Land use and land cover change (LULCC) is considered one of the major drivers of climate change, although climate change can also foster direct or indirect influences leading to LULCC. The objective of the presented study is to offer a strategic observation frame as the land use and land cover (LULC) transitions are grouped to define the cover flows (CFs). The Küçük Menderes River Basin (KMRB), which is located in the west of Turkey was examined as the case study. Through LULCC modelling via the employment of multi-layer perceptron (MLP), cellular automata (CA), and Markov Chain methods, future LULC maps were projected up to the horizon of 2050. Hydrologic responses of the basin to LULCC were determined by the developed hydrologic model, which is generated by the Soil and Water Assessment Tool (SWAT). The superimposed impacts of the examined effects of LULCC have been investigated by the CF types. This way, the individual impacts of the CFs have been assessed. In the case of the KMRB, projected annual runoffs for the year 2050 cover map represent a 9.06% reduction and the major responsible CF type for this reduction is the conversion from forest to non-irrigated agricultural land cover by 22.90%.
Future water availability of the KMRB is investigated by the defined cover flows.
About a 9% reduction in surface water volume is expected by the year 2050 at the KMRB outlet.
Conversion to non-irrigated agriculture is the major cause of the water volume reduction in the basin.
The Mediterranean region is one of the most affected areas worldwide by climate change. The region is expected to be more vulnerable to climate change-driven pressures and challenges in the future. As water resources decline on a global scale, examining the causes and the possible solutions is more relevant and motivating for researchers. One of the major drivers affecting the surface flow regime is land use and land cover change (LULCC) (Nie et al. 2011; Kashaigili & Majaliwa 2013; Khoi & Suetsugi 2014; Tan et al. 2015). Only the growth of the local population and, to a lesser extent, its increase in consumption were assumed to drive LULCC by the changing needs of the people and increasing demand in urban, industrial, and agricultural areas (Lambin et al. 2003). So far, the researchers mainly investigated LULCC in the context of transitions of pristine forests to agricultural uses (deforestation) or the destruction of natural vegetation cover by overgrazing, which leads to desertification (Otterman 1974). These transitions were assumed to be irreversible and spatially homogeneous and to progress linearly. Over the last many decades, numerous pieces of research have contributed to the development of a comprehensive understanding of the phenomenon, related measurements, and predictive models on the subject. The impact of land change on regional climate by the modified surface albedo has been recognized since the mid-1970s (Charney & Stone 1975; Sagan et al. 1979). Disrupted local evapotranspiration affects the imbalance of the water cycle, which is aggravating climate change-related threats. While deforestation (Chang 1993; Lal 1993) and forest-to-grassland or cropland transitions are associated with an increase in annual streamflow (Munoz-Villers and McDonnell, 2013), transitions between grassland and cropland are often not examined thoroughly because they are structurally similar vegetation types, even though their impact could be significant (Nosetto et al. 2012).
Projecting and evaluating the LULCC effects on the water resources could be a valuable tool in the target of mitigating climate change-related challenges. Here, hydrologic models that help simulate the dynamics of hydrologic response to LULCC were proven to be greatly significant in assessing water resource potentials under non-stationary conditions caused by natural and anthropogenic impacts. One of the most powerful instruments to this end is the Soil and Water Assessment Tool (SWAT) model, being a comprehensive, widely used means that is capable of simulating basin hydrology (quantity and quality) and forecasting the hydrologic and environmental impacts of land use, land management practices, and associated changes.
In the case of Küçük Menderes River Basin (KMRB), the basin has been struggling with existing water scarcity and pollution caused by the increasing demands of agriculture and industry sectors and domestic use. Such pressures have been aggravated by climate change in recent years, which for the basin resulting in decreased precipitation, water quality, water availability, increased temperatures, and evaporation. Several studies have covered the topics of water availability and water quality conditions and their impact on the environment in the KMRB. In the ‘Pollution Prevention Action Plan’ of the Turkish Ministry of Environment Urbanization & Climate Change (2016), the Ministry has reported that industrial (especially food industry), and agricultural activities, untreated urban discharges, and marble processing plants are the major polluters in the basin. Regarding water scarcity, Peksezer Sayit & Yazicigil (2012) have investigated the potential of artificial aquifer recharge in the basin because of the declining groundwater levels, which has been also highlighted by Yagbasan (2016). Eris et al. (2020) have examined the meteorological droughts that occurred in the basin and concluded that the KMRB has experienced severe prolonged droughts, which increased in severity and decreased in frequency in time. Despite these valuable scientific contributions aiming toward a better understanding of the water resources of the KMRB, a comprehensive evaluation of LULCC and the future water availability in the basin has been lacking.
The presented study mainly employs the SWAT modelling tool to evaluate the different LULCC scenarios projected in the KMRB. There are several SWAT modelling applications regarding water resources assessment, where the land-use changes and climate change impacts are included for the development of a better representation of hydrological processes (Anand et al. 2018; Busico et al. 2020; Fan et al. 2021; Samal & Gedam 2021). Adding to that, the presented study provides a detailed examination process, where each relationship between specific LULCCs and their impact on surface water resources can be assessed. LULCC modelling and projections included spatial estimates of major change processes that were represented by so-called land cover flows (LCFs) (Gómez & Páramo 2005) and adapted to hydrologic simulations.
MATERIALS AND METHODS
Study area and datasets
The study utilizes two main softwares, which are TerrSet and ArcSWAT. In the LULCC modelling part of the study, the Coordination of Information on the Environment (CORINE) database of Copernicus services supplied the LULC inputs (EEA 2021). Since the KMRB is a severely modified basin, digital elevation models (DEMs) online could not provide the correct topology aimed for the study. Therefore, DEM for the basin in 10-m resolution was developed over the course of the study by 1:5,000 maps and point elevations that were acquired from the State Hydraulics Works (DSI). The distribution of protected areas (PAs) and Urban Development Areas (UDAs) was produced from the 1:100,000 scaled Izmir-Manisa Landscaping Plan acquired from the Turkish Ministry of Environment and Urbanization. The remaining set of inputs for LULCC modelling was generated by the ArcGIS software tools using DEM and CORINE data sets mainly.
In order to synchronize projected LULC maps based on CORINE and ArcSWAT hydrologic model, CORINE and ArcSWAT LULC codes were matched in pairs (Table S1, Supplementary material). Some of the CORINE codes were grouped to avoid a larger number of transitions in LULCC modelling. Perennial vegetation such as orchards, olive trees, vineyards, etc. was classified under the ‘OLIVE’ ArcSWAT code since the dominant perennial vegetation of the basin was the olive trees, and the rest could be negligible. On the other hand, irrigated and non-irrigated agriculture were distinguished as AGRR_W and AGRR_nonW, respectively.
Land Change Modeller (LCM), which is a tool of TerrSet software created by Clark Labs, was used for LULCC analysis and modelling. The main employment of the tool was to evaluate the LULCCs to estimate the land-use category via an artificial neural network (ANN) and Markov Chain assessment (Ansari & Golabi 2019). The development of the model included: (1) spatial change analysis for historical LULC data, (2) relationships of the driver variables with historical LULC data, (3) transition sub-model development (ANN and Cellular Automata-Markov Chain), (4) validation of the sub-model, and (5) future predictions.
While defining the relationship between driver variables and LULCs, Cramer's V coefficient was employed to analyze the correspondence between variables and LULCC processes (Shooshtari & Gholamalifard 2015). Cramer's V coefficient has a range between 0 and 1. A value near one indicates a higher potential of being an instructive variable (Wang et al. 2016). Cramer's V coefficient value close to 0.15 or higher is considered convenient, whereas those of 0.4 or exceeding are evaluated as good (Shooshtari & Gholamalifard 2015). Calculated values of the Cramer's V coefficients of the driver variables are given in Table S2 (Supplementary material).
A transition is the outcome of developments in different domains. A transition can be described as a set of linked changes, which may boost each other yet take place in completely different areas, such as technology, the economy, institutions, behaviour, culture, ecology, etc. Since transitions are multi-dimensional with various dynamic layers, several developments must come together in several domains for a transition to occur (Martens & Rotmans 2005). This complexity gives the transitions their unstable, stochastic characteristics. Thus, there is a strong concept of uncertainty in LULC transitions. In the study, transitions of the LULCs between 2006 and 2012 were also evaluated to grasp the dynamics and fundamental character of LULC alterations. Major transitions in the basin are the decrease in forests (23 km2), barren lands (11 km2), and the increase in pastures (30 km2) (Table S3, Supplementary material).
Sub-model development of the model has been undertaken by the multi-layer perceptron (MLP) method. MLP is a sort of ANN method, which is widely used for modelling complicated processes and structures (Pijanowski et al. 2014; Gibson et al. 2018). As a non-parametric method, MLP has the benefit to model the complex practical interrelations of LULC classes. Unlike other models (e.g., the Weights of Evidence and Logistic Regression), MLP is revered as a more appropriate tool for estimating future LULCCs (Kleinbaum & Klein 2002). The neural network outputs can state the alterations of varied LULC categories more successfully than single possibilities attained by the Weights of Evidence method (Wang et al. 2016). The MLP has the verified advantage of handling non-linear associations without the need for transformed variables (Pérez-Vega et al. 2012). This method has been utilized in various studies to support dual and multi-class LULC alteration models (Almeida et al. 2008; Lin et al. 2011).
The SWAT is a spatially distributed, continuous-time, a process-based hydrological model developed by the research service (ARS) of the United States Development of Agriculture (USDA). The main elements of the SWAT model are meteorology, hydrology, soil properties, plant growth, nutrients, pesticides, bacteria, pathogens, and land management. The watershed is normally first divided into sub-basins based on topography, then hydrologic response units (HRUs), the smallest watershed unit of the SWAT model, which contains similar land use and soil type combinations within the sub-basin. The hydrological process at the sub-basin level is modelled by obtaining HRUs. Various inputs such as the digital elevation model (DEM), soil type, land use, and slope affect the size of the HRU.
In the SWAT model setup for the presented study, the locations of flow gauge stations and storage structures were defined as outlet points, having a total of 50 sub-basins generated. Basin hydrology was simulated considering the monthly river flows values observed at the most downstream outlet and in the period 2000–2012, where the meteorological and hydrometric data were commonly available (Figure S1, Supplementary material). While the first year (2000) has been defined as the model warm-up period, the simulations were completed for calibration of the 2001–2007 period and for the validation processes of the 2008–2012 period. Based on this selection of the period for calibrating and validating the SWAT model, the 2006 CORINE data for LULC was used accordingly.
Within the scope of the study, soil types and properties for the study area were defined in the model and a soil map was generated for the model area on a 1-km resolution. The HRU distribution was derived from the overlaid layers for land uses, soil characteristics, and slopes. As meteorological data, precipitation, maximum temperature, minimum temperature, relative humidity, wind and solar radiation values of the existing meteorology stations in and around the basin available with long-term daily-based measurements were incorporated into the model. The demand pressures that exert on surface water potential in the basin were included in hydrologic simulations through the definition of irrigation water requirements of the agricultural parcels at the level of HRUs, which are greatly supplied from the groundwater sources.
Defining cover flow models
The LULCC effect on water resources is a highly significant topic, yet most researchers focus only on the impact of a limited number of LULC types. The study examines in this respect all available LULC transitions that may appear in the study area. Since the number of total transition types is too high, certain LULC transitions were grouped to define the cover flows (CFs) (Table 1). The remaining set of transitions (coloured in grey) was either not observed or indeed negligible (i.e., less than 10 km2). The cover flow matrix presents both the CORINE nomenclature and associated ArcSWAT codes to describe the transitions among CFs, which were inspired by the process of defining land CFs between the CORINE land cover classes (Gómez & Páramo 2005). Here, internal transitions are regarded as ‘No Change’ (NC).
The transitions represent, individually or in groups, certain LULCCs indicating drivers such as urban residential sprawl, deforestation, etc. Categorizing these CFs gives the advantage of examining them strategically (Table S4, Supplementary material). Twelve models were defined to assess the impact of the CFs on surface water resources by ArcSWAT. Toward predicting the future LULC distributions, the year 2050 was chosen as a benchmark as it represents a midpoint between near and far future periods (i.e., 2025–2050, 2050–2075, 2075–2100). Thus, this group of models considers only the isolated impact of LULCCs between 2006 and 2050 by ignoring other variables such as meteorological, physical, etc. that would eventually contribute to superimposed impacts on water availability. Models were arranged in accordance with their total area sizes transitioned. Smaller change patterns that represent certain types of CFs were grouped together (Table 2).
|Model no. .||CF type .||Total area transitioned (km2) .||Model no. .||CF type .||Total area transitioned (km2) .|
|1||CF4.2||164.8||7||CF2.3, CF2.4.1, CF3.1.1, CF3.2.1, CF3.3.1||70.3|
|Model no. .||CF type .||Total area transitioned (km2) .||Model no. .||CF type .||Total area transitioned (km2) .|
|1||CF4.2||164.8||7||CF2.3, CF2.4.1, CF3.1.1, CF3.2.1, CF3.3.1||70.3|
RESULTS AND DISCUSSION
In predicting future land-use change, the sub-modelling of transitions has been practiced in LULCC studies at various spatio-temporal scales. The generality of LULCC models embraces a vicinity within which a trial is created for modelling the probabilities of land-use transitions from one category to another (Eastman et al. 2005). In the study, the model employed the MLP CA-Markov Chain algorithm multiple times for the evaluation of earlier (2006) and later (2012) LULCs, calibration period (2018), and driver variables. Out of a total number of 68 LULC class transitions, eight most topic-related ones were selected for assessments in the transition sub-model due to the limitations of including land cover change processes in each simulation trial. Thus, the evaluated results of the sub-model, which has the most effective results for modelling, were confirmed (Table S5, Supplementary material). Here, the skill measure was calculated as the precision of transition projection minus the precision anticipated by chance (Mas et al. 2014).
Furthermore, the validation was assessed by the relationship between the predicted LULC map and the actual LULC map for the year 2018. ROC analysis was carried out for each land type, and AUCs of the LULC classes are calculated (Table S6, Supplementary material). In Figure S2 (Supplementary material), true-positive and false-positive ROC curves were given for each LULC class. ROC curve apparently approaches true-positive values, which are higher than false-positives, in a way to indicate high accuracy.
The 2006 and 2012 LULC maps function as the input for LULCC model calibration. 2018 LULC map has also been utilized to validate the estimated map for the year 2018. After validating the transition sub-model and the projection efficiency of the LULC map, the maps for future periods were estimated. Through employing the LULC map for 2018 as the base year along with the considered transition potentials and the probability matrix, the future LULC maps were predicted for the selected years 2025, 2035, and 2050 (Table 3). For each future prediction, the sub-model including MLP neural network analysis was implemented separately to identify the influences of the transitions that might be involved within the matrix of Markov chain transition probabilities for future estimation (Ahmed & Ahmed 2012).
|Category .||Area (km²)|
|2025 .||2035 .||2050 .|
|Category .||Area (km²)|
|2025 .||2035 .||2050 .|
The study results were checked against the Global Land Cover for the Year 2050 of the Living Atlas of the World (Pisut 2020). The corresponding prediction layer shows the global land cover that was modelled for the year 2050 at a pixel resolution of 300 meters. In a comparison effort between this global source and the study results generated, ROC analysis was conveyed, in a way to result in the ROC figures given in Table S7 (Supplementary material).
In order to model the monthly discharge at the outlet flow station, the period 2000–2012 was taken into consideration in ArcSWAT, and by using the calibration algorithm (SUFI-2) in the SWAT CUP utility, performing sensitivity analysis and completing the required calibration and verification processes, hydrologic model setup was carried out. Out of a total number of 46 input parameters, 14 were found to be the most sensitive within the study extent (Table S8, Supplementary material). The selected parameters were then calibrated according to the SWAT CUP results by either the absolute (a) or relative (r) method. In the absolute method, the original parameter values were calibrated by adding or subtracting a calibration value. For the relative method, the original parameter values were calibrated by multiplying the original value by ‘1 + calibration value’.
The estimated change areas that are linked to different CFs were adapted individually onto the 2006 map of the land covers in order to reflect in each trial the impact of individual CFs over the hydrologic scenario simulations developed for the target year 2050 and thus estimate the simulated hydrological responses of the basin against each CF separately.
The results of the study showcase both the individual and accumulated contributions of the CFs to the basin hydrologic response. With every CF considered through the estimated representations over the 2050 LULC map, volumetric water quantities are calculated at the basin outlet with a reduction of 9.06% under the combined impact of all CFs. In this generic view, some individual CF definitions define more distinctive figures such that conversions from forest to non-irrigated agriculture (CF3.2.2) causes a 22.90% reduction in water volume computed. Although deforestation causes an increase in surface runoff, in this case, conversion to non-agriculture (or to agriculture in general) causes a decrease in surface runoff since; the precipitation is primarily consumed by the non-irrigated crops, and the agricultural activities (tilling, mowing, grazing, etc.) help destroy the microbiotic crust over the soil surface, and help increase soil infiltration rate, and therefore causes the decrease in surface runoff (Pingping et al. 2013; Amami et al. 2021). It also should be noted that while streamflow is given in volumes, average discharge values of the model results might highlight the water scarcity of the basin indeed. The reference model represents an average discharge of 3.90 m3/s and the most impactful model (CF3.2.2) represents an average discharge of 3.17 m3/s. Thus, the significant discharge difference determined in the study was calculated as 0.73 m3/s due to the low streamflow observed in the KMRB.
The second dominant impact for the reduced water quantities with associated rates of 11.09% is mainly described by the internal conversions of agriculture or forest/pasture conversions to agriculture. The remaining but still significant change pattern (below 10%) in the loss of available water estimated for the basin is accompanied by the simulated changes through the conversions (1) from irrigated to non-irrigated, (2) from annual to perennial crops, and (3) from all available land cover types to barren soils. On the other hand, there are apparent volumetric increases in the basin water budget via the processes of natural land creation from agriculture and barren lands with the highest percent rates (about 15%). Another indication of the volumetric increases in the water toward the horizon of 2050 is represented with minor significance, for example, by the soil sealing impact due to relevant factors like the urban sprawl, and/or deforestation.
The hydrological responses of the basin help conclude that the current intensive agricultural activities play a significant role in the evolution of the available water resources in the basin since the highest impact rates are associated with water resources through a number of processes mainly relating to the agriculture sector (the conversions to agriculture (CF3.2.2) or internal conversions among agriculture (CF3.1.2)). The remaining set of land cover transitions seems to make fluctuating impacts via variable increases or decreases in the available water quantities. Furthermore, the most prominent contribution of the presented study arises in association with its comprehensive framework, methodological detail level, and precise data instruments involved in the overall approach where each CF is defined and their impact assessed individually and/or grouped. The study is believed to contribute to the integrated efforts for land-use change analyses and hydrologic modelling perspectives in a novel understanding developed through the adaptation of the CFs onto the hydrologic response estimations.
The findings of the presented study have been obtained by the IKLIMRISK (2021) (Hydrological risks and water quality change for sustainable water resources management under climate change) project, which was sponsored by the 1003 Programme of the Scientific and Technological Research Council of Turkey (TUBITAK).
DATA AVAILABILITY STATEMENT
All relevant data are included in the paper or its Supplementary Information.
CONFLICT OF INTEREST
The authors declare there is no conflict.