Research On Small Sample Enhancement Method For High Resolution Remote Sensing Image Classification

The use of high spatial resolution remote sensing images (relered to as "high-resoiuion images") to implement land cover classification is of great significance to precision agriculture,urban planning, resource surveys, disaster prevention and other aspects. As for supervised elassifiation, labeling enough training samples is crucial to optimize the classification model and evaluate the performance of algorithm. However, in the process of label sampling, many factors such as field verifcation, equipment and instrument errors, and the limitations of human eye recognition, will result in long period, high cost and low efficiency. With the improvement of spatial resolution of remote sensing images, the texture structure of ground objects is more refined, and the performance of ground objects in the same area is complex and diverse. Therefore, the heterogeneity of similar ground objects is a major challenge in the process of labeling sampels.In addition, remote sensing images are typical class of unbalanced data, that is, the amount of data to be identified in different categories is quite different, which causes defects on traditional uniform sampling, random sampling, under-sampling, over-sampling and other sampling methods, leading to the generalization of classifier or network strcture model, and resulting in poor classification accuracy.This study has carried out the following four techniques to cope with the above mentioned problems:1.For the geospatial characteristics of remote sensing images, we propose an adaptive neighborhood based on a few initial sample points to obtain potential samples around the seed samples, on the basis of the law that geoscience is related to the principle of similaritv, Then update the location of seed pixels according to the mean shift theory to avoid mixed pixels and noise pixels.2.Due to the diversity of spectral features in high-resolution images,the same feature has different manifestations.We propose a sample enhancement method based on histogram distribution characteristics.Specifically,when automatically labeling sampels,the heterogeneity within the ground species is taken into account effectively.Meanwhile,selecting representative training samples can avoid redundant samples that increase time cost of classification model.Therefore,this study integrates histogram,double window flexible search and boxplot technology into an iterative algorithm to dynamically mine training samples and constantly update classification results.3.To address the unbalanced distribution of remote sensing data and training samples that lead to the generalization of classification model.We proposed a training sample enhancement method based on pixel and object perspectives.Firstly,detect and remove abnormal pixels by obtaining spatial information of pixels and objects.Secondly,label unknown pixels as training samples according to given information.Finally,update and optimize the classification results by iterating.If the difference of the classification results of a certain class of ground object is less than that of the iterative parameters,the training samples of this class will not be increased until all the classes meet the iterative conditions.4.To address insufficient extracted features of regular scale and single scale from high-resolution images.A multi-scale adaptive algorithm is proposed to extract spatial information of high-resolution images and enhance the expression ability of space-spectral features,so as to avoid the error of feature representation by introducing other ground object information into the rule window and overcome the incomplete spatial information extraction caused by a single scale.This algorithm aims to improve the competitiveness of ground features under small samples by adaptive multi-scale enhancement features.In order to verify the effectiveness of the above methods,corresponding experiments are designed respectively.Experiments show that by taking into account the neighborhood features of known samples can effectively enhance the sample representation ability,improve the sample quality and the target recognition accuracy.Compared with the current mainstream methods,the methods in this paper are more competitive and universal.In addition,this paper interprets the application potential of the proposed methods in solving practical problems by interpreting the urban image of Suzhou High-tech Zone and the rural land cover image of Shangrao,Jiangxi.