An experimental comparison of pixel-based and object-based classifications with different machine learning algorithms in landscape pattern analysis – Case study from Quang Ngai city, Vietnam

• Published in: IOP conference series. Earth and environmental science Vol. 1345; no. 1; pp. 12019 - 12033
• Main Authors: Vu Viet Du, Quan, Minh Pham, Tam, Manh Pham, Van, Duy Nguyen, Huu, Huy Nguyen, Quoc, Thanh Pham, Viet, Cao Nguyen, Huan
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.05.2024
• Subjects: Accuracy; Algorithms; Classification; Image classification; Land cover; Machine learning; Pattern analysis; Pixels; Spatial distribution; Support vector machines; Vietnam
• ISSN: 1755-1307; 1755-1315
• DOI: 10.1088/1755-1315/1345/1/012019

Abstract In landscape pattern analysis, the choice of an efficient method for image classification is widely studied, but the features are mostly extracted from digital values by the traditional approach (Pixel-based) and modern approach (Object-based). In this study, we compared the performance of two supervised classification algorithms (Maximum likelihood classifier - MLC and Support vector machines - SVM). We used SPOT-5 image data from 2011 to analyze the landscape pattern of a complex territory in Quang Ngai City, Vietnam. We collected 215 ground-truth samples and classified them into seven landscape classes. The results showed that the overall accuracy of the classification of Pixel-based (MLC), Object-based (MLC), Pixel-based (SVM) and Object-based (SVM) was 67.9%, 74.0%, 72.1%, and 82.3%, respectively. The combination of the object-based approach and SVM algorithm had the best classification result, which reflected the current spatial distribution of land cover types accurately. In the next step, we computed landscape metrics from detailed images to compare quantitative parameters with the actual verification data sources. From these metrics results, we discussed how classification methods could affect landscape structure, and possible ways to improve the accuracy of landscape-pattern identification.