طراحي و پياده سازي سيستمي جهت شناسايي انسداد راه‌ها با استفاده از منطق فازي و ماشين‌ بردار‌ پشتيبان (مطالعه موردي: زلزله بم)

Design and Implementation of a System for Roads Occlusion Detection Using Fuzzy Logic and Support Vector Machine (Case Study: Bam Earthquake)

تشخیص راه­های تخریب شده در اثر آوار و تعیین درجه­ی انسداد آنها به منظور امداد رسانی سریع و بهینه پس از بلایای طبیعی امری حیاتیست. در این تحقیق از یك رویكرد ابتكاری برای تشخیص و ارزیابی راه­های تخریب شده به صورت نیمه اتوماتیك استفاده شده است كه از تصاویر پیش و پس از زلزله به همراهِ نقشه­ی برداری راه­ها بهره میجوید. منظور از تخریب و انسداد در این پژوهش؛ آوار حاصل از ساختمانها و همچنین حضور خودرو­های پارك شده بر سطح راه بود كه طبیعتا این حضور امر امداد­رسانی را با اشكالاتی همراه میكند. ویژگی­های مختلفی به منظور طبقه­بندی سطح راه­ها همچون "شاخص سایه" و "شاخص گیاهی" بكار گرفته شد و سپس از یك الگوریتم ژنتیك برای یافتن ویژگی­های بهینه كه فاقد همبستگی و یا خطا باشند استفاده گردید. در این كوشش به منظور دستیابی به دقت كلی بالاتر در تشخیص تخریب از بین دو طبقه بندی كننده نظارت شده MLL و SVM و با روش آزمون و خطا؛ طبقه بندی كننده­ی ماشین بردار پشتیبان انتخاب شد كه نتایج بهتری نسبت به طبقه بندی كننده­ی بیشترین شباهت در مرحله ی تشخیص تخریب داشت. در انتها از یك سیستم استنتاج فازی "ممدانی" برای ارزیابی درجه­ی تخریب یك قطعه از راه استفاده شد و با تعیین یكسری ضوابط آماری روی تعداد قطعات با درجه تخریب معین در یك راه مشخص؛ میزان تخریب به كل آن راه تعمیم یافت. روش مذكور بر روی تصاویر سنجنده­ی Quick Bird از شهر بم ایران تست شد و نتایج دقت بالای 93 درصد و ضریب كاپای 0.91 را در مرحله­ی تشخیص تخریب نشان میدهد. همچنین بیش از 80 درصد از راه­های مورد مطالعه برچسب تخریب صحیح در سه سطح گرفتند كه با داده­های چشمی و تفسیر بصری توسط اپراتور همخوانی داشت. همچنین نقشه­ی قطعات مسدود راه ها با قابلیت ارائه به پروژه­های امداد­رسانی بدست آمد كه دقت بهتر و جزئی تری را عرضه مینمود. این امر نشان از كارایی بالای الگوریتم پیشنهادی دارد.

Damaged and devastated roads recognition and determination of their damage degree seem to be vital when they are affected by a natural disaster like an earthquake. This damage and obstacle is as a consequence of debris caused by collapsed building adjacent to the roads. Moreover, it is essential due to the emergency nature of facing such phenomenon. This study makes use of a new approach for the semi-automatic detection and assessment of marred roads in an affected urban area which is utilizing pre event roads vector map and both pre and post disaster Quick Bird satellite images. In this research, we need to explain a definition for the Damage. As a matter of fact, this damage or obstacle can cause any sort of disturbance on the functionality of the roads network such as conducting rescuers, retrieving survivors, and reconstruction operations. Indeed, in most of urban areas, the width of roads is not that wide, particularly in the third world countries and undeveloped areas. Thus, any trivial obstacle or extra object can cause a noticeable disturbance for transportation. Therefore, this damage is defined as both Debris engendered by collapsed buildings or any other urban structures, and the observation of parked cars on the surface of the roads in devastated areas. To illustrate, this method consists of two main steps; damage detection (by classification), and damage assessment. In this case, many different features are considered for classification of roads surface. These features are such as shadow index, NDVI, and GLCM based features. Furthermore, an appropriate Genetic Algorithm (GA) is designed and used to analyze and find the best set of optimal features. Given that there would be a potential defected band or any correlation among the features, this issue gives useless and unessential information to the classifier and increases the computations time and decrease the accuracy. Afterwards, with making use of these optimal features set and after trial and error between two well-known and prevalent classifiers (SVM and MLL), the supervised Support Vector Machine classifier was selected. It is because of gaining higher overall accuracy and enhancing the damage detection consequently. Thus, SVM is applied to the optimal features to detect damage (damage detection step). Since, a road is a slim object and to analyze the obstacle of this slim object more meticulously, it is needed to divide it into smaller parts. After dividing each individual road to several and equal partitions, a designed ‘Mamdani’ fuzzy inference system (FIS) is represented for the road damage assessment step. These three damage levels are including Low, Medium, and High damage levels. It is based on each small partition. That says, each partition goes into the Fuzzy inference system as a point and the output is the partition damage level index. Afterwards, some statistic criteria are considered on the number of different damaged partition and the damage level is generalized on each individual road. Therefore, each single road gets a damage label and lead to a roads damage level map. The proposed method was tested on QuickBird pan sharpened image from the Bam earthquake and the results indicate that an overall accuracy of 92% and a kappa coefficient of 0.9 were achieved for the damage detection step, and 82% of the roads were labeled correctly in the road damage assessment step. The obtained results show the efficiency and accuracy of the current algorithm.