution of water bodies and grasslands, contributing 19.74% and 16.34%, respectively. The distance to tertiary roads stands out as the principal driver for the expansion of impervious surfaces, accounting for 34.75% of the changes observed, with road construction often spearheading urban expansion.

3.2. Spatiotemporal Variation of Water Yield

Utilizing the water yield module of the InVEST model, the annual water yield within the GRB from 1990 to 2022 was calculated, as illustrated in Figure 5a, revealing significant fluctuations ranging from 3.52 × 10¹⁰ to 10.02 × 10¹⁰ m³. The highest and lowest water yields were recorded in 2011 and 2012, respectively. Specifically, the water yields for the years 1990, 2000, 2010, and 2020 were 6.94 × 10¹⁰, 6.93 × 10¹⁰, 9.45 × 10¹⁰, and 6.64 × 10¹⁰ m³, respectively. The temporal changes in water yield can be divided into three distinct phases, collectively forming an M-shaped pattern: an initial gradual increase from 6.94 × 10¹⁰ to 9.41 × 10¹⁰ m³ from 1990 to 2002, a period of marked variability from 2003 to 2016, and a phase of minor changes from 2017 to 2022, with water yields stabilizing around 6.50 × 10¹⁰ m³.

The spatial distribution of water yield within the GRB displays significant variability in the multi-year average, characterized by higher yields in the northern and eastern sectors and lower yields in the southern and western areas (Figure 5b). In the southern part of the GRB, average water yields remain consistently low, typically ranging between 300 and 800 mm. Conversely, the northeastern part of the basin is distinguished by high water yield zones, with sporadic high-yield areas surrounding water bodies and built-up regions across the basin.

Temporal changes in water yield are notable. Between 1990 and 2000, variations in water yield across the basin were minimal; most central regions experienced slight increases, from 0 to 300 mm, while other areas saw decreases ranging from −300 to 0 mm (Figure 6a). From 2000 to 2010, there was a general increase in water yield, particularly from southwest to northeast, with notable increments of approximately 300 to 600 mm recorded in the northeastern regions (Figure 6b). However, from 2010 to 2022, except for a few minor areas in the southwest, the overall water yield declined, with reductions intensifying from southwest to northeast, culminating in a maximum decrease of up to 1783 mm (Figure 6c). Overall, the spatial analysis of water yield in the basin indicates a declining trend (Figure 6d).

3.3. Driving Factors of Water Yield

The Geodetector model was utilized to elucidate the factors influencing water yield across five distinct temporal snapshots (2000, 2005, 2010, 2015, and 2019), as depicted in Figure 7. All analyzed factors demonstrated a p-value of 0, signifying their high statistical significance. AET and LULC displayed pronounced spatial variability in their explanatory capacity concerning water yield, with their q-statistics oscillating between 0.51–0.71 and 0.42–0.61, respectively. PRE emerged as another notable factor, with its q-statistic varying from 0.23 to 0.62. Apart from the DEM, which exhibited elevated q-statistics in 2010 and 2015, the other nine factors generally registered q-statistics ranging from 0 to 0.2. Except for the year 2010, AET and LULC consistently demonstrated the highest q-statistics. The above analysis identifies AET and LULC as the most important factors impacting water yield within the GRB.

Water yield in the GRB is also significantly influenced by the complex interactions among various factors, which collectively exert a more substantial impact on water yield than any individual factor. This synergy between factors enhances dual-factor and nonlinear effects on water yield. In 2019, the interactions involving PRE with other factors achieved a q-statistic exceeding 0.6, in contrast to the q-statistic of 0.31 observed for single-factor effects (Figure 8). Under the continuous development and influence of natural and socioeconomic factors, the spatial distribution of water yield has become increasingly complex. Notably, the interaction between AET and PRE was especially pronounced, consistently registering a q-statistic above 0.98 over a five-year period. Similarly, the interaction between LULC and PRE, with a q-statistic exceeding 0.95, also plays a critical role in modulating water yield within the GRB. This indicates that over the past 20 years, changes in the water yield of the GRB have been primarily driven by AET, PRE, and LULC. Interactions among other factors are weaker compared to AET, PRE, and LULC, with the q-statistic mostly below 0.3 across five years, suggesting that the effects of these factors on the spatial distribution of water yield are not significant. In 2010, the interactions between PRE and other factors were more significant than in the other four years, which is likely related to that year’s anomalous PRE and consistent with the results from the single-factor interaction analysis.

3.4. LULC Prediction and Effects of Its Future Scenarios on Water Yield

According to projections from the PLUS model, four LULC scenarios for 2030 have been delineated: ND, UD, EP, and CP (Figure 9). Within the ND scenario, cropland and impervious surfaces are anticipated to experience the most pronounced augmentations relative to 2020 metrics. Cropland is projected to undergo an expansion of 1647.22 km², representing a 7.24% increment, while impervious zones are expected to increase by 630.50 km², a substantial 21.39% growth (

Table 7. Area of LULC under different development scenarios in 2030 and rate of change in 2020–2030.

Type	Year	Development Scenario	Cropland	Forest	Shrub	Grassland	Water	Barren	Impervious
Area(km2)	2020	-	22,745.86	53,995.90	14.50	32.68	1206.68	1.98	2947.10
	2030	CP	24,789.98	51,579.21	33.23	75.16	1239.55	0.89	3226.68
		ND	24,393.08	51,553.19	59.83	122.53	1237.58	0.88	3577.60
		UD	24,258.65	51,519.56	54.32	48.59	1211.17	0.88	3851.54
		EP	22,067.88	54,149.93	11.18	28.33	1236.60	0.88	3449.90
Rate of Change (%)	2030	CP	2.19	0.12	−38.82	54.69	2.34	1.13	−16.22
		ND	7.24	−4.52	312.71	274.93	2.56	−55.44	21.39
		UD	6.65	−4.59	274.68	48.67	0.37	−55.49	30.69
		EP	−9.53	5.04	−81.32	−76.88	−0.08	0.00	−3.57

). Additionally, there are forecasted expansions in the areas designated as shrub, grassland, and water, with increments of 45.33 km², 89.85 km², and 30.90 km², respectively. Conversely, forest coverage is projected to witness a considerable contraction, diminishing by 2442.71 km² or 4.52% by the year 2030.

The UD scenario illustrates a notable expansion in impervious surfaces, increasing by 904.44 km² or 30.69% relative to 2020 levels. While there are also expansions in cropland, shrub, and grassland, these are less significant compared to the ND scenario (Table 7). In this scenario, forest areas exhibit the most substantial reduction, declining by 4.59%, primarily due to conversions to impervious and cropland areas. In the EP scenario, forest coverage significantly increases by 5.04% from 2020, amounting to an expansion of 154.03 km². In contrast, other LULC types witnessed reductions, with cropland and impervious areas notably restricted, declining by 9.53% and 3.57%, respectively. The CP scenario demonstrates the most drastic reduction in impervious surfaces among all scenarios, decreasing by 16.22% to 350.92 km², reflecting effective CP policies aimed at controlling urban expansion. Concurrent increases in cropland, forest, and grassland by 2.19%, 0.12%, and 54.69%, respectively, indicate targeted policy impacts on these land uses. Minimal changes in other LULC types emphasize the focused nature of these interventions.

Using the InVEST model, the water yield results for four different land use scenarios in 2030 were generated (Figure 10). The PRE and AET data were based on the average values from 1990 to 2022. All other parameters remained constant, with only the LULC data being altered (Figure 11). This indicates that the UD scenario results in the highest water yield, measuring 6.87 × 10¹⁰ m³. In contrast, the EP scenario shows the lowest annual water yield at 6.80 × 10¹⁰ m³. The ND and CP scenarios produce water yields of 6.85 × 10¹⁰ m³ and 6.83 × 10¹⁰ m³, respectively. Notably, the UD scenario, characterized by increased urbanization and the expansion of impervious surfaces, contributes to the enhancement of water yield.

4. Discussion

4.1. Driving Mechanism of Water Yield

The spatial and temporal differentiation characteristics of water yield in the GRB are intricately linked to climate change dynamics. Temporally, annual fluctuations in water yield are well-aligned with variations in PRE, as depicted in Figure 12a, but show minimal correlation with changes in AET, as seen in Figure 12b. Consequently, interannual variations in water yield are predominantly driven by PRE. Spatially, water yield variations are influenced by a confluence of climatic factors. In the northeast region of the basin, where PRE levels are high and AET levels are low, water yields are correspondingly high. Conversely, in the southern region, where PRE is low and AET is high, water yields are notably low (Figure 13). These observations are consistent with the findings of Chen et al. [52], highlighting the climatic impact on water yield. Therefore, in the future, if the PRE decreases continuously and the AET increases in the GRB, with other factors unchanged, the water yield will also decrease.

The spatial differentiation in water yield results from the synergistic effects of multiple factors. Climatic factors and LULC are the principal contributors to spatial heterogeneity in water yield. According to Dai and Wang [53], in regions with minimal topographical variation, such as plains, tablelands, and hills, AET, PRE, and LULC emerge as the most significant factors, with the interaction between PRE and LULC being particularly influential, corroborating the results of this study. Changes in LULC types are closely related to human activities, indicating that the spatial heterogeneity of water yield is influenced by both climate and human activities. Therefore, when adjusting LULC management strategies in the GRB, it is crucial to focus on the impact of LULC changes on water yield services. Moreover, the interactions between different factors have a more significant impact on water yield than individual factors alone, highlighting the need to consider the synergistic effects of multiple factors in maintaining and optimizing water yield functions in the GRB in the future.

Significantly, the influential factors of water yield in 2010 diverge from those observed in other years such as 2000, 2005, 2015, and 2019. In 2010, PRE was identified as having the most robust effects on the spatial heterogeneity of water yield. This year recorded an average PRE of 1913 mm across the study area, which is considerably higher than the average of approximately 1600 mm noted in other years, as illustrated in Figure 12. Furthermore, the spatial pattern of PRE in 2010 was distinctly different from other years, with higher PRE in the northeast and lower amounts in the southwest. This variation in PRE distribution was closely related to the observed spatial distribution of water yield, highlighting the critical influence of annual climatic variations on hydrological outcomes.

Despite variations in LULC from 2000 to 2019, the overarching spatial pattern of LULC has demonstrated considerable consistency. Lang et al. [54] and Dou et al. [55] suggested that climatic changes exert a more profound impact on water yield than LULC. In this study, we found the criticality of the distinctive PRE patterns in 2010 to be an important element in the observed variances for that year. The unique meteorological conditions in 2010, particularly the heightened PRE levels, are identified as pivotal factors causing the spatial heterogeneity in water yield, underscoring the predominance of climatic fluctuations over land cover transformations in influencing hydrological dynamics. Therefore, water resource management departments aiming to achieve a balance in regional water yield services cannot rely solely on natural rainfall. They must also implement the timely adjustment of reservoir storage and release plans. Additionally, rational management of water resources is essential [56]. For water resource users, the rational use of water is crucial. In agricultural irrigation, users should manage water scientifically based on crop water requirements and soil moisture conditions.

4.2. Water Yield under Different LULC Scenarios

LULC can indirectly modulate water yield by influencing AET and infiltration rates [51]. Over the period from 1990 to 2022, urbanization in the GRB accelerated, propelling Ganzhou into a more developed industrial hub in Jiangxi Province, buoyed by its ecological assets and geographic advantages [57]. This urban expansion was concurrently linked to an increase in water yield [35].

This study pinpointed AET as a pivotal factor affecting water yield. In areas with dense forest cover, elevated rates of AET and infiltration contribute to reduced water yields, assuming other factors remain constant [58,59]. Under the UD scenario, the extensive transformation of forest into grassland, shrub, and impervious surfaces—each characterized by lower evapotranspiration capacities—resulted in an increase in water yield [60]. Therefore, in future urban planning and socio-economic activities, it is essential to optimize land use structure. Additionally, the rational development and utilization of land are crucial. These measures aim to achieve harmonious coexistence between humans and the environment [31].

It is crucial to recognize that while LULC significantly influences the spatial distribution of water yield and is the primary driver of its spatial variability within the study area, its impact on the overall annual water yield is comparatively modest. The variances in total annual water yield across the four land use scenarios were minimal (Figure 11), which is consistent with studies conducted in the Dongjiang River Basin [61] and the Qilian Mountain Basin [62].

4.3. Limitations and Prospects

Utilizing the InVEST and Geodetector models, this study investigates the spatiotemporal variation and factors driving water yield in the GRB from 1990 to 2022. Additionally, it integrates the PLUS model to forecast changes in water yield under four scenarios in 2030. The findings of this study possess a certain level of accuracy and scientific validity, yet some issues remain.

All models have inherent limitations, as noted in the InVEST user manual and related studies [63]. The InVEST model employs a limited set of parameters and simplified equations to simulate water yield. While this approach facilitates convenient result acquisition, it overlooks the interactions between surface water and groundwater in its calculations. Moreover, the model does not account for the influence of human activities [64], leading to potential discrepancies between simulated and actual results. The PLUS model’s multi-scenario design also exhibits certain flaws. To predict future changes in LULC types under various land planning policies and to distinctly delineate the effects of different scenarios, this study adopts an aggressive assumption strategy. For instance, in the ecological protection scenario, the conversion of forests, grasslands, and shrubs to non-natural LULC types is prohibited. However, actual LULC changes are complex, and the predictive results of this study can only provide directional guidance. This study focuses solely on changes in water yield under LULC change scenarios, neglecting the influence of other factors. Future work should address the coupled impact of various factors on water yield, which warrants further investigation.

5. Conclusions

This study utilized the InVEST model to simulate water yield in the GRB in Jiangxi Province from 1990 to 2022, analyzing the spatiotemporal dynamics of water yield over this period. Additionally, the Geodetector method was employed to assess important factors influencing water yield, while the PLUS model was applied to project future LULC for 2030 under four different scenarios, thereby facilitating the clarification of potential changes in water yield under these conditions.

The results reveal that between 1990 and 2022, there was a significant reduction in forest area by 2313.16 km², whereas impervious surfaces and cropland increased by 1749.80 km² and 866.15 km², respectively. From 2010 to 2020, the topography (as revealed by the DEM) emerged as the dominant factor influencing all six LULC types: cropland, forest, shrub, grassland, water, and barren areas.

Over the 33-year period, the water yield in the GRB initially increased, followed by a decline, stabilizing at approximately 6.50 × 10¹⁰ m³ from 2017 to 2022, predominantly influenced by PRE. Spatially, water yield typically exhibited higher values in the northeast and lower values in the southwest, with a scattered distribution in the central built-up areas. The basin, predominantly comprising plains and hilly areas with low topographic relief and some mountainous regions, showed that AET, LULC, and PRE were the primary factors affecting water yield.

Under the scenarios of ND, UD, EP, and CP, the UD scenario projected the highest water yield for 2030, estimated at 6.87 × 10¹⁰ m³. In contrast, the EP scenario resulted in the lowest water yield, which may be attributed to differing AET and infiltration capacities associated with various LULC types. These insights could provide a theoretical basis for regional planning and decision-making regarding water resource management.