Research On Object-oriented High-resolution Remote Sensing Image Classification And Change Detection Method

With the development of satellite remote sensing technology,high-resolution remote sensing images have gradually become the main data source for monitoring land use and land cover changes.However,the increase in image resolution has brought both opportunities and challenges for land use change detection.Traditional pixel-based change detection methods often suffer from the "salt and pepper" effect due to insufficient utilization of image information,making it difficult to meet accuracy requirements.Therefore,this paper explores and studies the method of high-resolution remote sensing image change detection based on object-oriented analysis.Image classification is one of the important factors affecting the accuracy of change detection.Based on an analysis of the basic principles,main processes,and research status of object-oriented image classification and change detection,this paper systematically reviews the main problems faced by these technologies.For example,difficulties in determining the scale of image segmentation,and the inability of a single segmentation scale to consider the extraction of information from objects of different sizes.Additionally,low-dimensional features cannot effectively distinguish similar objects,and too many features can cause the "curse of dimensionality" problem.To address these issues,this paper proposes research into the optimal scale for image segmentation and feature selection,and a multi-level fusion classification and change detection method.After optimizing the extracted image features,multiple segmentation levels are constructed to enable information extraction for objects of different sizes at the optimal segmentation scale.Overall,this paper aims to address the challenges facing object-oriented image classification and change detection technologies by proposing innovative solutions and strategies to improve the accuracy and efficiency of these methods.The specific research content and conclusions are summarized as follows:(1)Research on the selection of the optimal segmentation scale.An optimal segmentation scale selection method based on image texture features is proposed.The method calculates the GLCM texture homogeneity and entropy of the segmented object to construct a model evaluation function,and achieves the optimal segmentation scale selection by combining image spectral and texture information.By comparing the optimal segmentation scales of different image scenes with different methods,it is demonstrated that the optimal segmentation scale selection method proposed in this paper is more applicable to each scene and has higher segmentation accuracy.(2)Image classification under the local optimal segmentation scale.The weight of the heterogeneity factor was determined using the control variable method,and the optimal segmentation scale was selected to classify the study area by combining with the optimal segmentation scale evaluation function established in the previous paper.It is proved that the classification accuracy of different sizes of features at different segmentation scales is different,and the classification accuracy of the corresponding features at each optimal segmentation scale is locally optimal.(3)Optimization of feature space and multi-scale fusion change detection method.The initial feature space is constructed by extracting spectral,texture,geometric shape and thematic index features from the image object layer under multi-scale segmentation,and the correlation between the features is used for PCA feature dimensionality reduction.Based on this,a multi-scale fusion classification post-change detection method is proposed to extract random forest-based change detection results for two temporal phase images and compare the change detection accuracy of different size features with that under a single optimal scale,while applying this paper’s multi-scale fusion method to decision tree-based and support vector machine-based change detection.The experimental results show that the multiscale fusion change detection is more accurate than the single optimal scale in extracting change information and detecting categories more accurately.