شناسایی ساختمان های تخریب شده بعد از زلزله سرپل ذهاب با استفاده از آنالیز چند سنسوری و یادگیری ماشینی

ارزیابی سریع مناطق تخریب شده بعد از هر فاجعه طبیعی مخصوصاً زلزله از اهمیت ویژه ای در پاسخ به بحران برخوردار است که در صورت استفاده از الگوریتم های مناسب با چشم انداز ماهواره های بلادرنگ می تواند کمک شایان توجهی در کاهش تلفات زلزله داشته باشد. زیرا که گسیل نیروهای امداد و نجات هدفمند به سمت و سوی ساختمان های تخریب شده خواهد بود و در نتیجه زمان بلاتکلیفی آنها کاهش خواهد یافت. در این مطالعه از 3 تصویر رادار با گشودگی مصنوعی (sar) اخذ شده از ماهواره alos2 با قدرت تفکیک مکانی 10 متر در دو پلاریزاسیون hh و hv و همچنین 2 تصویر اپتیکی مرتبط با ماهواره worldview-2 با قدرت تفکیک مکانی 0.46 متر در چهار باند برای ارزیابی تخریب لرزه ای ناشی از زلزله سرپل ذهاب (اِزگله) سال 1396 شمسی با پنج الگوریتم naive bayes (nb)، knearest neighbors (knn)،support vector machine (svm) ، regression tree (rt) و random forests (rdf) مورد بررسی قرار گرفتند. مجموعاً 24 پارامترهای در روش یادگیری ماشینی برای داده های sar از آنالیز بافت، مقادیر ضریب بازپراکنش و همدوسی تولید شده تداخل سنجی راداری (insar) مورد استفاده قرار گرفتند. در تصاویر اپتیکی نیز 20 پارامتر صرفاً از مولفه های مستخرج از آنالیز بافت استفاده شدند. کلاسه بندی نتایج بر اساس دو گروه ساختمان های تخریب شده و ساختمان های تخریب نشده انجام پذیرفت که صحت کلی کلاسه بندی برای هر دو دسته نشان می دهد که الگوریتم rdf قابلیت و صحت بالاتری برای ارزیابی تخریب ارائه می نماید.

Identifying Collapsed Buildings after Sarpol-e Zahab Earthquake Using Multisensor Analysis and Machine Learning

Earthquakes and their consequences should be studied in detail in order to reduce the number of casualties in future events. From the beginning of the twenty first century until now more than 800000 deaths were reported, in which most of the casualties are located in AlpHimalayan seismic belt. Bam earthquake in 2003 in central Iran, with more than 26000 casualties, Indian Ocean earthquake in 2004, with approximately 200000 casualties, Sichuan earthquake in 2008 in China with more than 96000 casualties, and Haiti earthquake in 2010 in Haiti with approximately 321000 casualties are only a few given examples that how devastating the earthquakes can be. Instant deaths right after a strong earthquake is primarily because of physical contact of rubbles material with exposed people, but the second phase of casualties emerge due to injuries, suffocation of trapped people among the rubbles and wasted materials, and collateral hazards such as fire. Although the instant deaths look inevitable, second phase casualties can be decreased by addressing rapid disaster response based on recent remote sensing earth observation systems to bring the quality of search and rescue teams to an actionable level, especially for nighttime earthquakes. In SAR remote sensing imagery, addressing of seismic damage states initiated with simple indices such as difference and correlation of SAR backscatters of pre and postevent images, difference of coherence value of interferometric phase analysis, and their combination. Furthermore, regression analysis of SAR backscattering of pre and postevent images together with seismic intensity were also applied for deeper understanding of the earthquake damages. In the recent developments of earthquake damage assessment, combination of multitemporal dualpolarized SAR data, combination of multitemporal ascendingdescending SAR data and only postevent SAR data are common methods to decrease the level of uncertainty. In the optical remote sensing, damage assessment was initiated by visual comparison of pre and postevent images. However it is possible to apply methodologies based on only postevent images if lower accuracy is needed. Therefore, visual interpretation of optical images, rather than automated change detection, is widely used in practice for building damage detection. Saito et al. (2004) visually interpreted collapsed buildings using three IKONOS images taken before and after the Gujarat earthquake, and confirmed the quality of the results by ground survey data. Further, Saito and Spence (2005) compared the visual interpretation results from only postevent QuickBird images with those from pre and postevent images, and revealed that the building damage tended to be underestimated when only postevent images were available. Adams et al. (2005) used a visualization system integrated pre and postevent QuickBird imagery to direct rescuers to the hardest hit areas and support efficient route planning and progress monitoring in the emergency response phase of the Bam earthquake. By comparing the pre and postevent QuickBird imagery visually, Yamazaki et al. (2005) classified the damaged buildings caused by the Bam earthquake into four damage grades (EMS98). Comparing the results to field survey data revealed that the preevent imagery was more helpful in detecting lower damage grades through visual interpretation.Here various machine learning based techniques for performance understanding of the classifiers in an urban scale is presented.This study covers a comprehensive seismic damage assessment of Sarpole Zahab town in western Iran which was affected by an earthquake M 7.3 on 12 November, 2017. The damage concept is evaluated using both synthetic aperture radar (SAR) and optical images. Two preevent and one postevent dualpolarized high resolution SAR images of ALOS2 satellite, and one preevent and one postevent very high resolution optical images of WorldView2 satellite (4 bands) are contributed in the comprehensive seismic damage assessment. In SAR dataset, twentyfour influential parameters are extracted from interferometric phase correlation (differential coherence), differential intensity, and differential texture analysis of HH and HV channels, whereas in optical dataset, twenty influential parameters are derived from differential texture analysis of red, green, blue and infrared (IR) bands. For the derived parameters of each dataset, principal component analysis (PCA) and machine learning based algorithms (i.e. random forests, support vector machine, naive Bayes, knearest neighbors and regression tree) are carried out in order to extract the damage maps and their related accuracy with respect to the calibration data which is acquired from United Nations Institute for Training and Research (UNITAR).