Object-oriented extraction research under multiple scenarios based on Sentinel-2A satellite images

Abstract

Abstract In order to solve the problem of information extraction from Sentinel-2A satellite remote sensing images, four scene images of mountainous landscape with complex forests, flat cultivated rural landscape suburban coastal landscape and complex urban landscape were selected as experimental objects, and three algorithms of KNN, CART and SVM of object-oriented method are used to extract the main feature information, so as to compare the extraction accuracy of different methods in different scenes. The experimental results show that the overall accuracy of object-oriented feature classification of Sentinel-2A image in four scenes is greater than 75.47%, and the kappa coefficient is greater than 0.72 except that the overall accuracy of KNN in Scenes 4 is 0.66. Through the comprehensive comparative analysis of the four scenes, it can be seen that the average overall accuracy of rural landscape of Sentinel-2A image is 84.65%, the average kappa coefficient is 0.79, which is better than the average overall accuracy of urban landscape classification of 78.1%, the average kappa coefficient is 0.7, and the flat terrain scene is better than the mountainous area with large topographic elevation difference. In terms of classification methods, SVM algorithm has better classification accuracy than CART algorithm, and CART has better classification accuracy than KNN.