Studying Effectiveness of Landsat ETM+ Satellite Images Classification Methods in Identification of desert pavements (Case study: South of Semnan)

Extended abstract 1- Introduction The process of identifying landforms is a subject that has been researched by many researchers. All the definitions of geomorphology emphasize the study and identification of landforms. Understanding landforms and how they are distributed are some sort of essential requirements in applied geomorphology and other environmental sciences (Shayan et al., 2012). On the other hand, remote sensing is a powerful tool for studying different ecosystems of the earth to produce valuable temporal and spatial data (Rezaei Moghaddam and Saghafi, 2006). Arekhi (2014) used ETM+ digital data to map the land use of the Abbas plain. To classify images, artificial neural network, supporting vector machine and maximum likelihood were used. Based on the results, neural network classification method has the highest accuracy of land cover mapping. Also, De laet et al. (2007), by studying the development and stability of desert pebbles in Turkey, concluded that desert pebbles in this area are formed likely (in situ) by mechanical erosion of the surface fragments and minimal Tophonomic effects, in sediment with diameters greater than 2 cm. The purpose of this study is to evaluate the efficiency of Landsat imagery in identifying and classifying desert pavement comprehensively using satellite imagery classification strategies. 2-Methodology The studied area with an area of ​​47645/98 ha, in Semnan city is located in 53° 28¢ to 53° 53¢ 43° east and 35° 20¢ to 35° 40¢north. ETM+ multispectral satellite data were selected for this study because of spatial, temporal and especially radiometric resolution. The data of this sensor comprises seven spectral bands, obtained from the USGS site. In order to distinguish different types of desert pebbles in terms of cover density using EDRISI Selva and ENVI 4.5 software and ETM+ sensor images of Landsat satellite, 4 methods including supervised classification methods, maximum likelihood, Mahalani distance, minimum distance from mean and Parallel surfaces were used. Each classification method was compared for classification accuracy using overall accuracy coefficients, kappa, user accuracy and producer accuracy. The error matrix table was also presented for each method. In this matrix, the degree of compatibility of each class with the ground reality is shown, in which the degree of overlap of one class in the other classes can be observed. The error matrix diameter and the percentage of correctly classified classes and other cells show the number of assigned errors (column of each class in the error matrix) and deleted row of each class in the error matrix) (Lillesand et al. 2004, Ahmadpour et al. 2011). Finally, the spatial mapping of each method was plotted in Arc GIS 10.2. 3- Results Supervised Maximum Likelihood (MXL) classification method According to the results, the mentioned method has less interference than the other classification methods and except for the class with 40-70% pavement density which has 22.41% interference with the class with 70-90% density, the rest of the applications have less than 8% interference. According to the results, desert pavement with 20-40% density had the highest percentage of 56.62%.   Minimum Distance to Average (MinDis) supervised classification method The overall kappa coefficient for the minimum distance from the mean is 75.54% and the overall accuracy is 81.61%. According to the results, this method has less interference in two classes of 20-40 and 70-90% of desert pavement density than the other two classes. According to the results, desert pavement with a density of 40-70% had the highest percentage of area with 41.88% (9935.79 ha) and the exposed rock with 537.7 ha (1.12%) compose the lowest area in the region.   Mahalani Distance Monitoring Classification Method (MahD) In this method, there is a high degree of overlap between the classes, with the highest interference being between 40-70% and 70-90% of the desert pavement, which accounts for 25.86% of pixels for the 70-90 class. According to the zoning, the desert pavement with a density of 20-40% with an area percentage of 53.55 have the widest class in this method. Also, the lowest percentage of area with 13.39% (6376.43 ha) of the whole area is related to the rock outcrop.   Supervised Classification Method of Parallel Surfaces The overall kappa coefficient for this method was 21.06% and the overall accuracy was 41.25%. The estimated coefficients indicate the inefficiency of the model in separating the pavement classes. The lack of recognition and separation of the two rock outcrops and pavement classes with density of 20-40% has increased the area belonging to the two classes of pavement and the mismatch with real conditions and ground reality, so that, with pavement with 70-90% density and the area covers of 34/99,995 ha (73.52%) of the region. 4- Discussion & Conclusions Given that for optimal use of multispectral data, it is necessary to identify the best band composition. Choosing the best bonding combination is difficult and time consuming visual comparison. Therefore, a technique called Optimal Determination Factor (OIF) can be used to determine the most appropriate band composition to produce the best false color image and to determine the most appropriate bands for digital classification (Alavi Panah, 2000). The results of applying the optimum index in the present study showed that the best band composition for the detection and separation of the desert pavement in the south of Semnan is a combination of 6-4-3 with the optimum index value equal to 71.45 that are located in visible and infrared thermal band (VNIR +TIR). Climatic and geographical differences, satellite harvest time, physicochemical and biological properties of landforms and other effects in the area can produce different results. On the other hand, the classification methods used in this study differ in terms of the structure and complexity of the algorithm, which were calculated to evaluate the performance of the methods, kappa coefficients and overall accuracy. The classification results show that the coefficients of the classification results accuracy obtained from the used methods are considerable. Since the classes, bands, and other conditions used for all methods are the same, the difference in accuracy depends only on the computational algorithms of the methods. In all the investigated methods (except for the least distance from the mean), the class with desert pavement of 20-40% has the lowest kappa coefficients and accuracy, indicating the low ability of the investigated methods in spectral resolution of this class. In general, spectrally separable classes show pretty high accuracy in all methods, and vice versa. The use of other object-based classification methods as well as better spatial and spectral resolution sensors images and the incorporation of properties such as particle diameter can be effective in mapping desert pavement classes.