این پژوهش با هدف استخراج نقشه کاربری اراضی شهری، با استفاده از مقایسه الگوریتمهای مختلف طبقهبندی پیکسلپایه و شئگرا میباشد. در این راستا الگوریتمهای طبقهبندی پیکسلپایه ماشین بردار پشتیبان، حداکثر احتمال، شبکه عصبی مصنوعی، حداقل فاصله از میانگین، سطوح موازی و فاصله ماهالانوی مورد استفاده قرار گرفتند. در ادامه به مقایسه روشهای مذکور با طبقهبندی شئگرا جهت تهیه نقشه کاربری اراضی شهر زنجان با استفاده از تصویرماهوارهایSentinel-2 با قدرت تفکیک مکانی 10 متر پرداخته شد. به منظور انجام پردازش تصویر مورد استفاده از نرمافزارهای ENVI 5.3، SNAP،eCognition و ArcGISاستفاده شده است. برای مقایسه عملی نتایج، در هر دو روش از دادههای آموزشی یکسان برای طبقهبندی استفاده گردید؛ سپس مهمترین روشهای ارزیابی صحت شامل دقتکلیو ضریب کاپای طبقهبندی استخراج شد. نتایج بدست آمده، نشان میدهد که از بین روشهای طبقهبندی پیکسلپایه مورد استفاده در این مطالعه، روشهای طبقهبندی حداکثر احتمال و روش حداقل فاصله تا میانگین با ضریب کاپای به ترتیب 95/0درصد و 85/0 درصد از دقت قابل قبولی برخوردار هستند. همچنین مقایسه نتایج حاصل از طبقهبندی پیکسلپایه و شئگرا نشان داد که روش شئگرا با اعمال پارامترهای مؤثر در طبقهبندی و توسعه قوانین جهت اطلاح طبقهبندی اولیه شئگرا با ضریب کاپای 95/0 درصد از نظر دقت در استخراج نقشه کاربری اراضی از روشهای پیکسلپایه از اولویت برخوردار است.
عنوان مقاله [English]
Extracting Land Use Map based on a comparison between Pixel-Based and Object-Oriented Classification Methods Case Study: Zanjan City
Currently, two general methods are used for classification of digital satellite images: pixel-based and object-oriented processing. Unlike pixel-based Methods, object-oriented techniques employ different geometric, spatial, spectral, and form-based algorithms, and selecting the most efficient algorithm in this process requires a lot of experience in image processing. In addition, multiple algorithms usually offer different results and this in many cases makes the selection of efficient algorithms difficult. In general, pixel-based classification includes supervised and unsupervised methods. Examples of these methods include maximum likelihood, neural network and support vector machine. Maximum likelihood method is one of the most effective methods used for image classification. Object-oriented methods take advantage of knowledge-based algorithms, and thus overcome problems pixel-based method faces because of not using geometric and textual information. In order to achieve high classification accuracy, two methods of pixel-based and object-oriented classification are compared in this research. On the one hand, integrated planning and management of urban areas, and on the other hand, collecting reliable information regarding land use makes this kinds of studies indispensable.
Present study seeks to extract urban land use map. Thus, necessary data was received from Sentinel-2. Moreover, ENVI 5.3, eCognation 9, SNAP, ArcGIS 10.3, Google Earth, and land-use data were also used to process images and analyze data. In SNAP, atmospheric correction process was performed on images collected from the study area using SEN2COR plug-in. Samples collected from each class of Sentinel-2 satellite image were mapped on the image area. Pixel classification algorithms, support vector machines, maximum likelihood, artificial neural network, Minimum Distance to Mean (MDM), parallelepiped and Mahalanobis distance were used. Finally, land use classes (residential, gardens and green spaces, wastelands and passageways) in the study area were mapped using different classification algorithms. For object-oriented classification using nearest neighbor algorithm, the satellite image was first segmented in eCognation software using the Multiresolution Segmentation Algorithm. Parameters such as scale, shape and compactness were also studied in the image segmentation stage. Through trial and error, an appropriate value was selected for parameters used in segmentation. For practical comparison of the results, the same educational data was used in both object-oriented and pixel-based classification methods. Then, the most important methods for assessing accuracy including overall precision and kappa coefficient were extracted.
Results & Discussion
As one of the most important methods used for extracting information from remotely sensed images, classification allows users to produce various types of information such as coverage maps, and land-use maps. Classification of satellite data includes segregation of similar spectral sets and classification of sets with the same spectral behavior. Regarding the resolution of images used (10 m) in this study, only 4 land-use classes possessed the required resolution capability for pixel-based classification of Sentinel-2 satellite images. These classes include built-up (residential) area, waste land, urban green space and street network. In this regard, support vector machine, maximum likelihood, artificial neural network, Minimum Distance to Mean, parallelepiped and Mahalanobis distance were used for classification. Classification results indicate that compared to other pixel-based methods, maximum likelihood method and Minimum Distance to Mean method show a precision of 85% or higher.
In present study, geometric properties of land use classes (including scale, shape, and compactness) were used for segmentation and this process was performed by multiresolution method. For this purpose, results of image segmentation process were analyzed based on different parameters (with different scales) and spatial resolution of the image. In this way, appropriate values for segmentation were selected based on the specific features of the study area (an urban environment) through trial and error. Then, the proper image segmentation was selected and prepared for the classification stage using the above mentioned parameters. In the next step, 20 effective parameters including statistical indices, mean score of bands, NDVI index, standard deviation of the bands and geometric index were used for classification.
The present study took advantage of six pixel-based methods (Support Vector Machine, Maximum Likelihood, Neural Network, Minimum Distance to Mean, Parallelepiped, and Mahalanobis) along with object-oriented classification method to produce a land-use map for Zanjan city. The accuracy of classification in different methods were compared and statistically analyzed using overall accuracy coefficient, kappa coefficient, user’s accuracy, and producer’s accuracy. The results of statistical analysis of the accuracy coefficients indicated that Minimum Distance to Mean and Maximum Likelihood method -with a Kappa coefficient of 90% and 85% respectively- are acceptable methods for land use mapping. Moreover, comparing pixel-based and object-oriented methods, it is possible to conclude that object-oriented approach with a Kappa coefficient of 0.95% and overall accuracy of 97.9% shows a higher potentiality. Nearest Neighbor algorithm is one of the most important reasons for achieving this high accuracy in object-oriented classification. In addition to the spectral information, this method uses information collected about issues like texture, form, position, and content for the classification process.
Methods used in this study prove the accuracy of objective-oriented technique by employing effective parameters and developing rules to modify the initial classification of object-oriented technique. Another advantage of object-oriented method (as compared to pixel-based methods) is that apart from spectral information and statistical data, it is possible to apply several other indicators such as shape, texture, color, dimensions and altitude of the phenomena in the final land use map produced by this method. Finally, it should be noted that object-oriented classification has been developed for high resolution spatial data.