پديد آورندگان :
وحيدي ،ميلاد دانشگاه صنعتي خواجه نصيرالدين طوسي , صاحبي ،محمودرضا دانشگاه صنعتي خواجه نصيرالدين طوسي , بابايي كفاكي ،ساسان دانشگاه آزاد اسلامي-واحد علوم و تحقيقات
كليدواژه :
رادار با روزنه مجازي , ابرطيفي , انتخاب ويژگي , بهينه‌سازي ژنتيك , طبقه‌بندي كننده SVM
چكيده فارسي :
جنگل به عنوان يكي از منابع سخت تجديدپذير در محيط‌زيست به شمار مي‌آيد. كسب اطلاعات از اين منابع، همواره مورد توجه سازمان‌ها و مديران بخش‌هاي منابع طبيعي بوده است. سنجش‌ازدور به‌عنوان علمي قوي و تا حدودي مقرون‌به‌صرفه، توانايي در اختيار قرار دادن اطلاعاتي از قبيل نوع گونه‌هاي اصلي، تخمين زيست‌توده، شناسايي و طبقه‌بندي تك‌درختان و غيره را از منابع جنگلي دارد. نحوه بهره‌گيري تصاوير سنجش‌ازدوري به منظور بهبود نتيجه طبقه‌بندي، امروزه مورد توجه محققين مي‌باشد. وجود جنگل‌هايي با تنوع گونه بالا و همچنين شباهت طيفي و ساختاري گونه‌هاي جنگلي، ضرورت استفاده توأمان از تصاوير رادار و اپتيك را بيشتر مي‌كند. از اين رو، هدف اين مقاله ارائه الگوريتمي مي‌باشد كه از داده‌هاي پلاريمتري و ابرطيفي به‌صورت توأمان استفاده مي‌كند. به طوري كه اطلاعات ساختاري و سطحي از تصوير پلاريمتري و اطلاعات رنگي، طيفي و شيميايي از تصوير ابرطيفي استخراج شود. الگوريتم از دو مرحله اصلي تشكيل شده است. در مرحله اول قطعه‌بندي تصوير و تفكيك مناطق جنگلي از غيرجنگل انجام مي‌شود. در مرحله دوم، ويژگي‌هاي مختلف از دو مجموعه داده، براي هر قطعه استخراج مي‌شوند. ويژگي‌هاي پلاريمتري در دسته‌هاي ويژگي‌هاي اصلي، المان‌هاي تجزيه‌كننده‌ها و تفكيك‌كننده‌هاي SAR و همچنين، ويژگي‌هاي اپتيك شامل ويژگي‌هاي اصلي، ويژگي‌هاي مرتبط با محتواي شيميايي گونه‌ها و نسبت‌هاي بازتابندگي مي‌باشند. به منظور انتخاب بهينه ويژگي‌ها از الگوريتم انتخاب ويژگي ژنتيك و همچنين از الگوريتم ماشين بردار پشتيبان به‌منظور تهيه تصوير طبقه‌بندي‌شده بهره گرفته مي‌شوند. نتايج حاكي از آن مي‌باشد كه استفاده از الگوريتم ژنتيك و تمام ويژگي‌ها دقت كلي 78/82 درصدي و ضريب كاپاي 36/79 را حاصل كرده است. ويژگي‌هاي مبتني بر محتواي شيميايي درخت و شاخص‌هاي طيفي و بازتابندگي در ناحيه مادون قرمزكوتاه به همراه درجه پلاريزاسيون، مؤلفه Kd تجزيه كروگاگر، مؤلفه H از تجزيه H/Alpha و Lambda از تجزيه كلود-پوتير به عنوان ويژگي‌هاي مؤثر در طبقه‌بندي معرفي شدند.
چكيده لاتين :
Forest has been introduced as one of the resolvable sources in environment. The forest regions considerably effect on metrological condition and CO2 content of the regions. Hence, Management and preservation of these sources is so critical and important for forestry organizations. Since, propounding tree species maps are essential and useful issues for managers and also because of vast area of the forest, remote sensing could be powerful tool for forest mapping in a large terrain. Optical and synthetic aperture radar (SAR) are two prevalent remote sensing imaging systems which have high capacities to provide different information, such as recognizing type of tree species, biomass and CO2 content estimation, tree chemical combinations and tree species classification which commonly could be utilized in forest regions management. In this paper, because of same spectral and structural behaviors of trees to each other and existence of various types of trees in forest, a new algorithm has been developed to classify tree species by using Hyper-spectral and POLSAR images. The algorithm consists of two main stages, first the co-registered image is segmented to certain groups of homogenous pixels and forest and non-forest segments are separated to each others. Since in SAR images every object of land surface has certain scattering mechanism and because of volume mechanism of forest regions, we utilize polarimetric signatures so that recognize volume scattering mechanism in image. Second, specific features of each segment are extracted from Hyper-spectral and POLSAR images. Reflectance in each band, continuum removals features in special spectral ranges and spectral indices related to chemical contents tree structures, stress and spectral ratios are some extractable features from Hyper-spectral image and on the other hand, POLSAR features include original features, decomposition methods, and SAR discriminators. Although both hyperspectral and SAR images provide large number of features, but some of them have correlation to each other and became as extra features which should be removed in trend of process. For this reason, non-parametric feature selection algorithm has been proposed to select effective features among all. To choose the optimum features, genetic algorithm is applied and then, trained SVM algorithm in feature space with optimal dimensions classifies image to certain tree species. To find out ability of each type of features in classification, the features are divided to groups with certain number of features and overall accuracy and kappa coefficient related to each group is calculated. The results demonstrated that combination of GA algorithm and all features has more accuracy than others algorithm which was about 82.78% in overall accuracy and 79/36 in kappa coefficient. Spectral indices related to chemical behaviors of trees and reflectance in SWIR regions have better performance than other mentioned optical features. Some effective polarimetric features, such as lambda, degree of polarization, Alpha/H of cloude-potier decomposition, span and Kd of krogager decomposition were some freatures with high ability in tree species classification. SAR features play major roles in three parts of the result; first, constructed Polarimetric signature of SAR images was useful for displaying volume mechanism and separation of forest and non-forest segments. Second, mentioned features of SAR images had high ability to distinct pine species of needle leaf species from broad leaf species. The last application of SAR images was in helping the algorithm to classify short leaf loblolly and loblolly from each
