Introduction

Remote sensing provides a broad view of landscapes and consistent record through time, making it an important tool for monitoring and managing land resources (Phinn et al. 2002). Change detection is the process of identifying differences in the state of an object or phenomenon by observing it at different times (Coppin et al. 2004). Accurate change detection of land use/land cover (LULC) has become a key issue for monitoring local, regional, and global environments and resources, providing the foundation for a better understanding of relationships and interactions between humans and natural phenomena in order to improve management and use of resources (Lu et al. 2004). Remote sensing has been used as a powerful tool in change detection applications and provides cost-effective multi-temporal satellite images. Since the early 1970s, these data, along with remote sensing analytical approaches, have been of considerable interest for periodic monitoring of large LULC in natural and built environments (Phinn et al. 2002; Griffiths et al. 2010; Jayanth et al. 2016). Considering the importance of remote sensing in evaluating changes in landscape cover, this technique was selected for the present study. A popular and commonly used approach for image analysis is digital image classification. The purpose of image classification is to label the pixels in the image with meaningful information from the real world. Through classification of digital remote sensing images, thematic maps bearing information such as land-cover type and vegetation type can be obtained. Subsequently, change detection may be carried out based on these thematic maps (Dewan and Yamaguchi 2009; El-Asmar et al. 2013; Rawat and Kumar 2015; Ghebrezgabher et al. 2016). Traditionally, pixel-based techniques usually were used to create LULC classifications (Erbek et al. 2004; Weih and Riggan 2010; Taati et al. 2015). The pixel-based techniques include either a supervised or unsupervised classification or a combination of these (Enderle and Weih 2005; Weih and Riggan 2010). These methods only analyze the pixels’ spectral signatures without any consideration to the spatial or contextual information of each pixel. eCognition® software package has been developed utilizing object-based classification. The package uses a segmentation process and learning algorithm to analyze both the spectral and spatial/contextual properties of pixels procedures (Weih and Riggan 2010). This process provides a semi-automatic classification that could be more accurate than traditional pixel-based methods (Weih and Riggan 2010; Whiteside et al. 2011; Jawak et al. 2015; Butt et al. 2015; Sinha et al. 2015). There is increasing attention on object-based classification (Du et al. 2013; Long et al. 2013; Im et al. 2013; Aslami et al. 2015; Hussain and Shan 2015), which has already been reviewed by Blaschke (2010). Moreover, an ArcGIS Python model was developed by Wells (2010) utilizing polygon extraction from aerial images and producing land-cover map. As for the advantages of object-based classification comparing with pixel-based methods, we can briefly mention (1) higher potential for processing high-volume data in lower time since it uses segments (a group of similar pixels) not individual pixels; (2) utilizing segments or objects representing characteristic texture features which are ignored in traditional methods; (3) doing the classification in two steps (segmentation and classification) providing the chance of obtaining the higher accuracy due to applying all spectral, textural and contextual information in the same time for classification (Gao and Mas 2008), and (4) eCognition package provides a more user-friendly environment for experts in all steps of classification and post-classification. This technique is considered in this study due to its success in several research using Landsat images, such as Lewinski (2006), Al Fugara et al. (2009), Roostaie et al. (2012), Amalisana et al. (2017) and Phiri and Mogenroth (2017).

Myint et al. (2011) used a QuickBird image over a central region in the city of Phoenix for object-based classification. They selected 0.1 and 0.5 for the shape and compactness factors, respectively, and selected 10 different objects per class at scales of 10, 25, 50, and 100 to perform a discriminant analysis for segmentation, finally settling on a scale of 100 for their study. They then classified the image using a nearest neighbor classifier with an accuracy of 90.40%. There are some other studies using Landsat series, the appropriate accuracy assessment criteria (Campbell et al. 2015), and several segmentation parameters which obtained higher overall accuracy, including Li et al. (2015), Olmanson and Bauer (2016), and Cheţan et al. (2017).

Over the study area, the traditional land management and land-use planning with rapid population growth particularly during recent decades resulted in unrestrained settlement/industrial/ recreational areas in addition to agricultural land expansion which pose crucial challenges for environment managers in various aspects especially in Ardabil County with a traditional economy. Attentive monitoring at fine temporal/spatial scales would reveal a fluctuating landscape, punctuated by changes in movement of people, perturbation from environmental disasters, and shifts in activities. On the other hand, the lack of financial support to scientific research have made researchers in Iran to find and evaluate the free, relatively high spatial/temporal remotely sensed data for land and environment monitoring (Akbari et al. 2007; Solaimani et al. 2010; Amini Parsa and Salehi 2016; Halimi et al. 2017; Mirzaei et al. 2018). Thus, the applied, using free satellite images and new cost-effective processing techniques and resulted information, can enable regional/local planners as well as policy makers to identify and manage LULC changes better. Moreover, the methodology can be evaluated by other researchers in other regions for similar aims.

Already, we have compared Artificial Neural Network (ANN), Support Vector Machine (SVM), and Maximum Likelihood (ML) methods with object-based classification technique for the preparation of LULC map using Landsat-8 images in Ardabil, Namin, and Nir counties, in the Ardabil province in the northwest of Iran, and concluded that the object-based technique provided with more reliable results in comparison (Aslami et al. 2015), following in this study we have conducted to detect LULC changes in the study area using object-based image analysis and Landsat data.

Materials and methods

Study area

The study area is located at 47° 46′ to 48° 42′ E and 37°48′ to 38° 38′ N (Fig. 1), covering an area of 4726 km2. The Ardabil plain is situated at the center of the study area, with agriculture as its primary land use (mainly potato and wheat cultivation) in addition to residential and industrial areas. Mount Sabalan is the third highest summit in Iran with a height of 4811 m above sea level and is situated in the west of the study area, which is dominated by rangeland (shrubland, grass-shrublands, grasslands, and approximately 32 wetlands); it serves the Shahsavan nomads as their summer rangeland (Sharifi et al. 2013; Ghorbani et al. 2018). Highlands at the south and northeast are also primarily rangeland (mainly shrub-grasslands). Areas of low altitude in the east are covered by scattered forest, which is dominated by Euro-Siberian species such as Quercus castaneifolia, Quercus macranthera, Fagus orientalis, Corylus avellana, and Acer campestre (Teimoorzadeh et al. 2015). Since 1993, when new country political divisions have made the area and some other surrounding counties as a new province (Ardabil province, which has separated from East Azerbaijan province), population concentration has been increased in the study area as the capital city of Ardabil province; consequently, new industrial, commercial, and recreational sets have been developed. Moreover, due to the new policy for tourists’ attractions in the province, more than 6 million tourists dominantly not respecting the natural environment annually visit the area. Their need for more cross-road service complexes and ski resorts on one hand and development of Sarien city due to its different spa facilities on the other hand have made the situation even worse because of building lots of restaurants and shopping centers in different natural areas. All these resulted in expanding border cities (Ardabil, Namin, Nir, Sarien, Hir, Abibagloo, and Samian) and their population needing more land for agriculture, cultivating rangeland mainly for planting irrigated/rainfed wheat, barley, and potatoes, and building farmland (mainly irrigated farms) (Ghorbani et al. 2011). The elevation of the study area ranges from 709 to 4811 m and experiences harsh winters and moderately cool summers with average annual temperatures ranging from − 3 to 13 °C, and annual precipitation from 215 to 766 mm.

Fig. 1
figure 1

Digital elevation model (DEM) of the study area (Ardabil, Namin, and Nir counties) in Ardabil province, Iran, and the world, with topographic variation

Full size image

The overall framework of the study is presented in Fig. 2.

Fig. 2
figure 2

The framework of the employed principal procedure

Full size image

Image selection

The seasonality and phenological patterns (Reed et al. 2003) of the study area were considered, according to observations at an altitude of 4102 m, and no considerable seasonal variation was found. However, phenological stages were found to be different, with four discernible seasons, in addition to different temperatures, and types and amounts of precipitation at different elevations. It was thus decided that the best time of the year for images was late April to early May. In the case that moisture would affect the data acquired from these images, dates 15 days earlier were also considered; however, there was no noticeable precipitation during this period. Finally, images including Landsat Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), and Operational Land Imager (OLI) were selected (Table 1).

Table 1 Characteristics of the selected images
Full size table

Image preprocessing

Preprocessing is critical in change detection studies because it is assumed that the spectral properties of non-changed areas remain stable, and inadequate preprocessing can increase the potential for error by causing detection of false change in the spectral space (Lu et al. 2004; Wulder et al. 2006).

Initially, a digital elevation model (DEM) map of the study area was produced (the DEM was extracted using 1:25000 topographic maps of National Cartographic Center of Iran, with 10 m horizontal and vertical accuracy) using ArcGIS10 (Fig. 1). The image preprocessing stages, including geometric correction, radiometric controlling, topographic normalization, and image enhancement, were conducted before image processing (Chander and Markham 2003; Lillesand et al. 2008). The obtained image was registered to the UTM projection with the WGS84 datum. However, according to the collected 16 ground control points (GCPs) using Garmin Oregon 550 GPS and some geographical information systems (GIS) layers such as registered topographic map, the acquired images still required rectification. The affine transformation model which is widely used for geometric corrections was used to geometrically correct the images to align accurately with the collected GPS points. Therefore, images were geometrically corrected (UTM/WGS84), and then controlled using 16 known GCPs (crossroads). According to the concept of change detection, radiometric correction is a necessary step in image preprocessing. For radiometric correction, the following equations, taken from the official Landsat site, were used for the three selected images. Equation 1 was used for the TM image and Eq. 2 for the ETM+ image to convert the digital number (DN) to radiance:

$$ {L}_{\uplambda}= Gain\times QCAL+ Bias $$
(1)
$$ {L}_{\uplambda}=\left(\left({Lmax}_{\uplambda}-{Lmin}_{\uplambda}\right)/\left( QCAL max- QCALmin\right)\right)\times \left( QCAL- QCALmin\right)+{Lmin}_{\uplambda} $$
(2)

where Gain is rescaled gain (m2*ster*μm), Bi`as is rescaled bias (m2*ster*μm), QCAL is the quantified calibrated pixel value (DN), Lλ is the spectral radiance of the sensor’s aperture in watts/(m2*ster* μm), Lminλ is the spectral radiance scaled to QCALmin in watts/(m2*ster* μm), Lmaxλ is the spectral radiance scaled to QCALmax in watts/(m2*ster* μm), QCALmin is the minimum quantized pixel value in DN, and QCALmax is the maximum quantized pixel value in DN. Equation 3 was used to convert radiance to Top of Atmosphere (TOA) reflectance for TM and ETM+ sensors:

$$ {\rho}_{\mathrm{p}}=\frac{\uppi .{\mathrm{L}}_{\uplambda}.{\mathrm{d}}^2}{\mathrm{E}{\mathrm{SUN}}_{\uplambda}.\cos {\uptheta}_{\mathrm{s}}} $$
(3)

where ρp is unit less TOA or planetary reflectance, d is earth-sun distance in astronomical units, ESUNλ is the mean solar exoatmospheric spectral irradiance, and θs is the solar zenith angle in degrees. OLI images were corrected using Eq. 4, which directly converts DN to TOA reflectance:

$$ \uplambda \uprho =\frac{M_{\rho }{Q}_{cal}+{A}_{\rho }}{\sin \left({\theta}_{SE}\right)} $$
(4)

where Mρ is multiplicative rescaling factor, Aρ is additive rescaling factor, which both are band-specific; and θSE is local sun elevation angle.

Topographic normalization was performed because of extreme height differences in the study area. For this purpose, the Minnaert approach in Eq. 5 was used (ERDAS Imagine 9.2. Field Guide™, 2008):

$$ {BV}_{normal\lambda}=\frac{BV_{observed\lambda}\cos e}{k\left(\cos\ i\right)\ \left(\cos\ e\right)} $$
(5)

where BVnormalλis the normalized brightness value, BVobservedλis the observed brightness value, i is the incident angle, e is the slope angle, and k is the empirically derived Minnaert constant.

Object-based classification

Object-based methods use information from texture, shape, and tone in addition to numerical values (Bontemps et al. 2008; Dronova et al. 2011). The most evident difference between pixel-based and object-based image analysis is that, firstly, in object-based analysis, the basic processing units are image objects or segments, not single pixels. Secondly, the classifiers in object-based image analysis are soft classifiers that are based on fuzzy logic. Soft classifiers use membership to express an object’s assignment to a class. The membership value usually lies between 0.0 and 0.1, where 0.0 indicates absolute improbability assignment to a class and 1.0 indicates complete assignment to a class. The degree of membership depends on the degree to which the objects fulfill the class-describing condition. One advantage of such soft classifiers lies in the possibility of expressing uncertainties regarding the description of the classes (Bontemps et al. 2008; Dronova et al. 2011). Fuzzy classification is usually used in eCognition software, where it presents a type of nearest neighbor classification.

A segment is a group of adjacent pixels in an area in which likeness, such as number value or texture, is the most common criterion between them. Image segmentation is very important in supplying the basic building blocks for object-based image analysis and the accuracy with which this phase is carried out influences the quality of the object-based classification. Segmentation can take into account factors such as shape, texture parameters, and the compactness coefficient (Zhang et al. 2005). Compactness parameter determines the segments density and the shape parameter is used for determination of softness/sharpness of the segments. Appropriate scale is another important factor for segmentation (Liu and Xia 2010). Different segmentations on an OLI image of part of the study area are presented in Fig. 3. Scale parameter determines size of the segments as shown in Fig. 3. The parameter values are selected based on “try and error” experimentally. This means we used different values for each and compare results until the best result is obtained. The object-based classification of images from 1987, 2002, and 2013 in this study makes use of different values of scale, compactness, and shape parameters for each image, in addition to the nearest neighbor algorithm. A comparison of segmentation and classification gave resulting values for the images equal to 20, 0.1, and 0.6 for 1987 image; 30, 0.1, and 0.5 for 2002 image; and 40, 0.6, and 0.5 for 2013 image, respectively for scale, compactness, and shape parameters.

Fig. 3
figure 3

Results of three segmentations using different scale values for part of the study area. a 30. b 60. c 130

Full size image

Accuracy assessment

A validation data set was used to assess the accuracy of each classification. There were 500 validation points taken from Google Earth for the map from 2013, 235 points for the map from 2002, and 170 points for the map from 1987 (100 points of those were initially collected by a Garmin Oregon 550 GPS, and then controlled using Google Earth imagery, because of the road accessibility limitation rest of the 400 points were collected only by Google Earth images). To overcome low-resolution problem of Landsat images by choosing ground control points (GCPs) based on some criteria (according to Campbell et al., 2015): (1) the sample unit must be ≥ 90 m × 90 m in size (3 × 3 Landsat pixels) (most units were significantly larger and then the collection was done at or near the center), (2) the entire area within the unit must be visually (and spectrally) homogeneous, and (3) the areas must have the maximum heterogeneous (variability) between units. Multivariate techniques were used to perform the accuracy assessment, including an error matrix, overall accuracy, and a Kappa coefficient of agreement (Congalton and Green 2008) by class and overall classification. The error matrix was also used to calculate the producer’s accuracy, the user’s accuracy, and overall accuracy. The overall accuracy is defined as the proportion of pixels correctly classified divided by the total classified pixels, as shown in Eq. 6:

$$ \mathrm{OA}=\frac{\sum_{\mathrm{k}=1}^{\mathrm{N}}{\mathrm{a}}_{\mathrm{k}\mathrm{k}}}{\sum_{\mathrm{i},\mathrm{k}=1}^{\mathrm{N}}{\mathrm{a}}_{\mathrm{i}\mathrm{k}}}=\frac{1}{\mathrm{n}}\sum \limits_{\mathrm{K}=1}^{\mathrm{N}}{\mathrm{a}}_{\mathrm{k}\mathrm{k}} $$
(6)

where N is the total number of pixels classified and \( \sum \limits_{\mathrm{k}=1}^{\mathrm{N}}{\mathrm{a}}_{\mathrm{k}\mathrm{k}} \) is the sum of the diagonal pixels. The producer’s accuracy relates to the probability that a reference sample was correctly mapped and measures the omission error (1—producer’s accuracy). In contrast, the user’s accuracy indicates the probability that a sample from the map actually matches the reference data and measures the commission error (1—user’s accuracy). As noted by Foody (2004), the Kappa coefficient (formally estimated by\( \widehat{K} \)) is based on the comparison of predicted and actual class labels for each case in the validation data set, and is calculated using Eq. 7:

$$ \widehat{K}=\frac{p_0-{p}_c}{1-{p}_c} $$
(7)

where p0 is the proportion of cases in agreement and pc is the proportion of agreement expected by chance.

LULC change detection over the selected time periods was analyzed using a post-classification comparison (PCC) technique with IDRISI Selva software and all information concerning LULC change was summarized.

Results and discussion

The land-use maps from 1987, 2002, and 2013 derived from Landsat TM, ETM+, and OLI images can be seen in Fig. 4 and their overall accuracy and Kappa statistics in Table 2. The result of accuracy assessment shows high ability of object-based classification in producing LULC map.

Fig. 4
figure 4

Land use maps of a 1987, b 2002, and c 2013 for the study area produced using the object-based method

Full size image
Table 2 Summary of accuracy assessment
Full size table

Rangeland and dry farming are the predominant land uses in the study area for all 3 years, followed by irrigated farming. Forest cover appears relatively insignificant. As shown in Table 3, there has been considerable change in the study area from 1987 to 2013. Dry farming and urban areas have increased, while rangeland, forest, and meadow lands have decreased significantly. Irrigated farming area has decreased at first but has increased after 2002.

Table 3 Areas of land-use types
Full size table

Measurements and locations of these changes are presented in Tables 4, 5, and 6, and Figs. 5, 6, and 7, respectively. Each label in Table 4 corresponds to a class in Fig. 6. The numbers in the legends of the figures correspond to the LULC classes, listed under the ID label in Table 3.

Table 4 Results of land use comparison between 1987 and 2002 (ha)
Full size table
Table 5 Results of land use comparison between 2002 and 2013 (ha)
Full size table
Table 6 Results of land use comparison between 1987 and 2013 (ha)
Full size table
Fig. 5
figure 5

LULC change from 1987 to 2002. (The numbers in the legends correspond to the LULC classes, listed under the ID label in Table 3)

Full size image
Fig. 6
figure 6

LULC change from 2002 to 2013. (The numbers in the legends correspond to the LULC classes, listed under the ID label in Table 3)

Full size image
Fig. 7
figure 7

LULC change from 1987 to 2013. (The numbers in the legends correspond to the LULC classes, listed under the ID label in Table 3)

Full size image

In addition, 97,156.6 ha of rangelands were converted to dry farming lands, constituting the most significant change in the study area, as shown in Fig. 8.

Fig. 8
figure 8

Land converted to dry farming from 1987 to 2013

Full size image

Medium-resolution satellite images such as Landsat images have been used for monitoring and mapping LULC from local to regional scales. Change detection and LULC mapping have increasingly been recognized as some of the most effective tools for urban and environmental resource management. Our results demonstrate the reliability and accuracy of using object-based remote sensing in LULC mapping and change detection in a heterogeneous region. The results of Aggarwal et al. (2016) and Petropoulos et al. (2012) have also demonstrated the high capacity of this method. Considering the appropriate accuracy assessment criteria and classification parameters using Landsat imagery and object-based classification, Cheţan et al. (2017) reported a range of 80 to 93% for overall accuracy using object-based classification and different variables. Moreover, the overall accuracy obtained by Olmanson and Bauer (2016) study ranges from 92.2 to 96.4% which even was improved using Lidar data (was peaked at 98%). The recent examples which are supporting our results show the high potential of object-based technique in similar classification and fields with the lowest cost. Moreover, comparison of the results of previous studies, such as Azizi et al. (2016) and Mirzaei et al. (2018), which used pixel-based image analysis for LULC mapping and change detection in the Ardabil province, with the results of the present study, allows us to conclude that object-based image analysis has more advantages in this regard.

By considering LULC change from 1987 to 2013, the high amount of change in different land uses such as rangelands to dry farming lands is highlighted, indicating population increase, mismanagement, and inappropriate land policy in this region. Shalaby and Tateishi (2007), and Dewan and Yamaguchi (2009) have also reported the large changes in LULC associated with human development. This issue requires more attention from officials and rangeland managers for suitable land management. For example, assigning rangeland to agriculture goes contrary to the principles of sustainable development and can cause an increase in soil erosion and pest outbreaks, in turn causing soil and water pollution through the use of chemical fertilizers and pesticides. The effects of such land-use allocations are the focus of numerous studies, such as Hale et al. (2008). It is thus evident that past land policies and administration of these areas require in-depth scrutiny and review.

Conclusion

In this study, medium resolution satellite Landsat images were classified using the object-based method, and LULC changes were detected through PCC. Validation of the classified maps indicates the high capacity of the object-based method. Change in different land uses, such as rangelands to dry farming lands, is considerable, showing links to human development, mismanagement, and inappropriate land policy in this area. In the north of the study area in 2013, rangelands have decreased considerably, having been converted to dry farming land. This issue requires more attention from officials and rangeland managers for suitable land management. If such trends of rangeland loss continue, effects such as soil erosion, especially on slopes, water and soil pollution, and pest outbreaks are predicted to increase considerably.