مقايسه كارايي برخي الگوريتم‌هاي طبقه‌بندي در مطالعه تغييرات لندفرم‌هاي بياباني دشت يزد-اردكان

Comparison the performance of some classification algorithms in study of desert landforms changes in Yazd-Ardakan plain

شناخت لندفرم­ها و تغييرات آن‌ها از نيازهاي اساسي ژئومورفولوژي مناطق بياباني به‌منظور مديريت آن‌هاست. يكي از روش‌هايي كه در شناخت تغييرات لندفرم ها مورداستفاده قرار مي‌گيرد طبقه‌بندي پديده‌ها و عوارض زمين با داده‌هاي سنجش‌ازدور در زمان‌هاي مختلف مي‌باشد. هدف از اين پژوهش، كارايي سه الگويتم نظارت‌شده حداكثر شباهت، فاصله ماهالانوبيس و حداقل فاصله براي طبقه‌بندي لندفرم‌هاي بخشي از دشت يزد-اردكان و بررسي تغييرات آن‌هاست. بدين منظور از تصاوير سنجنده هاي TM، ETM و OLI ماهواره Landsat مربوط به سال‌هاي 1987، 2000 و 2016، در بازه زماني 30 ساله استفاده شد. ابتدا تصحيحات لازم بر روي تصاوير انجام گرديد و سپس با بازديدهاي ميداني و بر پايه نقشه‌هاي توپوگرافي، ژئومورفولوژي و تصاوير گوگل ارث، لندفرم‌هاي ژئومورفولوژي منطقه موردمطالعه تعريف و نمونه‌هاي آموزشي انتخاب گرديد. نتايج ارزيابي دقت تصاوير طبقه‌بندي نشان داد الگوريتم حداكثر شباهت در سال‌هاي 1987 و2000 و 2016، به ترتيب با دقت كلي 91/50، 93/22 و 93/35 و ضريب كاپا 0/87، 0/89 و 0/89 نسبت به دو الگوريتم ديگر، عملكرد بهتري در تهيه نقشه ژئومورفولوژي و تغييرات آن داشت. بر اين اساس به‌منظور شناخت وضعيت تغييرات لندفرم ها از روش آشكارسازي تغييرات "پس از طبقه‌بندي" از روش الگوريتم حداكثر شباهت استفاده شد و نقشه ماهيت تغييرات تهيه شد. همچنين نتايج طبقه‌بندي و تحليل تغييرات لندفرم ها طي 30 سال با روش الگوريتم حداكثر شباهت نشان داد كه سبخا و لندفرم‌هاي موجود بروي دشت‌سر پوشيده به‌جز تپه‌هاي ماسه‌اي مواج داراي روند نزولي تا سال 2016 بودند و در تغييرات لندفرم‌هاي بياباني منطقه عوامل طبيعي توأم با عوامل انساني نقش داشتند. به‌طوركلي اين نتايج نمايانگر آن است كه در محدوده موردمطالعه در طي 30 سال، نه‌تنها بر وسعت اراضي بياباني افزوده نشده است، بلكه از محدوده اراضي بياباني كاسته شده و بر ديگر كاربري‌ها افزوده‌شده است. همچنين در شرق و غرب منطقه مرز لندفرم ها تغيير كرده است.

Introduction Today, using remote sensing (RS) instead of the visual interpretation to identify landforms and their change detection is a necessity. One of the steps in identification of change detection is the classification of remote sensing data. There are various methods for classification of satellite images used for geomorphological landforms mapping and identification of their variations that each has advantages and limitations. The most common classification methods can be noted the maximum likelihood classification. Other classification methods such as the minimum distance, Mahalanobis distance and neural network have attracted much attention. The main aim of this case study is the classification and surveying of geomorphological landforms changes in part of Yazd-Ardakan plain using Landsat images over 30 years. The study area is located at 31° 47′ 16″-32° 13′ 20″ northern latitude and 53° 40′ 37″-54° 27′ 04″ eastern longitude. This area is part of the Saduqh county in Yazd province. The area of the study area is 1563/11 square kilometer. Methodology The research method in this study is survey-analytical. To study the changes of desert landforms, satellite imageries of TM, ETM and OLI of Landsat satellite in 1987, 2000 and 2016 were used. First, the radiometric and atmospheric corrections was performed using Flaash algorithm, and then the geomorphological landforms were introduced and the training samples were selected by field observations, topography and geomorphology maps and Google Earth images. To classify the landforms, three supervised classification algorithms were used, including maximum likelihood, minimum distance and Mahalanobis distance. Then, the accuracy of classified maps was evaluated using the overall accuracy and the kappa coefficient metrics. Finally, to evaluate the changes of landforms, the "Post classification" method was used and the change detection map of nature was made. To analyze the database, ENVI 5.3, ArcGIS 10.4.1 and Excel 2013 softwares were used. Results and discussion Totally, 15 landforms were identified in the study area Which include Alluvial Fan, Glacis Pediment Plain, Clay Pan, Glacis Dennoyage Plain, Inselberg, Glacis Epandage Plain, Kalut, Erg (barchans, longitudinal dunes, barchanoid), Hill, Mountain, Salt Dome, Sebkha and Sand Sheet. After image correction, geomorphological maps were prepared. The results showed that the best algorithms for classification of landforms is the maximum likelihood method in 1987, 2000 and 2016 years, with the overall accuracy and the Kappa coefficients of 91/50, 93/22 and 93/35 and 0/87, 0/89 and 0/89, respectively. Finally, to investigate the nature of the changes, the comparison method of "post classification" was used that was applied on maximum likelihood algorithm as it had a more favorable outcome. Then, the changes of landforms were calculated in terms of its area and percentage. The results showed that the dominant class was the Glacis Epandage Plain in 1987 and 2016, but this class had downward trend in 2000. Then, hills with 17/58 percent of the total area is ranked next highest area in 1987, but this class had downward trend with 11/58 percent in 2016. In 1987, Barkhan class had the lowest area with 0/17 percent but this class had downward trend with 0/11 percent in 2016. Conclusion The results of the landforms classification showed that the maximum likelihood algorithm, offerd a more detailed classification method than the minimum distance and Mahalanobis distance algorithms. The area and the percentage of landforms changes over 30 years showed that landforms such as Barchan, Clay Pan, Longitudinal Dunes, Barchanoid and Kalut had a downward trend because they were located in the context of the development of the city. Also the area of sand dunes landforms like Longitudinal Dunes, Barchan and Sand Sheet, had a slight increase until 2000. The results showed that Sabkha area, has been a downward trend over 30 years, The matching of results with the results of previous studies showed that with the increasing of the wells in Yazd-Ardakan plain, it was provided to cultivate large areas of saline lands. Therefore, the natural and human factors were involved in changes the desert landforms in the study area.