Comparison of Accuracy of Difference Supervised Classification Methods for Land Use Mapping (Case Study: Alamarvdasht Watershed)

       Satellite data is one of the fastest and the least expensive methods available to researchers to prepare land use maps (Pal and Mather, 2005). Analysis of this data can provide accurate insights into human interaction with the natural environment. In particular, the use of multispectral image analysis can help humans identify land cover (Brian and Michael, 2005). The use of different parts of the magnetic energy spectrum to record the properties of phenomena and the possibility of using hardware and software have made the use of satellite images particularly popular (Richard and Jia, 2006). In general, classification methods can be divided into supervised and unsupervised methods (Ommen, 2008). In the monitored method, we can refer to the maximum probability methods, the minimum distance from the mean, the Mahalanui distance, the parallelogram, the neural network and the support vector machine. In fact, the classification process is the conversion of data into comprehensible information (Rakis, 2011). The maximum probability method is one of the most efficient methods of classifying images (Jensen, 2005). In most research studies, this method has been introduced as the most accurate classification method (Riahi Bakhtiari, 2000; Hovang and Tonshend, 2002). In this method, the user must be careful that the classification follows the normal Gaussian distribution, and this method is more suitable for multispectral classes.

Landsat 8 was launched on an Atlas-V rocket from Vandenberg Air Force Base, California on February 11, 2013. The satellite carries the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS) instruments. The capabilities and advantages of the OLI sensor compared to the ETM + Landsat sensor are as follows: better spectral resolution with narrower bandwidth ranges and 2 more spectral bands (obtaining information in 9 spectral bands), quadrupling the geodetic recording accuracy absolute images, changing the geometry of the image from Whisk broom to Push Broom and thus obtaining150 more information scenes per day (400 images per day), improving radiometric resolution from 8 bits to 12 bits and the possibility of better description of the ground cover and increasing the Signal Ratio to Noise (SNR) (Biranvand and hashim, 2015). In this research, satellite images of Landsat 8 sensor on 13/03/2020 and Envy 4.5 software for processing satellite images and classification of images, and from GIS 10.3 software for creating educational map, ground reality map and layer format conversion were used. Google Earth Pro 4.2 software was used to collect land points and Excel and Office 2018 software were used to generate data tables. According to the purpose of the study, only from visible bands, near infrared, infrared, short wavelength (with cell size of 30 m) and panchromatic band (with cell size of 15 m), to extract spectral values ​​corresponding to the plot of ground samples and statistical analysis have been used. In this study, in order to control the quality of data and increase awareness of atmospheric, geometric and radiometric errors, the data were first studied (Kiani et al, 2014). In the images prepared for this study, due to the newness Landsat 8 satellite and also the non-mountainous area, no device error was observed. The results of the evaluation of the classification accuracy with four algorithms of maximum likelihood classification, minimum distance from the mean, Mahlon distance and parallel level were shown in Tables 2 and 3. In order to better show the effects and reduce the number of bands in the data and to compact the most information of the main bands in the fewest number of bands, principal component analysis was used and the appropriate band combinations were selected. According to research on Google Earth software data, the reseults showed a high confirmation of the accuracy and precision of the software data (Pakravan et al, 2012). In this research, Google Earth software has been used to determine the accuracy of the mentioned classifications. Then, in order to prepare an educational sample from the image of the main components, the three main bands are combined, after which a number of areas or levels are selected as a sample to be used to classify information (Firoozinejad et al, 2014). Also, after carefully selecting the educational samples, the classification of the classes was done using the Training Sample Manager tool and the study area was examined in the classification with 7 user classes (Gong et al, 1996). According to the sampling, 7 land uses were identified in the study area, which include agricultural land, medium rangeland, poor rangeland, barren land, rock outcrop, plains and canals, and residential areas due to errors in the study area occurred with the wasteland as a class. According to the purpose of the study, the baseline image was classified with four monitored classifications, i.e., maximum probability, minimum distance from the mean, Mahalano distance and parallel surfaces. Then it was applied in order to remove single and scattered pixels on the surface of the classified image and also to obtain the desired image of the 3 * 3 model filter. Then, using the obtained results, the accuracy and precision of the classification and further evaluations were calculated using Envy software.

        Preparing land use maps in the study of surface and basement resources and information on current conditions and planning for sustainable management in the future are among the basic principles. Today, the use of remote sensing data and quantitative statistical methods are very common to prepare land use maps (Arkhi, 2014). Easy-to-reach, access to remote and mountainous areas, low cost of data extraction in a short time, wide coverage and reproducibility are some of the benefits of remote sensing data that has been widely used since the last decade (Sofali and Khodarahmi, 2011). According to Tables 2 and 3, the highest producer and land use accuracy is related to agricultural lands, poor rangeland, medium rangeland, barren lands and rock outcrop related to the maximum probability method, which indicates a high percentage of pixels related to the mentioned land uses. In all methods of land use classification, medium rangeland and barren lands had the highest producer accuracy. Mahalanui classification and minimum distance from the average in all land uses were close to the percentage of producer accuracy and only in agricultural lands there was a big difference, which shows the similarity of the two methods mentioned. Table 4 does not classify the parallel surface classification method for 10.2% of pixels. The reason for not classifying some pixels is that the parallel surface method uses the minimum and maximum pixel values ​​for classification. Therefore, the numerical value of the pixels may not be in the range of minimum and maximum classes and may not be known in the range of classes. Lack of classification of pixels means that this method is not used much in research (Yousefi et al, 2014). According to the results of this study, it can be concluded that the maximum likelihood method with a strong statistical basis distinguishes the boundary between classes better than other classification methods (Ahmad et al, 2013). Due to the high spectral resolution of the OLI sensor, the maximum likelihood method provides the best results for the supervised classification of OLI sensor data and the results of this study are consistent with the results of other researchers (Yousefi et al, 2014; Ahmadpour et al, 2014; Ahmad et al, 2013; Nazari et al 2013; Firouzinejad et al, 2014; Kiani et al, 2014).

     One of the most essential information needed by natural resource managers and decision makers is land use maps. Today, remote sensing technology provides a good opportunity to prepare user maps. The value and efficiency of land use maps depend on their accuracy and precision. The purpose of this study was to investigate the efficiency of supervised classification algorithms in preparing land use maps. For this purpose, Landsat 8 satellite OLI images were taken from Alam Rudasht basin on 12/23/1398 and after geometric, radiometric and atmospheric corrections, principal component analysis was performed and appropriate band compositions were selected. The four monitored classification algorithms of maximum probability, minimum distance from mean, Mahalano distance and parallel levels were compared to prepare land use map. Then, in order to remove single and scattered pixels on the surface of the classified image and also to obtain the desired image, a 3 * 3 model filter was applied. The ground reality map was prepared using satellite images to determine the accuracy of the classification. Results of maximum probability algorithms, minimum distance from mean, Mahalano distance, parallel levels with overall accuracy of 88.32, 72, 76.65, 53.3 and kappa coefficient of 0.87, 0.68, 0.73 and 0.45. Finally, the maximum likelihood method was calculated with an overall accuracy of 88.32 and a kappa coefficient of 0.87. The most accurate method is to prepare a land use map.