Object-based Land Cover Classification And Mapping Based On High Resolution Remote Sensing Images

With the development of remote sensing image acquisition technology and object-oriented image analysis,the research of land cover classification based on object-oriented image analysis technology has gradually been adopted by most researchers.However,the multi-level classification of image segmentation objects and the optimization of classification feature space still need to be further studied,especially the effective combination method of classification features and multi-level image classification.This paper adopts WorldView II and GF-1multi-spectral data and integrates the Optimization of classification features into object-based multi-level image classification to realize land cover classification mapping.Through the research,the following conclusions are obtained:(1)In this paper,the principle of ReliefF feature selection algorithm is taken as the reference,and through its modification,the feature optimization work of extracting single-class ground objects in multiple categories is realized.The modified ReliefF algorithm can better determine the preferred classification feature set of the classification object category.In order to verify the effectiveness of the method,three classifiers were applied to compare and analyze the difference between the classification results based on the traditional object classification and the method of this paper.Experiments show that the modified ReliefF algorithm can effectively improve the classification accuracy of categories to be classified while reducing feature dimensions.(2)Multi-level image object classification is an inevitable choice for realizing the integrity of image objects in various categories.This article compares and analyzes the differences between object classification based on multi-level imagery and traditional object-based classification.The classification results show that the classification of multi-level image objects ensures high-precision classification,and can make the boundary information of each image classification more accurate,and thus the image classification and drawing results are more beautiful.(3)The accuracy of classification based on WorldView II images in the same study area is significantly higher than that based on GF-1 images.The highest accuracy on World View II images is to increase the classification accuracy by 7.38% and at least 1.17%.However,the highest scoreon the GF-1 image can only increase the classification accuracy by 3.04%,and the lowest is 0.34%.The main reason is the high spectral resolution and high spatial resolution of WorldView II images.(4)It is feasible to determine the classification order of multi-level image objects based on the J48 decision tree.Before classifying multi-level image objects,it is necessary to determine the classification order of each classification category.Based on the J48 decision tree,the classification order of multi-level image objects of each category is determined according to the principles of “top-down,easy to difficult”.Experiments show that multi-level image object classification based on this scheme can achieve higher classification accuracy.(5)Image shadow objects have high separability.By analyzing the spectral curves of similar features under shadow and non-shadow,it is found that similar features have similar spectral characteristics when covered with shadows and without shadows.With the help of this feature,the spectral information of the shadow segmentation object is used to realize the reclassification of shadow objects,and the real land cover classification information under shadow coverage is obtained.The classification results show that the shadow classification results can be obtained with high precision by directly applying the spectral information of shadow objects,and the image objects covered by the shadow have good separability.