تلفيق آناليزهاي شيءگرا و پيكسل مبنا به‌منظور شناسايي درختان و ساختمان‌ها در مناطق شهري از داده‌هاي لايدار و نوري

Fusion of Pixel-Based and Object-Based Analysis in Order to Buildings and Trees Detection in Urban Area From LiDAR and Optic Data

شناسايي عوارض شهري مهم مانند ساختمان و درخت از تصاوير هوايي و ماهواره‌اي به همراه داده‌هاي لايدار توجه محققان سنجش‌ازدور و فتوگرامتري را در چند دهه اخير به خود متوجه ساخته است. روش‌هاي طبقه‌بندي و شناسايي مختلف به‌منظور استفاده در مناطق شهري با مشكلات پيچيده‌اي مانند شناسايي ساختمان و درخت‌هاي كوچك، مرز نامناسب تاج درختان، وجود سقف‌هايي از ساختمان‌ها با پوشش گياهي، ساختمان‌ها محصورشده توسط درختان و پوشش‌هاي گياهي قرارگرفته در سايه، مواجه هستند. در اين تحقيق جهت بهبود مشكلات ذكر شده در مرحله اول پيش‌پردازش داده‌هاي لايدار، توليد مدل رقومي زميني و مدل رقومي سطح نرمال شده انجام گرديده است؛ در مرحله دوم ويژگي‌هاي هندسي، طيفي، بافتي و تركيبي متناسب با نوع مشكلات ذكر شده توليد و با استفاده از الگوريتم SVM_GA ويژگي‌هاي مناسب انتخاب گرديده است. از ميان ويژگي‌ها توليدي شاخص طيفي تركيب يافته پوشش گياهي با مناطق سايه و ويژگي‌هاي هندسي توليدشده از داده‌هاي لايدار جهت شناسايي درختان نقش بسزايي بازي مي‌كنند. در مرحله سوم طبقه‌بندي ماشين بردار پشتيبان در دو سطح شيءگرا و پيكسل مبنا با استفاده از ويژگي‌ها منتخب و مناسب با استفاده از روش SVM_GA به‌منظور شناسايي عوارض شهري ازجمله درخت و ساختمان‌هاي شهري استفاده گرديده است. در مرحله چهارم نتايج هريك از دو سطح طبقه‌بندي با استفاده از روش‌هاي پس پردازش بهبود داده‌شده‌اند. در مرحله پنجم با توجه به قدرت هر سطح از طبقه‌بندي در شناسايي عوارض سعي در بهبود شناسايي مرز ساختمان‌ها، درختان و شناسايي ساختمان‌ها و درختان كوچك و كم ارتفاع را با استفاده از تلفيق نتايج آن‌ها را داريم و درنهايت نتايج حاصله از شناسايي در هر سطح و هر مرحله با داده‌هاي مرجع مقايسه و ارزيابي شده است كه شناسايي ساختمان‌ها و درختان در سطح شيءگرا به ترتيب داراي صحت شناسايي 0/961 و 0/65 و دقت كلي 0/97 و 0/943 هستند. در سطح پيكسل مبنا داراي صحت شناسايي 0/953 و 0/632، دقت كلي 0/961 و 0/94 و بعد از تلفيق داراي صحت شناسايي0/971 و 0/718 و دقت كلي 0/975 و 0/957 مي‌باشند. نتايج نشان مي‌دهد كه تلفيق دو سطح از طبقه‌بندي شيءگرا و پيكسل مبنا باعث بهبود نتايج شناسايي به‌ويژه در شناسايي ساختمان‌ها و درختان كوچك و كم ارتفاع مي‌گردد. نتايج اين روش با توجه به بهبود شناسايي‌هاي انجام‌شده در مرز ساختمان‌ها و شناسايي عوارض كوچك تأثير مثبتي را مي‌تواند در استخراج عوارض داشته باشد.

Photogrammetry and Remote sensing researchers pay attention to urban phenomenon detection; such as building and tree from aerial and satellite image with LiDAR data in the recent decades. Different classification and detection methods in order to use in urban area faced with complex problems, such as small trees and buildings, unfavorable boundary crown, roof covered with vegetation , buildings that are surrounded by trees and vegetation located in the shadow. In this research to improve the problem mentioned above, in the first step the following works has been carried out, preprocessing LiDAR data That in the preprocessing step, noise and outliers of the LiDAR data detected by using Frist and Last Range images and then to calculate the proper height for them used the weighted averaging method than the distance, then at this step generation digital terrain model and normalized digital surface model that for generating digital terrain model the method local base filtering is used that has two stage 1- select training data 2- applying filter with using training data that all part of the first step is implemented in the MatLab 7.12 . In the second step geometrical, spectral, textural and combined features have been produced according to the mentioned problems that to produce these features used from normalized digital surface model and first range of LiDAR data. Then these features by using the SVM_GA algorithm for detecting suitable they are used. SVM_GA algorithm implementation is done in tow rounds, that in first round same features are evaluated with each other and in the second round selected features in first round are evaluated with each other. Finally after three times running SVM_GA algorithm 27 features are selected. From the production features vegetation index combined with shadow area and geometrical features produced from LiDAR data have crucial role for trees detection. In the third step support vector machine classification has been used in order to trees and building detection in the object-based and pixel-based levels with using suitable and selected features. Training data for pixel level classification with using SVM classifier are selected semi-automatically in four classes include road, tree, building and vegetation. For this purpose, the following features are used nDSM, Laplacian and Combined_IndexNDVI . for object-based classification with SVM classifier we use training data that they were selected manually and for segmentation of image we use multi resolution segmentation in eCognition 8 software by using the flowing layers, green, red, NIR and Frist range that segmentation parameters are set manually. In order to improving the results of each levels of classification in fourth step we improved with post-processing methods that in this step for both level of classification we have used erosion and dilation morphological filter with different size, so that in object-based level classification, segments of buildings divided into two group: high and low buildings so that for each group of buildings employed morphological filter with different size Separately and then for results of object-based classification we have used a conceptual process for improving object-based level classification. Due to high spectral similarity between two group buildings-roads and vegetation-trees and miss-classification, accuracy, correctness and quality of results have reduced. To improving results have been used neighborly relations and conceptual thresholds in five steps.In fifth step due to the power of each level classification, we try to improve buildings and trees detection with use to fusion object-based and pixel-based classification results, that in this step for building and tree classes have been used different fusion algorithms. The main basis of fusion for building’s class voting results of two level classification and for trees class height pixels in segments and again voting. Results show that fusion of pixel-based and object-based classification improve buildings and trees detection accuracy especially in small and low objects of trees and building also improve crown detection. Fusion the classification results for trees class has more improvement compared to buildings class. Object-based classification level due to using the segments and similarity between trees and vegetation leads to pixels of vegetation not be detected, that in the segments of trees and with use to fusion method for both level of classification, can be detected low and little trees. Results of this method with regards to improving detection in boundary of trees and buildings have positive effect for object extraction. Also suggested method can be detected building with low height and area. After post-processing for detected trees and building classes specified pixel level can be show more capable than object level in detect small object. Finally, results of detection and each levels classification evaluated with reference data, That buildings and trees detection in object level have correctness 0.961 and 0.65 and they have overall accuracy 0.97 and 0.943 respectively, and in pixel-based level they have correctness equal to 0.953 and 0.632 and overall accuracy 0.961 and 0.94 respectively. After the fusion they have correctness 0.971 and 0.718 and they have overall accuracy 0.975 and 0.957 respectively.