Impact of Land use dynamics on the water yields in the Gorgan river basin

Abstract

This research investigates the future dynamics of water yield services in the Gorgan River Basin in the North of Iran by analyzing land cover changes from 1990 to 2020, using Landsat images and predicting up to 2040 with the Land Change Modeler and InVEST model under three scenarios: continuation, conservation, and mitigation. The results indicate significant shifts in agricultural land impacted water yields, which fluctuated from 324.7 million cubic meters (MCM) in 1990 to 279.7 MCM in 2010, before rising to 320.1 MCM by 2020. The study uniquely assesses the effects of land use changes on water yields, projecting a 13.6 % increase in water yield by 2040 under the continuation scenario, a 3.9 % increase under conservation, and a 1.6 % decrease under mitigation, which limits changes on steep slopes to prevent soil erosion and floods. This underscores the interplay between land use, vegetation cover, and water yield, emphasizing strategic land management for water resource preservation and effective watershed management in the GRB.

Graphical abstract

Highlights

• •Land use changes modeled for 3 future scenarios.
• •Water yield impact evaluated under future land use scenarios.
• •Watershed runoff impact analyzed by means of InVEST model.
• •Mitigation scenario limits the soil loss and flood hazard.

1. Introduction

Land use and land cover characteristics have been shown to significantly influence land surface conditions [1], which influences hydrological processes, notably groundwater infiltration and surface runoff [2]. In addition, a close association has been found between the hydrological response of watershed areas and non-uniform alterations in land use or changes in vegetation cover [3]. Notably, land use transformation is a widespread and rapid process, often detrimentally affecting essential natural resources, such as soil and water resources [4]. For example, Acharya et al. [5] demonstrated that rapid urbanization in the Kathmandu Valley Watershed significantly impacts water dynamics, notably increasing surface runoff and reducing groundwater flow, necessitating immediate and effective water management interventions. The Gorgan River Basin (GRB) is also experiencing extensive land cover changes, which necessitate an evaluation of these alterations and their potential implications for water resource development and management [6].

The examination of the consequences of alterations in land use on surface runoff can be facilitated using ecosystem service models. Ecosystem services encompass the advantages individuals gain from ecosystems and are categorized into four distinct groups: provisioning, regulating, supporting, and cultural services. Their purpose is to establish a tangible connection with human well-being. Regulating services constitute the benefits stemming from the natural regulatory processes of the ecosystem. Notably, water yield falls under this category [7]. Ecosystem services linked to the safeguarding of water resources, inclusive of those delivered by the self-regulating functions of forest ecosystems are regarded as among the most valuable ecosystem functions. The value associated with these self-regulating functions falls within the realm of indirect non-use values, a category that regrettably lacks a genuine market for estimating these values. As a result, the significance of these functions often goes unnoticed until they are compromised. The harsh reality is that failing to acknowledge these services as complimentary offerings in economic equations may lead to escalating the degradation of these vital functions. However, the role played by forests in preserving water resources is of paramount importance, especially in nations confronting constraints on water resources [8].

In light of the growing recognition of the significance of ecosystem services, there is a growing demand for software and models that can offer decision-makers insights into the provision of ecosystem services and assess the repercussions of land management practices on these services.

Furthermore, a multitude of studies has delved into the consequences of alterations in LULC on the hydrological attributes of rivers, highlighting their considerable impact on modifications in runoff and river flow. Guo et al. [9] conducted a comprehensive examination of the impact of both seasonal and annual weather fluctuations, as well as changes in land cover, on runoff within a lake basin in China, employing a modeling approach. Their findings underscored the substantial influence of land cover changes on seasonal streamflow and the modification of the annual hydrograph.

Bi et al. [10] analyzed the interaction between precipitation and land use over 50 years in a Chinese watershed. Their research identified a significant increase in land use changes from 1954 to 2008, alongside a notable reduction in runoff by up to 49.63 %, largely due to decreased forest cover. Similarly, Nasta et al. [11] undertook an investigation to assess the enduring repercussions of land use alterations on the hydrological performance of an ecosystem situated in a Mediterranean region of southern Italy. Their research affirmed the pivotal role of forest cover in diminishing runoff and concurrently enhancing evapotranspiration. Chiang et al. [12] explored the consequences of alterations in land surface on ecohydrological processes within the Chenyulan watershed in Taiwan. In a parallel study during the same year, Samie et al. [13] evaluated the impact of land use changes on runoff, utilizing classified satellite data. Their primary objective was to ascertain the influence of land use changes on runoff within the Chenar Rahdar watershed.

The forest ecosystems within the GRB hold a pivotal position in the regulation of surface runoff, mitigating flood occurrences, safeguarding surface soil integrity, and combatting soil erosion. Additionally, these ecosystems contribute to temperature moderation and the reduction of greenhouse gas emissions. In this research, we focus on the transformations in land cover within the GRB and their influence on the control of surface runoff, specifically in the context of water yield as an ecosystem service. Unfortunately, there is not any preventive regulations or legal provisions to prohibit deforestation. Accordingly, the objective of this study is to investigate spatial changes in land use and land cover (LULC) in the Gorgan River Basin under different management scenarios using the Land Change Modeler (LCM) tool and analyze the relationship between land cover dynamics and vegetation changes on water yield services in the study area. The first scenario, labeled as the "continuity scenario," posits that land use changes will persist, following patterns like previous years, without any imposed constraints. The second scenario, known as the "agriculture-conservation scenario," is designed to halt the conversion of agricultural lands into residential zones, specifically preventing villa construction. Finally, the third scenario, termed the "limited change scenario," permits alterations in land use within the region but enforces restrictions on areas characterized by slopes exceeding 15 % and vegetation cover density surpassing 50 %.

2. Materials and methods

2.1. Study area

The research area for this study encompasses the GRB, which extends from the Gorgan River to the entrance of the Voshmgir Dam, covering a total area of 7548 square kilometers. Situated in the northern part of Iran, its geographical coordinates span from approximately 54° 42′ to 56° 28′ latitude and 36° 43′ to 37° 49′ longitude. Within this region, the highest point within the watershed can be found in the Khosh-Yeylagh area in the southwest, boasting an elevation of 2898 m, while the lowest point is located at the Voshmgir Dam, situated at an elevation of merely 10 m above sea level. The GRB, on average, maintains an elevation of 890 m and features an average slope of 18 % [14]. The population in the Gorganrud River Basin increased from 1,054,518 in 2011 to 1,091,231 by 2016, with a 3.5 % growth rate over five years. It is projected to reach around 1,400,000 by 2040, assuming a national policy-driven increase in the growth rate to about 6 % every five years [15].

The Gorgan River, recognized as the longest river within the watershed, originates from mountains reaching a peak of 2297 m and stretches across a length of 333 km. The annual precipitation within the watershed exhibits a range of 195–946 mm. The lowest average annual temperature registers at around 11 °C, while the highest average annual temperature is recorded at the evaporation station in proximity to the Gorgan Dam, approximately at 18.1 °C. Fig. 1 provides a visual representation of the GRB's location within the country.

Fig. 1.

Location of the Gorgan River Basin in northern Iran.

2.2. Methodology

2.2.1. Overview of the research approach

This paper is structured into four main sections. Firstly, it entails the analysis of land cover changes within the GRB by employing satellite imagery spanning the years 1990–2020. Subsequently, the second section involves the modeling of these land cover changes through the utilization of a Multi-Layer Perceptron (MLP) in three distinctive scenarios: continuity, conservation, and development limitation. Finally, the last section is dedicated to the evaluation of the role played by land cover alterations in the regulation of surface water yield, an integral ecosystem service. The structural layout of the research phases is visually presented in Fig. 2.

Fig. 2.

Main processes of the study Diagram.

2.2.2. Data Collection and preparation

In order to generate the Land Cover/Land Use (LC/LU) map of the GRB, Landsat Thematic Mapper (TM) and Operational Land Imager (OLI) satellite images for the years 1990, 2000, 2010, and 2020 were acquired from the United States Geological Survey (USGS) website (Table 1). These image layers underwent preprocessing using IDRISI Terrset version 18.11 and ArcGIS 10.3 software tools. Subsequently, all layers were converted into raster format, featuring spatial coordinates based on WGS 1984 UTM-Zone-40N projection, and a pixel size of 30 m, corresponding to the geographical boundaries of the GRB. These processed layers were prepared for utilization in the modeling and analytical phases that followed.

Table 1.

Landsat Images specifications.

Year	Path	Row	Date Acquired	Sensor ID
1990	162	34	5/30/1990	TM
2000	162	34	October 6, 2000	TM
2010	162	34	June 6, 2010	TM
2020	162	34	6/17/2020	OLI_TIRS

To enhance the identification of features within the TM and OLI satellite images, field surveys were conducted, employing true and false-color combinations. Training samples, were meticulously collected. Random sampling was employed to evaluate the accuracy of the classification maps generated from satellite imagery. On-site sample locations were determined using a high-precision Global Positioning System (GPS) device, which allowed for the precise categorization of land cover types at each location (Fig. 3). This process identified five primary classes, comprising forest, agriculture, pasture, residential areas, and water bodies. For satellite image classification, an object-based classification technique, utilizing the nearest neighbor algorithm, was executed using eCognition software [16]. Object-based classification technique is a powerful technique used in remote sensing to analyze imagery. Unlike traditional pixel-based classification, object-based classification technique focuses on groups of pixels that represent real-world features, like trees or buildings. This approach offers a more accurate and context-dependent understanding of the image. eCognition is a popular software program that allows to perform OBIA workflows. Data Import and Preprocessing, Image Segmentation, Defining and Training Classification Rules, Classification and Evaluation and Refinement Based on the evaluation are breakdown of the process [17].

Fig. 3.

On-site sample locations were determined using a high-precision GPS.

The assessment of the accuracy of the classified map was accomplished by comparing the collected ground truth points (constituting 30 % of the dataset) with the classified map. This assessment involved the use of kappa coefficient and overall accuracy as evaluation metrics [18].

To run the water yield component in the InVEST model, annual precipitation maps, along with minimum and maximum air temperature maps for the study watershed, were procured from the WorldClim website. These climatic data covered the period from 1990 to 2020 and had a spatial resolution of 30 s (equivalent to 1 square kilometer). These maps were compared to observational data obtained from the Department of Natural Resources and Watershed Management of the province for validation purposes. Additionally, climatic parameters of precipitation and temperature for the year 2040 were downloaded from the MIROC6 model using the ssp245 scenario [19]. The main objective of the research was to assess the impact of land use changes on water yield over historical and future time periods. Therefore, considering various climate models for the future requires a comprehensive and separate study. On the other hand, assuming the constant effects of climate for future time periods and analyzing the impact of land use changes solely on watershed water yield, were obtained.

The input factor maps for the water yield model throughout the study years within the GRB are depicted in Fig. 4. It's important to note that the factors pertaining to limiting root depth and soil water content were assumed to remain constant across the study years, as per the data provided by the Department of Natural Resources and Watershed Management of Golestan province in 2021. Other essential input parameters for the water yield model, encompassing biophysical characteristics, are detailed in Table 2.

Fig. 4.

Input variables for water yield model during the study years and future in the GRB.

Table 2.

Biophysical characteristics of Land cover classes in GRB.

LULC Classes	LULC_veg	Kc	root_depth (mm)
Forest	1	1	7000
Agriculture	1	0.6	1000
Pasture	1	0.65	2000
Residential	0	0.2	500
Water	0	1	500

The data used includes land use maps, average annual precipitation and reference evapotranspiration, soil depth, available water capacity, and a biophysical table (Table 2) containing characteristics of each land cover class [20].

The input factor maps for the water yield model throughout the study years within the GRB are depicted in Fig. 4. It's important to note that the factors pertaining to limiting root depth and soil water content were assumed to remain constant across the study years, as per the data provided by the Department of Natural Resources and Watershed Management of Golestan province in 2021. Other essential input parameters for the water yield model, encompassing biophysical characteristics, are detailed in Table 2.

2.2.3. Analysis and prediction of Land use/land cover map

2.2.3.1. Change detection

IDRISI Terrset LCM was utilized for the analysis and prediction of changes. Land cover maps for the periods of 1990–2000, 2000 to 2010, and 2010 to 2020 were employed for change analysis and detection. The magnitude of change, net change, change map, and transition map were assessed using charts and maps [21].

2.2.3.2. Modeling Land cover change potential

The calibration of the LCM in the study involved a systematic approach to selecting descriptive variables, modeling sub-models for land use transitions, generating probability maps for change, and fine-tuning model parameters to improve the model's predictive capabilities for assessing land use dynamics in the Gorgan River Basin. In fact, the Modeling of transport potential is the calibration starting.

In the realm of predicting land cover changes, a prevalent approach includes modeling transition probabilities between various land cover classes [22]. In this research, the LCM software within TerrSet was employed to forecast land use changes over the forthcoming two decades. The LCM model incorporates the Markov chain model, similar to the CA-Markov model, for calculating the probability transition matrix. However, LCM goes a step further by integrating the capabilities of the Markov chain model with the MLP and logistic regression training through error backpropagation. This amalgamation proves to be effective in simulating intricate processes [21,23].

Within the LCM framework, the probability matrix and change area are initially derived using the Markov chain model. This matrix quantifies the likelihood of transitioning from a specific land use to other land uses within the designated study period. Notably, due to significant changes observed within the study area, four sub-models for land use change transitions were delineated. These sub-models encompass 1. Transition from forest to pasture, 2. Transition from pasture to agriculture, 3. Transition from pasture to residential areas, and 4. Transition from agriculture to residential areas. Each of these sub-models makes use of distinctive descriptive variables modeled through the MLP method, ultimately yielding probability transition maps for each land use change. Finally, probability maps for change, based on environmental variables, were generated for each sub-model, subsequently facilitating the creation of change prediction maps for the year 2040, under the purview of three predefined management scenarios.

2.2.3.3. Selection of descriptive variables

In the creation of each of these sub-models, various descriptive factors were employed, and they were modeled using the MLP method, ultimately leading to the generation of probability maps that display how each land use change spreads spatially. The variables included in the model are: 1) Elevation, 2) Slope, 3) Distance from residential areas, 4) Distance from agricultural lands, 5) Distance from forest areas, 6) Distance from roads, 7) Distance from rivers, and 8) Empirical probability of change. These variables played a significant role in modeling the transferability capability within this study.

Steep slopes may limit certain land uses, while elevation gradients can impact vegetation patterns and water runoff dynamics. Areas closer to residential zones may undergo urbanization, while proximity to forests or agriculture may drive land use transitions. Roads may facilitate land conversion, while rivers affect decisions related to water availability and ecosystem services. Empirical probability of change helps calibrate the model to accurately reflect actual change processes within the study [18,24].

The slope was calculated using Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM) with resolution of 1 Arc-Second. Residential areas were identified from the prior land cover map from the calibration period, and the distance from this land use class was calculated. Road and river data were extracted from 1:25,000 topographic maps and modified as needed based on satellite images. To create the empirical probability of change (a qualitative variable), a transfer probability map was generated, depicting the likelihood of change from all land cover classes to residential areas. Considering the substantial changes in the residential area of the study region and a higher Cramer coefficient [25], the variables of distance from residential areas and dynamic road variables were given special attention. These variables exhibited greater variability over time compared to others, while all other variables remained constant. The Cramer coefficient, which gauges the relationship between variables and land cover changes, was analyzed. A Cramer correlation coefficient higher than 0.15 is considered acceptable for this coefficient in transferability modeling [18,26].

For modeling transferability, the MLP neural network was employed. It's recognized as a robust and common method for this purpose. It employs the backpropagation algorithm for forecasting changes. One of the advantages of using MLP is its ability to model non-linear relationships between variables. The neural network consists of three layers: input, hidden, and output. It was trained to determine appropriate connection weights between the input and hidden layers and between the hidden and output layers for classifying unknown pixels. To evaluate the transferability modeling with MLP, various metrics were calculated, including artificial neural network performance, accuracy, training Root Mean Square Error (RMS), and testing RMS. This comprehensive approach assists in the assessment of the model's reliability and predictive capabilities [21,27,28].

2.2.3.4. Prediction of land use changes

The likelihood of a class transitioning to another type was calculated using the Markov Chain method, as previously employed in studies by Refs. [29,30]. Subsequently, the modeling process was executed using a hard prediction approach. To accomplish this, all transitions were considered, leading to the creation of two distinct lists: one comprising host classes (which would experience a decrease in land), and the other consisting of claimant classes (which would gain land) for each host. The specific quantities for these transitions were determined based on the outcomes of the Markov transition matrices, as described in prior research [21,27]. A multi-objective allocation process was then undertaken to distribute land among all the claimants of a host category. The results of reallocating land for each host class were subsequently superimposed to generate the outcome. This approach has been documented in previous studies [31].

2.2.3.5. Validation of the model

To assess the accuracy of the LCM model in generating a land use map for the year 2040, the maps for the years 2000 and 2010 were initially utilized to predict the land use map for 2020. This prediction process involved the creation of a probability change matrix based on the predefined sub-models and the probability transfer maps, resulting in the land use change map for the year 2020.

Subsequently, the predicted map produced by the model, referred to as the comparison map, was compared to the land use map obtained through object-based classification, serving as the reference map. This comparative analysis was executed using the Validation module within the LCM model. The evaluation encompassed the overall coefficient, as well as the computation of Hits and Misses errors, to assess the accuracy for each land use class. Additionally, pixel quantities were calculated to gauge the accuracy of location and false alarms.

2.2.4. Water yield Model

The water yield generation model in InVEST is performed using spatial analysis and modeling within the ArcGIS software. This model has been implemented in various locations and decision-making contexts, such as Indonesia [32].

The water yield generation model is based on the Budyko curve and the average annual precipitation. The water yield production or Y(x) is calculated annually for each land pixel X using the following relationship:

$$\mathrm{Y}\left(\mathrm{x}\right)=\left(1-\frac{AET\left(x\right)}{P\left(x\right)}\right).\mathrm{P}\left(\mathrm{x}\right)$$ (1)

Where AET(x) represents the annual actual evapotranspiration for pixel x and P(x) represents the annual precipitation for pixel x. The actual evapotranspiration and precipitation balance, AET(x)/P(x), for the vegetated land cover component is determined based on the Budyko curve, which is obtained using equation (2).

$$\frac{AET\left(x\right)}{P\left(x\right)}=1+\frac{PET\left(x\right)}{P\left(x\right)}-{\left[1+{\left(\frac{PET\left(x\right)}{P\left(x\right)}\right)}^w\right]}^{\frac{1}{w}}$$ (2)

That in which PET(x) is the potential evapotranspiration, which is defined according to equation (3) as follows:

$$\text{PET}\left(\mathrm{x}\right)=K_c\left(l_x\right).{ET}_0\left(x\right)$$ (3)

Where ET₀(x) is the evapotranspiration for pixel x and K_c(l_x) is the crop evapotranspiration coefficient for pixel x in land use. The reference crop evapotranspiration is calculated using the modified Hargreaves equation based on the guidelines provided by the InVEST model, and it is obtained as equation (5).

$${\text{ET}}_{\mathrm{O}}=0.0023{\mathrm{R}}_{\mathrm{a}}{\left({\mathrm{T}}_{\max -}{\mathrm{T}}_{\min }\right)}^{0.653}\left(\frac{{\mathrm{T}}_{\max +}{\mathrm{T}}_{\min }}{2}+17.8\right)$$ (4)

In the above equation, ET₀ represents the reference crop evapotranspiration in millimeters, Ra represents the solar radiation in mega-joules per square meter, Tmax and Tmin represent the maximum and minimum air temperatures in degrees Celsius [33]. The raster map of Ra is prepared using the Solar Radiation function in ArcGIS 10.3 software.

W(x), which was mentioned earlier in equation (2), is a non-physical or empirical parameter that defines the properties of climate and soil. Further explanation of this parameter is provided below equation (5).

$$\mathrm{W}\left(\mathrm{x}\right)=\mathrm{Z}\frac{AWC\left(x\right)}{P\left(x\right)}$$ (5)

In the above equation, AWC represents the available water for plants. Soil texture and effective soil depth are influential factors in determining this factor. Z is a seasonal factor that represents the seasonal distribution of precipitation and the intensity of rainfall. In regions where precipitation occurs more in winter, this factor is assigned a value close to 10, while in regions where precipitation is evenly distributed throughout the year or predominantly in summer, it is assigned a value of 1. General conclusions from studies indicate that this factor has an approximate value of 4 in rainy areas, 9 in temperate regions, and 1 in seasonal areas.

3. Results

3.1. Production of land use maps and accuracy assessment

Land use maps were created by classifying land into five distinct categories, namely forest, agriculture, pasture, residential, and water bodies. This classification was accomplished using the object-based classification method, specifically employing the nearest neighbor algorithm [16]. The classification process was executed using eCognition software. The accuracy of the generated maps was evaluated, and the results, indicating acceptable accuracy, are provided in Table 3. As a cross check and enhancing the reliability, LULC map of European Space Agency (ESA) World Cover at 10m resolution for 2020 might also be used for model validation/comparison as depicted in Fig. 5. In fact, ESA WorldCover data for cross-validation strengthens the model validation process by providing a globally consistent and independent reference dataset. This reduces the possibility of bias.

Table 3.

Results of the assessment of the generated land use maps.

Year	1990	2000	2010	2020
accuracy	0.96	0.9	1	0.87
kappa coefficient	0.95	0.86	0.83	0.87

Fig. 5.

ESA world cover map (10m) and Landsat 8 cover map (30m) for 2020.

Following an accuracy assessment that compared the generated maps with ground truth data, the kappa coefficient values for the land use maps of the years 1990, 2000, 2010, and 2020 were computed and yielded values of 0.95, 0.86, 0.83, and 0.87, respectively. These values reflect the degree of agreement between the generated maps and the ground truth data. Additionally, the overall accuracy of these maps was determined and resulted in values of 0.96, 0.90, 0.88, and 0.87 for the respective years. These high accuracy values underscore the reliability of the generated maps (Fig. 6).

Fig. 6.

Land use map for historical period (1990–2020). For comprehensive details and access to the generated maps, please see the article by Ref. [18].

3.1.1. Change detection

Fig. 7 illustrates the distribution of land use classes within the GRB for the study years. Based on the findings presented in Table 4, the analysis of land cover changes in the GRB reveals that over the study period from 1990 to 2020, the most substantial changes occurred in the agricultural class, with an expansion of 553.0 square kilometers, corresponding to 7.3 % growth. Following closely, the grassland class experienced significant changes, with a reduction of 513.7 square kilometers, equating to a 6.8 % decrease. The forest class also saw a decrease of 279.5 square kilometers, accounting for a 3.7 % reduction. Furthermore, residential and urban areas witnessed an increase of 224.1 square kilometers, reflecting a 3 % rise during this period. This shift indicates a notable spatial trend of residential development concentrated in the northern and northeastern regions of the area, as illustrated in Fig. 8.

Fig. 7.

The areas (in square kilometers) for different land use classes for the study years.

Table 4.

Area and percentage changes of land use classes in GRB during the study years.

LULC	1990		2000		1990–2000		2010		2000–2010		2020		2010–2020		1990–2020
classes	Area		Area		Change		Area		Change		Area		Change		Change
	Km2	%	Km2	%	Km2	%	Km2	%	Km2	%	Km2	%	Km2	%	Km2	%
Forest	1703	22.6	1679	22.3	−23.6	−0.3	1589	21.1	−90.2	−1.2	1423	18.9	−165.7	−2.2	−280	−3.7
Agriculture	1817	24.1	1751	23.2	−66.5	−0.9	2124	28.1	373.4	5	2370	31.4	246	3.3	553	7.3
Pasture	3886	51.5	3948	52.3	61.7	0.8	3530	46.8	−417.9	−5.5	3373	44.7	−156.8	−2.1	−513	−6.8
Residential areas	127	1.7	160	2.1	32.8	0.4	272	3.6	112	1.5	352	4.7	79.4	1.1	224	3
Water Area	14.7	0.2	11	0.1	−4.2	−0.1	33	0.4	22.9	0.3	30	0.4	−2.9	−0	16	0.2

Fig. 8.

Land use change map between 1990 and 2020.

Between 2000 and 2010, agricultural lands saw the most significant expansion, with an increase of 373.4 square kilometers, corresponding to a 5.0 % change for this land use class during this timeframe. Conversely, the most significant decrease in grasslands was observed during this period, with a reduction of 417.1 square kilometers, signifying a 5.5 % decline. The most notable change in forested areas occurred between 2010 and 2020, with a reduction of 165.7 square kilometers, representing a 2.2 % decrease. Urban and residential areas experienced their most substantial growth between 2010 and 2020, with an increase of 351.5 square kilometers. This was followed by the period from 2000 to 2010, which recorded an increase of 272.1 square kilometers in urban and residential development.

3.2. Modeling of transport potential (calibration)

The results of studying the correlation between influential variables and land use changes are presented in Table 5 using the Cramer coefficient. The calibration (2000–2010) and validation (2010–2020) of these variables are provided. All the input variables had a Cramer coefficient higher than 0.15.

Table 5.

Values of the Cramer coefficient for the variables influencing land use changes in the GRB during the calibration and validation periods.

Variable	2000–2010	2010–2020
Elevation	0.33	0.32
Slope	0.35	0.34
Distance from Forest in 2000	0.41	0.41
Distance from residential in 2000	0.26	0.26
Distance from agriculture in 2000	0.29	0.31
Distance from river	0.19	0.21
Distance from Road	0.17	0.18
Qualitative	0.44	0.43

The maximum and minimum values of Cramer coefficient were observed for the qualitative variable and distance from the river. Table 6 also includes descriptive variables and other features of the MLP model for constructing sub-models. The accuracy level is mentioned as a parameter to express the validity of the MLP model.

Table 6.

Specifications of different sub-models for land use conversion in the period 2000 to 2010.

Sub-Models	Number of variables	Model	Momentum factor	Hidden layer neurons	Iterations	RMS training	RMS Test	Skill Measure	Accuracy (%)
Forest to Pasture	8	MLP	0.5	8	10000	0.4012	0.398	0.73	91.3
Pasture to Agricultural			0.6			0.4417	0.3501	0.69	90.4
Pasture to Residential			0.5			0.2125	0.2184	0.63	84.3
Agricultural to Residentials			0.5			0.2344	0.2881	0.56	72.4

The results reveal that the sub model representing land changes from pasture to agricultural land had the most substantial impact on the accuracy of the MLP model. Furthermore, in the sub model illustrating the conversion of forest to pasture, the geographic slope variable played the most pivotal role. In the sub model depicting the transformation of pasture into agricultural land, the variable of distance from the agricultural land boundary exerted the greatest influence on the formulation of these sub models.

Table 7 provides the land use probability matrix for the period between 2000 and 2010. This matrix serves as a foundation for modeling and predicting the spatial distribution pattern of land use classes for the year 2040, considering the probability of future changes and accounting for various management scenarios. The probability change maps for all four sub models are depicted in Fig. 9, offering valuable insights into the anticipated land use alterations.

Table 7.

Land use transition probability matrix between the years 2000 and 2010.

	2010
2000	LC/LU Classes	Forest	Agriculture	Pasture	Residential	Water Area
	Forest	0.7474	0.0891	0.155	0.0077	0.0007
	Agriculture	0.03242	0.5585	0.32	0.0692	0.0181
	Pasture	0.0479	0.2905	0.6165	0.0414	0.0038
	Residential	0.0175	0.4191	0.3714	0.1791	0.0128
	Water area	0.0018	0.1568	0.0232	0.0094	0.8088

Fig. 9.

Map of probability of land use changes (transfer potential) between a: 2000 and 2010: Forest to Pasture. b: 2000 and 2010: Pasture to Agriculture. c: 2000 and 2010: Agriculture to Residential areas. d: 2000 and 2010: Pasture to Residential areas.

3.3. Validation of land use modeling

In this study, the land cover map of the year 2000 was considered as the first-year map, and the land cover map of the year 2010 was considered as the second-year map in the software. At this stage, the demand for changes was determined using a Markov chain, and the prediction of changes for the year 2020 was made (Fig. 10).

Fig. 10.

Prediction of land use changes using the Multi-Layer Perceptron Artificial Neural Network method. A: Actual land cover map of the year 2020; B: Predicted land cover map of the year 2020.

The assessment results of the Multi-Layer Perceptron Artificial Neural Network model for the year 2020 show a strong concordance between the model's output and the observed changes, as indicated by an ROC (Receiver Operating Characteristic) value of 0.91. The success-to-failure ratio for the Multi-Layer Perceptron model stands at 0.66, while the Figure of Merit (FOM) index is calculated to be 19.3, as detailed in Table 8. Furthermore, the Kappa statistic is estimated to be 0.82. These findings collectively underscore the model's high reliability and efficacy in predicting land use maps for future years, especially when considering various scenarios.

Table 8.

Evaluation results of the accuracy assessment between the predicted land use map for the year 2020 and the actual land use map for the year 2020 in the GRB.

year	ROC	$\frac{Hits}{FalseAlarms}$	$\frac{Hits}{FalseAlarms+Hits+Misses}$	Overall Kappa
2022	AUC = 0.91	0.66	19.3	0.82

3.4. Predicting land use changes for 2040 based on defined scenarios

Land use changes were predicted for the next 20 years based on three defined scenarios in the GRB. Table 9 and Fig. 11 illustrate the area of each land use category in 2020 and 2040 for all three scenarios.

Table 9.

Area of each land use in the year 2020 and under the three management scenarios in 2040.

Landuse	LC/LU 2020		2040 (Continuation Scenario)		2040 (Conservation Scenario)		2040 (Mitigation Scenario)
	(Km2)	(%)	(Km2)	(%)	(Km2)	(%)	(Km2)	(%)
Forest	1423.4	18.9	941.4	12.5	1184.9	15.7	1414.6	18.7
Agriculture	2370.2	31.4	3531.4	46.8	2849.9	37.8	27437	36.3
Pasture	3372.9	44.7	2443.7	32.4	3044.2	40.3	2756.1	36.5
Residential	351.5	4.7	598	7.9	435.3	5.8	430.8	5.8
Water Area	30.4	0.4	33. 3	0.4	33.3	0.4	32.7	0.4

Fig. 11.

Area of each land use in 2020 and 2040 for each of the three management scenarios.

3.4.1. Scenario 1 (continuation scenario)

Scenario 1, labeled as the "Continuation" scenario in this study, posits a trajectory of land use changes that closely resembles the patterns of the past two decades, devoid of any constraints. The modeling outcomes rooted in this scenario project that if the forthcoming changes mimic the trends of the prior 20 years, the extent of forested land would dwindle from 1423 square kilometers–941 square kilometers. This equates to a substantial 33 % reduction in the current forested area over the next 20 years, amounting to 6.4 % of the total watershed area. Under the purview of this scenario, a 12 % decline in pastureland is anticipated, causing it to contract from 3372 square kilometers to 2443 square kilometers. Conversely, agricultural land is expected to expand by approximately 1161 square kilometers during this period, encompassing about 15.4 % of the entire watershed area. A noteworthy transformation during this period is the escalation of residential areas by 246 ha, signifying a notable 27.3 % surge. These residential expansions are primarily concentrated around the existing cities and villages within the region and along the roadways. Considering these findings, it becomes apparent that this scenario forecasts a substantial loss of pasture and forested areas.

3.4.2. Scenario 2 (conservation scenario)

The second scenario, denoted as the "Conservation" scenario, introduces protective measures aimed at averting 80 % of land use conversions from agricultural land and pastures to residential areas, particularly villa construction. This scenario also seeks to curtail the likelihood of forest to agricultural conversions by 60 % and reduces the probability of pasture to agricultural transitions by 30 % within the study area. By implementing this scenario, it becomes possible to prevent the conversion of approximately 162 square kilometers, constituting 2.2 % of agricultural land, into residential areas. Furthermore, under this scenario, the forested area is projected to decrease from 1423 square kilometers in 2020 to 1184 square kilometers. This strategy helps prevent a 23.3 % reduction in the forested area when compared to Continuation Scenario. Additionally, this scenario anticipates a 4.4 % reduction in pasture area, decreasing it from 3372 square kilometers to 3044 square kilometers. The implementation of this scenario mitigates the loss of approximately 600 square kilometers, which corresponds to 8.0 % of the pasture area. Agricultural land, in this scenario, will experience a more modest 6.4 % increase, equivalent to 479 square kilometers, thus restraining 9 % of the projected increase in this land use (681 square kilometers). As for residential areas, their extent will contract from 351 square kilometers to 435 square kilometers, with a minimal 83 square-kilometer increase over the ensuing 20 years. The implementation of this scenario paves the way for achieving the highest level of protection for forest and pasture areas within the region.

3.4.3. Scenario 3 (mitigation scenario)

The third scenario, termed the "Mitigation" scenario, is defined as follows: land use changes within the region are permissible, but with a key restriction—these changes will not take place in areas characterized by a slope greater than 15 % and a vegetation coverage density exceeding 50 %. These constraints were imposed within the ArcMap environment by integrating the slope layer and vegetation coverage density layer of the region and subsequently applying them to the land use map devised for 2040, as per Continuation Scenario. As indicated in the results presented in Table 4, this scenario attains the highest level of protection for forested areas. Only 0.1 % of this land class, equivalent to 7.8 square kilometers, is anticipated to experience a decrease in the forthcoming 20 years. Furthermore, based on the outcomes of this scenario, a mere 0.2 % (82.6 square kilometers) of pastureland will undergo conversion into other land uses. According to this scenario, approximately 45.10 % of agricultural lands that are unsuitable for cultivation due to the region's slope will be repurposed for afforestation and pastureland through protective management measures, encompassing 788.4 square kilometers. Additionally, within this scenario, a 1.1 % increase (83 square kilometers) in residential areas is anticipated within the specified timeframe. This scenario is poised to provide the utmost safeguard for forested lands within the region.

3.5. The spatiotemporal variation of water yield

Fig. 12 represents the distribution maps of water yield in the GRB, and Table 10 illustrates the water yield values for each land use category in the study years as well as future scenarios.

Fig. 12.

Distribution Map of Water yield (mm) in GRB during the Study Years and Management Scenarios.

Table 10.

Water yield Quantity (million cubic meters) for Different Land Use Classes in GRB during the Study Years.

LC/LU	1990	2000	2010	2020	2040
					(Continuation Scenario)	(Conservation Scenario)	(Mitigation Scenario)
Forest	23.9	16.9	17. 8	18.6	7.9	11	15.3
Agriculture	142.2	107.9	122	148.3	190.4	166.1	152.8
Pasture	136.2	80	96	107.6	65.4	84.1	75.8
Residential	22.4	24.9	44	45. 6	99.9	71. 6	71
Total (sum)	324.7	229.6	279.7	320.1	363.6	332.8	314.9

Considering constant temperature and precipitation values for predicting future water yield relative to the year 2020, a decrease is observed under the first and second scenarios, while an increase is noted under the third scenario, which contradicts the results obtained from future climate projections. This discrepancy indicates a greater influence of future climate on water yield compared to land use changes (Table 11).

Table 11.

Water yield Quantity (million cubic meters) for the future Considering constant temperature and precipitation values in GRB.

Variable	Scenario1	Scenario2	Scenario3
Water Yield (MCM)	267	305	342

The prioritization of 56 studied sub-watersheds based on water yield and the total water yield quantity for each land use class per unit area are presented in Fig. 13.

Fig. 13.

A) The prioritization of 56 studied sub-watersheds based on water yield per unit area during the study period. B) The total water yield quantity for each land use class per unit area in 2020.

The research outcomes, encompassing the identification of factors affecting water yield in the GRB, along with the sensitivity analysis results of the InVEST water yield model, are detailed in Table 12, Table 13, respectively. It is worth noting that in Table 11, the year 1990 serves as the baseline year. In each model run, one specific input factor (as listed in the table) was established for the year 1990, and its changes were computed over the course of the research period.

Table 12.

Results of determining the impact of influential factors on water yield during the study period in GRB.

Variable factor	Baseline Calculation (m3)	Calculation affected by changes (m3)	Change (%)
Precipitation	324681952	478029237.9	47.2
Evapotranspiration		412346079	27
LC/LU		281823934.3	−13.2
Seasonality Factor (Z)		172405504.1	−46.9

Table 13.

Results of sensitivity analysis of the influential factors in the water yield model in GRB.

Variable Factor	Change (%)	Relative Sensitivity Coefficient
Precipitation	15	0.504
	−15	−0.554
Evapotranspiration	28.5	−0.254
	−28.5	0.341
Seasonality Factor (Z)	14	−0.067
	−14	0.057

In order to better compare the changes in land use with the water yield values from 1990 to 2040 and under different management scenarios, separate graphs of each land use and the amount of runoff from each class were plotted (Fig. 14).

Fig. 14.

Number of changes in land area and water yield for each land use from 1990 to 2040.

4. Discussion

The findings of the current study, as presented in Table 10, reveal a decreasing trend in water yield within the GRB. Specifically, water yield has declined from 324.7 million cubic meters (MCM) in 1990 to 279.7 MCM in 2010, with a subsequent increase to 320.1 MCM in 2020. Several key factors contribute to this trend, including a 14.5 % decrease in precipitation and an 18.8 % increase in average annual temperature during the 20-year period from 1990 to 2010 in the study watershed. This decrease in water yield aligns with [34], who highlighted a strong correlation between annual precipitation and water yield, emphasizing that increased water yield in the Qinghai Lake watershed in China is a result of augmented annual precipitation. The increase in water yield from 2010 to 2020 can be attributed to several factors, including the expansion of residential areas, a significant reduction in forested lands, and an increase in agricultural land [35].

Further investigation into annual water yield across different land uses within the GRB reveals that agricultural and forested lands allocate the maximum and minimum water yield, respectively, with average values of 130.1 MCM and 19.3 MCM per year during the study period. This finding is in line with [36], who also observed the highest water yield in agricultural land. The extensive coverage and presence of agricultural land in areas with steep slopes, resulting in reduced infiltration, increased surface runoff, and consequently increased water yield, can be considered the primary reason for the maximum water yield in this land use.

Pasture and residential areas are the second and third-ranking land use categories in terms of water yield within the GRB, with values of 104.9 MCM and 34.2 MCM, respectively. However, when considering water yield per unit area for comparative evaluations, residential, agricultural, Pasture, and forest land uses rank first to fourth, respectively, with average water yields of 1556.5, 649.7, 286, and 120.9 cubic meters per hectare in the study watershed (Fig. 13B). This discrepancy is attributed to the higher water losses in forestland use due to factors such as infiltration, evaporation, and transpiration driven by dense vegetation cover [37]. These findings are consistent with the study by Ref. [38], which revealed higher infiltration rates in forestland use compared to orchard and agricultural land uses. Zare et al. [39] also demonstrated that forestland use had the minimum water yield using the L-THIA model in the Kasilian watershed. The results of this study highlight the maximum water yield per unit area in residential land use, which is a consequence of impermeable surfaces leading to lower infiltration and higher surface runoff production in this land use category [40,41]. The relatively high-water yield, especially in agricultural land uses compared to other land uses, except residential areas, can be attributed to lower water infiltration in these land uses, as indicated by the findings [40] regarding low permeability in agricultural land use.

The decreasing trend in water availability during the historical period of 2020–1990 can be attributed to increased water consumption across various land uses due to population growth, meteorological factors such as precipitation and evapotranspiration changes [[42], [43], [44], [45]], as well as expansion of land uses with high water demand such as irrigated agriculture and residential areas [46]. On the other hand, the expansion of residential areas leads to reduced soil infiltration and water absorption, resulting in faster evaporation of water.

The shift in agricultural patterns from traditional systems to more industrialized and mechanized systems can result in increased water consumption and decreased water availability in the watershed. In fact, disparities in water resources for various uses such as agricultural, urban, and industrial purposes can lead to changes in water consumption and management, ultimately resulting in decreased water availability in the watershed. Urban redevelopment and expansion may alter the water usage patterns in different areas and lead to changes in surface and groundwater flows.

In the GRB, three management scenarios were modeled using the InVEST ecosystem services model to predict water yield changes by 2040. In the "Continuation" scenario, which assumes the continuation of current trends, water yield is projected to increase due to expanding agricultural and residential land uses, resulting in land degradation and increased soil erosion. The "Conservation" scenario, with restrictions on converting agricultural land and pastures to residential areas and forest to agricultural land, leads to a moderate increase in water yield. The "Mitigation" scenario, which places restrictions on land use changes in areas with high slopes and dense vegetation cover, shows a slight decrease in water yield, emphasizing environmental preservation. Among these scenarios, the "Mitigation" scenario is recommended for its focus on natural ecosystem protection and reduced risks of soil erosion and flooding in the region. The choice of scenario should balance environmental conservation and development needs.

Given the trend of land use changes over the mentioned time intervals, it is observed that there has been a reduction in the area of forest and pasture lands, while the area of agricultural and residential lands has increased. Due to fewer changes in land use in the second scenario (Conservation) compared to the first scenario (Continuation), less variation in hydrological variables is also reported in this scenario. To address the challenges of water management in different scenarios, we propose tailored water management measures for different land use change scenarios. For the baseline (trend) scenario, we recommend implementing sustainable land management practices, promoting reforestation, and monitoring water quality. In a conservation scenario, establishing riparian buffers, encouraging green infrastructure, and investing in water conservation measures are key. For a mitigation scenario, we suggest focusing on targeted land use planning, prioritizing land restoration in degraded areas, and developing adaptive water management strategies through stakeholder collaboration. These measures aim to mitigate the specific implications of Land Use and Land Cover (LULC) dynamics on water resources, enhancing overall watershed resilience and sustainability.

The analysis of ecosystem service provision in the GRB reveals that the continuation of current land use trends will lead to increased water yield in agricultural and residential areas but decreased water yield in forest and pasture areas. Spatially, sub-watersheds in the northern and northwestern parts of the GRB exhibit the lowest water yield, with an increase observed as one moves southward and eastward. Notably, despite significant land use changes, the priority of sub-watersheds in terms of water yield remains largely unchanged. Our findings indicate that precipitation is the most sensitive factor in the InVEST water yield model, which is consistent with previous studies highlighting the importance of precipitation in hydrological modeling. In terms of relative sensitivity, evapotranspiration and seasonality follow precipitation. The impact of seasonality can vary across watersheds, depending on factors such as precipitation patterns. Understanding these sensitivities is crucial for effectively managing water resources in the GRB and has implications for future policy decisions.

5. Conclusion

In conclusion, the comprehensive analysis of the GRB over the past three decades highlights a discernible shift in hydrological services influenced by land use changes. Overall, we found that water yield in GRB has experienced a notable decline, attributed to a combination of decreased precipitation and increased temperatures. This study has revealed that agricultural expansion and residential development have led to increased water yield, while forested areas have seen a decrease. In contrast, agricultural lands have demonstrated the highest water yield, likely due to their expansive presence in areas with steep slopes, which increases water yield.

The implications of land cover and land use dynamics on water management are significant and multifaceted. Based on our findings, the effect of land cover and land use change on water management can be categorized into several key consequences: Firstly, the study highlights that the transformation of forests and pastures into agricultural or residential areas have a significant impact on water yield in the GRB watershed. Secondly, regarding surface runoff regulation, the transformation of land cover, particularly changes in forest ecosystems, plays a vital role in regulating surface runoff and mitigating flood occurrences. Forested areas act as natural buffers, absorbing excess water and reducing the risk of flooding. Thirdly, this study underscores the crucial importance of understanding the ecosystem services provided by different land cover types for effective watershed management. By incorporating ecosystem services and water resource conservation strategies in water management plans, we can achieve sustainable water resource use and conservation. Lastly, our scenario analysis of sustainable management demonstrates that, by assessing various management scenarios and their impacts on water resources, decision-makers can make informed choices that balance environmental conservation with developmental needs.

The InVEST ecosystem services model projects different scenarios for 2040, with the "Mitigation" scenario emerging as the most favorable for ensuring environmental preservation. It underscores the delicate balance between maintaining natural ecosystems and pursuing developmental agendas. The spatial analysis indicates a consistent pattern of water yield across the GRB, with the most significant changes occurring in the northern and northwestern sub-watersheds. Precipitation remains the most sensitive factor influencing water yield, highlighting its critical role in hydrological modeling and water resource management in the basin. Ultimately, this study serves as a pivotal reminder of the intricate connections between land use patterns and hydrological services, reinforcing the need for careful consideration of these dynamics in watershed management and ecosystem conservation strategies Future research can benefit from a more comprehensive approach that considers the human element. This includes understanding how population growth influences land use changes and acknowledging that models might not always perfectly predict conservation outcomes. By taking these factors into account, decision-makers can develop land-use strategies that balance human needs with environmental protection and achieve sustainable development.

CRediT authorship contribution statement

Masoomeh Yaghoobi: Writing – original draft, Visualization, Methodology, Investigation, Formal analysis. Aram Fathi: Writing – original draft, Methodology, Conceptualization. Shahryar Fazli: Writing – review & editing, Visualization, Resources. Wenzhao Li: Writing – review & editing, Supervision. Elham Haghshenas: Validation, Software. Vahid Shokri Kuchak: Methodology. Hesham El -Askary: Writing – review & editing, Supervision, Funding acquisition.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:Hesham El-Askary reports financial support was provided by US Department of Education. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.