عنوان مقاله [English]
Karst zones are spread in different parts of the world and according to Ford and Williams (2007) about 20% of the continents' surface is covered with karst, which accommodates 1.4 billion of the world's population and help provide fresh water for human use. In addition to its benefits to human beings, karst can also be dangerous if it is not accurately identified. Karst hazards include subsidence and damage to urban and rural facilities, water escape from dam lakes, disruption of road construction activities and water tunnels, as well as the high potential for karst water pollution. The issues raised and the significant expansion of karst areas in different parts of the world, as well as in Iran (especially in the Zagros and Alborz), and the government's national plans for the exploitation of these water resources doubled the need for management plans to further protect them (Medadi, 2018; Morsali, 2018). From the point of view of water resources management, since karst landforms play a major role in feeding karst aquifers, and these aquifers supply water to the surrounding urban and rural communities, knowledge of their conditions, how to manage them and the dangers that these water resources facing is essential. One of the karst phenomena that is very important in both water supply and water security is the collapse and formation of sinkholes. Sinkholes play an important role in the absorption, infiltration, feeding and discharge of karst aquifers (e.g. Mohammadi et al., 2018). Despite their importance, few studies have been conducted to map sinkholes distribution. These studies are mostly based on airborne/drone-carried LiDAR technology (e.g Rajabi, 2018; Zhang et al., 2019), which is expensive and in most cases unavailable, especially in developing countries, or on different combiations and analysis of morphometric parameters derived from digital elevation model (DEMs) (Yong Je Kim et al., 2019).
In our work we present a new method for sinkhole detection and mapping by DEMs analysis of the Parav-Biston massif in the Kermanshah province of Iran. The study has been motivated by the paramount importance of this massiv, punctuated by plenty of sinkholes, and its carbonatic rocks in supply fresh water to the whole region (Vahdati et al., 2007; Mohammadi et al., 2018).
In this area, many studies analyzed springs, aquifers and the potential for underground water resources (e.g. Mahmudi and Maleki, 2001; Mohammadi et al., 2018; Zanganetabar and Ghadimi, 2019; Maleki et al., 2020; Khoshraftar et al., 2019; Ghadimi and Zanganetabar, 2019). Karst development in the Parav-Biston massif has been analyzed by spring water discharge with time, analysis of isotopic and chemical data and topography (e.g. Maghsoudi et al., 2010; Entezari and Aghaeipour, 2015; Dastranj et al., 2019). Also, some studies have been devoted to the analysis of relationships among lithology, fault and fractures, precipitation, altitude, temperature, slope, aspect and karst and sinkholes formation (e.g. Jafarbeyglou et al., 2011; Jahanfar et al., 2018).
Despite all these studies, a comphreensive and homogeneous mapping of sinkoholes in the Parav-Biston massif is still missing. Recently, the availability of TanDEM-X, 0.4 arc seconds (approximately 12 meters) resolution, DEMs has opened a new opportunity in the field of studies based on DEMs. Therefore, this paper seeks to evaluate the possibility of detecting sinkholes with the help of TanDEM-X data, and its derived information layers, and using data mining and Geogarphic Object-Based Image Analysis (GEOBIA) methods, and to map the distribution of sinkholes in the Parav-Biston region.
The study area is hosts a limestone massif part of the Zagros Mountains (in the Kermanshah province of western Iran), and is known as Parav-Biston massif. This mountain range is surrounded on all sides by open plains and its area is equal to 1033 square kilometers (Figure 1). Its maximum height is Paravo mountain at 3385 meters a.s.l. Parav-Biston massiv is composed by the carbonate rocks of Bistoon limestone formation, which has a high solubility, and due to its high altitude and heavy rainfall, has provided the basis for the formation and development of karst forms such as sinkholes, ovals and caves.The existence of the famous Paravo cave with a depth of 751 meters and the existence of 26 natural wells in this cave is a sign of the high volume of underground caves and karst development in the region. In the study area, the depth of sinkholes sometimes reaches tens of meters. Karst development and interconnection of sinkholes also generate a more developed and wider form of sinkholes known as Uvala.
Figure 1. Study area
In the present study, different data and methods were used with the aim of shaping a process to identify and map sinkholes. The data for the analysis and validation of the results have been prepared by fieldwork, archive information, GIS and remote sensing. Fieldwork (by camera, simple laser meters and GPS) and remote sensing allowed to collect samples for machine lerning and data mining methods and get familier with the morphometry of some sinkholes (location, maximum and minimum diameter, depth).
The TanDEM-X data have been prepared by the German Space Agency through interferometry. The dataset was created using X-band SAR data from two polar orbit satellites between 2011 and 2014. TanDEM-X data were prepared as a DEM layer in a GIS, from which various other layers have been derived, and the preprocessing and extraction steps of the required information layers were performed (Figure 2). This DEM has spatial resolution of 12 meters and vertical accuracy of 2 meters in flat areas and 4 meters in sloping areas (Krieger et al., 2007).
Figure 2. DEM and derived information layers. a) DEM, b) Hillshade, c) Slope, d) Aspect, e) Flow direction, f) Flw accumulation, g) Profile curvature, h) Plan curvature, i) Curvature.
Then, data segmentation and object detection were performed on the dataset. Next, based on collected sinkhole (200 samples 70/30 percent for training and validation), non-sinkhole samples, using WEKA Machine Learning (Hall et al., 2009; Frank et al., 2016) open source software, the necessary rules for distinguishing sinkholes from the GIS layers and their properties were extracted and created rules accuracy evaluated using K-fold cross-validation. Sinkholes were extracted by applying created decision tree as a supervised machine learning method to the data set. Finally, the accuracy of classification was assessed using 30% of collected samples. A summary of the implemented process is provided in Figure 3.
Figure 3. Methodology flowchart
Results and Discussion
The classification accuracy in the data mining based on K-fold cross-validation were 87.6%, which seems acceptable. According to the results, the variables used in defining the final rules were cumulative flow, slope, flow direction, hillshade, plan curvature, Max-diff index(“the absolute difference between the minimum object mean and the maximum object mean divided by the mean object brightness”(Minaei and Kainz, 2016)), Area pxl, profile curvature and slope direction through data mining in WEKA Macine Learning Software. After applying the data mining process to the samples, the resulting list of object properties and threshold boundaries have be used to identify synkholes and construct a classification tree by applying a set of rules on the information layers (Figures 4). The assessment of the overall accuracy of the classification output (Figure 5) was equal to 71.3%. Evaluation of the spatial distribution of the sinkholes identified in the results shows that most of the sinkholes are located in areas with an altitude of 2400 to 3500. In this regard, Mahmudi and Maleki (2001) and Jafarbeyglou et al. (2010) suggested the continuation of karst processes at altitudes above 2500 meters. Also, in the slopes of less than 10 degrees, the highest rate of development of sinkholes has occurred, which is 13% of the area with these conditions. Entezari et al. (2015) and Jahanfar et al. (2018) also announce the expansion of surface karsts in the region in the slopes of 0 to 5 degrees. In other degrees of slope, a sinkhole is rarely seen. Geographically, sinkholes have been identified in the northern and eastern regions more than the southern and western regions, which is consistent with the findings of Entezari et al. (2015).
Figure 4. Decision Tree provided by data mining in the WEKA Machine Learning Software to detect sinkholes.
Figure 5. sinkhole distribution map in the study area provided by implementing data mining and GEOBIA on the TanDEM-X data
The northern slopes are 13.25% and the eastern slopes are 11.54% of the study area. Also, the distribution of sinkholes is consistent with the ovals detected by Jahanfar et al. (2018). In general, the results show that the use of a combined data mining and GEOBIA approach with TanDEM-X data can provide acceptable results with sinkholes detection accuracy. The results of this study are in line with the results of Jafarbeyglou et al. (2010) who have identified sinkholes in the area. Compared to the pixel-based methods, despite the fact that each has its advantages due to the nature and shape of the sinkholes, the object-oriented method showed to be effective, also preventing “salt and pepper” effect in the results. Utilizing data mining capabilities and machine learning algorithms also greatly simplifies the process of defining rules, which is very difficult by trial and error. Another important point is that data mining helps a lot in choosing threshold limits for object detection. Also, the ability of the object-oriented method to combine visual/optical properties with texture, size and geometry properties in the classification process can be very useful and instructive (Blaschke et al., 2014, Minaei and Kainz, 2016).
The aim of this study was to extract and identify sinkholes using remote sensing data. The advantage of this study is the use of TanDEM-X data and the combination of data mining and GEOBIA methods, which increases the accuracy and better identification of the desired phenomena. Using a combined approach of data mining and GEOBIA along with TanDEM-X data was able to provide acceptable results with a sinkholes detection accuracy of 71.3%. Our results show that the sinkhole detection approach using TanDEM-X data provides an effective way to identify sinkholes in large or inaccessible areas, which may be very difficult and time consuming on land. Classification by using different features for specific classes as one of the advantages of the object-oriented method has made this method an efficient and reliable. This method is suitable for detecting sinkholes and on the other hand, it also makes it possible to identify areas close to aquifers in large karst massifs. Due to the emphasis of related research on the direct relationship of sinkholes with water resources in the study area and the need to protect them, it is suggested that newer methods and data were used to produce more accurate geographic databases. Future research could also examine the combination of sinkholes mapping to water intake vulnerability maps.
Special thanks to: German Aerospace Center (DLR), for providing the DEM: Elevation data is a TanDEM-X Digital Elevation Model, derived from TanDEM-X mission and provided by ©DLR 2017 in the scope of project DEM_OTHER1257.