Introduction

Land-use activities which are converting natural resources to other land uses are resulting in many changes in the natural ecosystem (Foley et al. 2005). Converting forest to rangeland and agriculture (Fathizad et al. 2015) and also rangeland to cropland (Faramarzi et al. 2010) has been recognized as the main problems of natural ecosystem in semiarid rangeland in western Iran. These land-use practices are altering the hydrological cycles, e.g., groundwater recharge/discharge and groundwater quality in the ecosystem (Dams et al. 2008; Lin et al. 2007; Scanlon et al. 2005). On the other hand, groundwater is considered as one of the main freshwater resources of the world (Siebert et al. 2010), particularly in arid and semiarid areas. Agricultural irrigation has the first place in groundwater usages in areas that have a limited surface water source. Therefore, modeling and detecting land-use changes and their effects on water resources are required for land-use planning and developing. Markov chain analysis has been widely used in modeling land use/cover changes since 1960s (López et al. 2001). This analysis is a process in which a set of changes occurs over time from a state to another (Thapa and Murayama 2010). In fact, Markov chain analysis predicts a future trend of land based on its past and present changes in land use/cover. Markov chains have several applications in modeling land cover and land-use changes in urban and non-urban regions (Wu et al. 2006). This model has been successfully simulated forest and rangeland ecosystems (Fathizad et al. 2015; Vance and Iovanna 2006). In addition, artificial neural networks (ANNs) have been used for modeling hydrology since 1990s (Singh and Woolhiser 2002). Application of ANNs in the watershed can provide a high accuracy in simulation of hydrologic parameters (Cui and Singh 2015; Singh and Woolhiser 2002; Tokar and Markus 2000). Furthermore, the groundwater levels have also been efficiently modeled by using ANNs (Daliakopoulos et al. 2005; Taormina et al. 2012). In this paper, in addition to simulating the groundwater levels and the land-use/cover changes in the study area, it was aimed to find the relationship between these factors in the future.

This research is carried out in four sections in the arid rangeland ecosystem in western Iran. The first section is aimed to study the trend of past land-use changes in 1988, 2001, and 2014. This gives away comprehensive information about past changes in the region. The aim of second section was to predict the changes of land use via Markov chain model to predict the changes of next 13 years up to year 2027. In the third section, groundwater level changes will be detected and predicted using artificial neural network in this period of time. In the final section, the purpose was to find a relationship between land-use changes and groundwater levels.

Materials and methods

Study area

The studied area of Dehloran plain is located in southeast of Ilam province (47°02′16″–47°25′07″E and 32°48′33″–32°18′48″) with an area of 543.5 km2 (Fig. 1). The region has an arid climate based on de Martonne aridity classification (de Martonne 1926). The mean annual precipitation and temperature are 242.4 mm and 25.5 °C, respectively. In Iran, natural resources have changed particularly in the form of deforestation (Fathizad et al. 2015) and conversion from rangeland to agriculture (Faramarzi et al. 2010; Khormali and Nabiollahi 2009). These land-use changes have had multiple impacts on soil characteristics which consequently has altered the ecosystem. In our study area, the conversion of rangeland to agriculture raises the expansion of the region desertification trend (unpublished reports of Ilam Nature Resources Office) which was the main reason behind choosing this area for the study.

Fig. 1
figure 1

Geographical location of the studied region in Iran and Ilam province

Full size image

Satellite images

In this research, satellite imagery of Landsat TM and EMT + sensors which were taken in June of 1988, 2001, and 2014 years was utilized. The multi-temporal image used in this research was corrected geometrically for classifying land cover maps, and a number of pixels were selected as training sites. Training samples as a ratio of total image, represent 1–5% of pixels (Richards and Richards 1999). Generally, 5–10% of total pixels of image were selected as training samples (Fathizad et al. 2015). In the present study, the number of classes was selected according to available maps and images, conditions of studied region and required classes for mapping land cover; Therefore, four classes were determined, which includes agriculture, rangeland, artificial forest, and residential and barren lands. For this, we applied four supervised methods of parallelepiped (PP), minimum distance from mean (MD), neural network (NN), and maximum likelihood (ML) to classify the satellite imagery (Lillesand et al. 2014). Then, the method with higher-order accuracy was used for the detection analysis. This method can obtain better results in image analysis than other methods of classification, e.g., classification of least distance from mean using mean and covariance of training data (John and Xiuping 2006). The maximum likelihood method was recognized as the best method of classification with about 86% accuracy among the used methods in this study. The precision of image classification was evaluated by random systematic sampling; for this, the ground truth map was prepared by using field data collection (Fig. 2). Moreover, flowchart of the methodology used in this study is illustrated in Fig. 3.

Fig. 2
figure 2

Ground truth map

Full size image
Fig. 3
figure 3

Flowchart of the methodology for the models of both land use and groundwater

Full size image

Markov chain analysis

Markov chain model is the best model to analyze land-use changes in urban and non-urban regions (Wu et al. 2006). These model outputs include likelihoods of status conversion, matrix of converted areas of every class and finally conditional probability images for conversion of different land uses (Gilks 2005; Weng 2002). In this research, cellular automata Markov was used to conduct prediction of land use/cover changes. For this purpose, two raster maps were used as inputs to the model for further analyses. CA–Markov Model is a combination of automatic cells, Markov chain, and multi-purpose allocation of lands, which has been applied to predict future changes of land use/cover. First, the likelihood calculation of class changes of the map was conducted in the format of occurred area changes between t0 and t1 with the implementation of the Markov chain model. The output of Markov model is non-spatial in terms of nature used resolution of automatic cells along with this model to predict spatial situation of uses in t1 + 1 time (Eastman 2009). The aim of implementing a filter is to determine the significance of cells of suitability maps based on their distance from the present use. In this research, filters were sufficiently tested, and finally the 3 × 3 proximity filter was used. Repeat number of automatic cell model execution for prediction of land use in every year equals the number of periods between onset and the end time of predication. For example, year 2001 as the base year included 13 repeats for prediction of land use until 2014. Therefore, a number of 13 more repeats were required for the prediction from 2014 to 2027. In every repeat, the significance of pixels of suitability maps was determined by the implementation of proximity filter according to the present land use, and pixels then, obtain new values. Land-use maps of 1988 and 2001 were used to simulate land use in 2014. Then, simulated land-use map of 2014 was compared with the real map from satellite imagery classification of the same year to validate the model of which the Kappa coefficient was achieved to be about 82%. Afterward, the land-use map of 2027 was predicted by using CA–Markov model in 2027.

Matrix of status conversion

The transition matrix of status conversion of land-use classes between two time periods was calculated by using the obtained land-use maps for each period. This chain requires two land use/cover maps to determine the probability of changes. These kinds of data are usually achieved using the processed satellite images (Tang et al. 2008). The first map is introduced as the “t” time map, and the second map is introduced as “t + 1” time map to the Markov chain model. Then, two values that represent the time intervals between the two land use/cover maps as well as the time interval between the second map and the map that simulates changes in that year is introduce into the Markov chain model. Accordingly, 1988 and 2001 cover maps of the first status conversion matrix were used to simulate year 2014, and 2001 and 2014 cover maps of the second status conversion matrix were used to predict the year 2027s. The value of land-use changes in future was calculated via Markov chain and using land-use maps of onset and the end period in the form of a change matrix transfer probability. It was then applied as a base for planning the future period.

Predicting groundwater levels

One of the important consequences of indiscriminate use of groundwater tables is the agglomeration of layers and soil settlement (Galloway 1984). Falling groundwater level causes the decrease of hydrostatic pressure of table and solid part or table construction loses its stability. In order to study the relationship between the increase of agriculture fields and artificial forests with the decrease of groundwater levels, the groundwater levels in 2027 were predicted. For this, the artificial neural networks (ANNs) with multi-layer perceptron (MLP) and back propagation network (BPN) algorithm were used. The ANNs have been broadly used for hydrologic process (Behzad et al. 2009; Cui and Singh 2015; Rogers et al. 1995) which has a better accuracy than some other methods such as regression as a detection model of nonlinear trends (Cross et al. 1995). In this research, groundwater levels of 13 wells were analyzed to predict groundwater levels in future. These data were obtained from Ilam Regional Water Authority. The inputted data to the model included precipitation, evaporation, temperature, and water levels of discharge/recharge months. In the current study, the cross-validation model (Lek and Guégan 1999) was used to verify the model. For this purpose, the whole data were divided into two data series in which 70 and 30% of data sets were used for training and testing of the model, respectively. Both the training and testing data sets are required to apply in the BPN algorithm which their input/output was considered by the real data (Lek and Guégan 1999). The predictive capability of the model is evaluated by the statistic values of root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) those are found the acceptable amounts for forecasting the groundwater levels (Table 1). The values of RMSE and MAE would equal zero in the ideal condition which indicates a perfect fit between observed and predicted data (Daliakopoulos et al. 2005; Mohanty et al. 2013). In this condition, the R2 statistic estimates an amount of one (Mohanty et al. 2013). Yoon et al. (2011) argued that the ANN model had a better ability than other models for calibration, because of a low RMSE value.

Table 1 Statistical results of precision of ANNs for predicting the groundwater levels
Full size table

Furthermore, falling groundwater of different land-use types was analyzed using ANOVA (α < 0.05). Shapiro–Wilk test was applied to assess normality of the data. As a post-hot test, Tukey’s HSD (α < 0.05) was used to compare means among the land-use types.

In addition, falling groundwater data for each year were extracted from groundwater level data, which is the amount of removed water from the aquifer in the period of irrigation season. In the present study area, falling groundwater was calculated according to the differences between groundwater levels from April to October. Next, the map of predicted falling groundwater levels was interpolated by using the inverse distance weighting (IDW) method in the geographic information systems (GIS) technology for the year 2027.

Results and discussion

Image classification

After geometric correction of multi-temporal images, land cover maps were classified into four classes of agriculture, grassland, artificial forest, and residential and barren lands. The most suitable method of classification was chosen to be maximum likelihood with the Kappa coefficient of 86%, which had the highest similarity to the ground truth data. The area of each land cover/use in 1988 (Fig. 4), 2001 (Fig. 5), and 2014 (Fig. 6) was determined based on the maximum likelihood classification method.

Fig. 4
figure 4

Land use/cover map of 1988

Full size image
Fig. 5
figure 5

Land use/cover map of 2001

Full size image
Fig. 6
figure 6

Land use/cover map in 2014

Full size image

Table 2 shows the change detection of the areas of each land use/cover class during 1988–2014. During this time, the areas of agriculture and artificial forests have had an increasing trend of 39.7 km2 (7.32%) and 55.5 km2 (10.22%), respectively. On the other hand, the rangeland area and residential and barren lands were decreased in area up to 70.8 km2 (13%) and − 24.4 (4.50), respectively. Therefore, it can be argued that rangeland and barren lands have been replaced by agriculture and artificial forests (Fig. 7).

Table 2 Area of different land use/cover classes for 1988, 2001, 2014 (numbers inside the parentheses show percentage)
Full size table
Fig. 7
figure 7

Simulation map of land use/cover changes in 2014

Full size image

Prediction of land-use changes

First, the model calibration was executed to predict land use/cover changes. Area-related statistics and the extent of occurred changes between the real and the predicted areas are presented in the Table 3. Next, the probability matrix of land cover transfer for the year 2014 was prepared. Then, the suitability map of land use/cover was introduced to cellular automata models of Markov to prepare the changes for 2014. The Kappa coefficient for model assessment was calculated to be an amount of 82%. Table 3 shows the value of the real and predicted area of land use/cover in 2014. Accordingly, the values of the predicted area are close to the real areas.

Table 3 Area of predicted and real land use/cover classes in km2 (numbers inside the parentheses show percentage)
Full size table

After finding an acceptable accuracy of 82% for modeling land use/cover changes in 2014, the model was applied to predict the extent of changes in 2027 (Table 4; Fig. 8).

Table 4 Area of predicted classes in 2027
Full size table
Fig. 8
figure 8

Map of predicted land use/cover changes in 2027

Full size image

Classification maps were used to understand the manner of changes in the region and show which classes have increased/decreased. The first status conversion matrix (Table 5) was calculated using the cover maps of 1988 and 2001, and then the second status of the conversion matrix (Table 6) was estimated using the cover maps of 2001 and 2014. These matrices include information about percentage of every class converting to other classes. Given numbers in the diameter of Tables 5 and 6 indicate the values of no change of every use/cover class. Other elements show the extent of class changes.

Table 5 Transition matrix during 1988–2001 achieved by Markov chain model (km2)
Full size table
Table 6 Transition matrix during 2014–2027 achieved by Markov chain model (km2)
Full size table

Prediction of groundwater level falling

According to this fact that farmers and residents of the region use the groundwater for irrigating the agricultural fields and artificial forests (reports of Natural Resource’s office, 2014) and most of the agricultural fields being irrigated, an increasing trend is expected in the exploitation of region wells. Figure 9 shows that the groundwater levels decreased about 6.1 m during 2000–2014. For further analysis, the artificial neural networks (ANNs) model was used to simulate the groundwater levels. The simulation of groundwater levels was conducted after obtaining an acceptable statistical precision, which are those of RMSE = 0.25, MAE = 0.20, and R2 = 0.90 (Table 1). The results of prediction analysis showed that the groundwater level will continue decreasing about 16.8 m during 2014–2027 (Fig. 9). This leads to be severe falling groundwater in the most parts of the study area (Fig. 10). Whenever the agricultural fields and the artificial forests are increased until 2027 (Fig. 8), falling groundwater will highly occur in most parts of the study area (Fig. 10). The simulated map showed that the high intensity of falling groundwater (levels of 0.62–0.72 m and 0.72–0.82 m) is located in the central, south, and southwest of the study area (Fig. 9) where agricultural and artificial forest areas are concentrated (Fig. 8).

Fig. 9
figure 9

Groundwater levels for two periods of 2000–2014 (detection) and 2014–2027 (prediction). The data used for the detection analysis were obtained from Ilam Regional Water Authority. The data collected (groundwater level of the wells) were averaged for each year in the study area

Full size image
Fig. 10
figure 10

Prediction of falling groundwater (m) in 2027

Full size image

Land use in relation to groundwater level

Table 7 summarizes the falling groundwater in different land-use types. Artificial forest had the highest falling number which was not significantly different with the agriculture. By contrast, the lowest falling value was related to the rangeland area which was significantly different among all land-use types, except residential and barren land. In fact, residential and barren land had no significant differences among other land uses.

Table 7 Mean and standard deviation of the falling groundwater (m) in different land use types during 2000–2027
Full size table

Assessing, monitoring, and modeling of land use is a great essential to understanding the environmental and management impacts of the current status, temporal and spatial changes of land uses. Prediction of land-use changes can be simulated by the Markov chain model as a predictive tool with an acceptable accuracy (Arsanjani et al. 2013; Fathizad et al. 2015; Zhang et al. 2011). The hydrological impacts of land-use changes are generally recognized as increasing surface runoff and decreasing groundwater level (Harbor 1994; Foley et al. 2005; Lin et al. 2007).

The purpose of the present research was to study the trend of land-use changes in three periods of 1988, 2001, and 2014 and to predict changes by using Markov chain. Afterward, the changes and simulation of the groundwater levels were considered in relation to land-use changes. The results of the land change detection indicate that the area of rangeland has decreased, while agriculture and artificial forest have followed an increasing trend during periods of 1988–2014 (Table 2). Conversion of rangeland to cropland has been recognized as the main reason of land degradation in semiarid rangelands of western Iran (Faramarzi et al. 2010). Rangeland area has decreased from 294.1 km2 in 1988 to 223.4 km2 in 2014 (Table 2). During this period, artificial forest has increased about 55.6 km2 (10.2%). This situation seems to have been developing in response to deforestation either in this area in the past or in an area under similar conditions; Since, Fathizad et al. (2015) have indicated that deforestation was the greatest destruction in the semiarid area in west of Iran. Converting forest both to agricultural lands (Mugagga et al. 2012) and to rangelands (Fathizad et al. 2015) were demonstrated as the main land-use changes in different areas. In the study area, agricultural area has increased from 119.7 km2 in 1988 to 159.5 km2 in 2014 (Table 2). At the same time, residential and barren land has decreased about 4.5% (Table 2) which seem to belong to decrease of barren lands while residential buildings are increasing in urban regions. In constant, Fathizad et al. (2015) found that both barren and residential lands had an increasing trend that was arguably due to severe deforestation in their study area. In our field, the increase of artificial forest and agriculture could be tended to the decline of barren lands. Increasing artificial area indicates more human activities for tourist presence where the pasture area was reduced (Falcucci et al. 2007).

Markov chain model provided ratio determination of different land-use changes and their prediction of future (Muller and Middleton 1994). In the present research, cellular automata model of Markov chain was used to predict future trends of land-use development in the study region. The transition matrix results indicated that artificial forest and rangeland areas had the highest and the lowest stability, respectively (Tables 5, 6). The prediction analysis showed that agricultural fields would reach 188.9 km2 (34.7%) with the rise of 5.3% in 2027. Increasing agricultural fields will be shown as a slow trend which could be due to the decreasing groundwater levels of the region wells. Extraction of groundwater in many regions has resulted in reducing groundwater levels (Foley et al. 2005; Lin et al. 2007), and this indenting discharge/recharge of an area can be applied as a useful method for preparing a successful land-use plan (Batelaan et al. 2003). In addition, the prediction data indicate that artificial forest will only increase about 0.80% in 2027. The statistical analysis indicated that falling groundwater in artificial forest and agriculture was significantly higher than rangeland area (Table 7). The main reason of this condition is that the rise of artificial forests requires irrigation and using the water of region wells that we will lead to a severe shortage of water due to indiscriminate exploitation in the coming years. This leads to less inclination of region residents to artificial forests because these forests require more water for irrigation. Stolte et al. (1997) proposed that the tree plantation decreased a piezometric level which is not a solution for the discharge areas. On the other hand, the prediction data showed that the rangeland area will decrease about 5.81% from 2014 to 2021 in the study area. Replacement of rangeland with cropland alters many ecosystem parameters, which is considered to be one of the main problems of a semiarid area (Scanlon et al. 2005). Water resources are significantly influenced by land-use changes (Bhaduri et al. 2000). It is predicted that the groundwater level will fall about 16.8 m from 2014 to 2017. These changes could be due to the replacement of rangeland to agricultural land and artificial forest where needs more water for irrigation. The predicted results showed that the most severe case of falling groundwater (Fig. 10) is centered on the cropland and artificial forest areas (Table 7; Fig. 8). Increasing irrigated croplands are resulting in the declining quality and quantity of groundwater (Foster et al. 2008; Saiko and Zonn 2000). Under this condition, conducting some developmental techniques seems to be required for improving groundwater conditions in the study area. For example, artificial recharge was suggested as an effective way to enhance groundwater levels (Liu et al. 2013).

Conclusion

Analysis of land-use changes showed that the agriculture fields and artificial forest have increased in the study area. The artificial forest is known as human presence for the tourist activities and also as a windbreaker against wind and water erosions, which are the natural proceedings in the arid and semiarid regions. Moreover, the inappropriate developments of agricultural fields are leading to the falling groundwater levels. Rangeland was the main natural resource which was converted to cropland and artificial forest. Furthermore, Markov chain model was applied to predict the changes in the future after obtaining an accuracy of about 82%. Rangeland and artificial forest had the lowest and highest stability, respectively, in our study area. It was predicted that transition of rangeland to agriculture was of the most important changes in the future. On the other hand, we predicted that the groundwater level will significantly be decreased in 2027, which could be related to the increasing area of cropland and artificial forest. Therefore, according to the predicted results, the local people, managers, and decision makers should prevent the conversion of rangeland to other land uses particularly cropland. These actions not only generate positive changes in natural resources of the region, but will also prevent the excessive drop of water levels in region wells via observance of fundamental law for the expansion of agricultural land.