The Study Of Disaster Target Automatic Classification Based On High-resolution Remote Sensing Images

China has a vast territory and is a country prone to natural disasters. Natural disasters have caused serious loss to the government and people. The remote sensing technology can be well used in the progress of disaster assessment and relief. It has the features of observing large area synchronously, and getting images data with objective authenticity. Using image classification technology to extract disaster target rapidly and direct relief operations in post-disaster, it can increase relief efficiency and minimize the loss of life and property effectively. Visual interpretation is the traditional way to classify remote sensing images. This approach has a slower interpretation speed and cannot meet the demand of quickly acquiring the image classification products in post-disaster. With the development of technology, more and more remote sensing images with high spatial resolution and high temporal resolution can be applied in many fields. Combined with disaster assessment application demand, this research is based on disaster target feature library, realizes the progress of disaster target automatic classification with high-resolution images, and applies to disaster rapid assessment.The main research contents of this paper include:(1) Studying disaster target feature library. This paper builds classification systems for different natural disasters based on their characteristics and the application requirements of disaster assessment, and designs disaster assessment application oriented disaster target feature library. It can effectively store the information of disaster events which have occurred, and use these information as prior knowledge to better serve the progress of disaster assessment and relief. This paper designs the general framework and the detailed physical structure of disaster target feature library, and analyzes the progress of feature information extraction.(2) Studying sample automatic selection algorithms. According to the structure of disaster target feature library, this paper proposes a sample automatic selection algorithm which includes two steps: initial sample selection and sample correction. The paper introduces the algorithm flow in detail, analyzes spatial-temporal proximity rules, and studies the technology of remote sensing image change detection and the knowledge about object spectrum characteristics. This paper also verifies sample automatic selection algorithm by experiment, and shows that it can increase the efficiency of image classification.(3) Studying feature selection and image classification algorithms. On the basis of sample automatic selection, this paper focuses on m RMR(Min-Redundancy and Max-Relevance) feature selection method. According to the principle of m RMR algorithm, this paper uses three computation methods to realize the feature selection process and uses two supervised classifiers(C5.0 decision tree and k-nearest neighbour) for image classification experiment based on feature selection results. This paper also compares m RMR algorithm with principal component analysis method. The experiment shows that m RMR algorithm has optimum effects in image classification progress.(4) Building disaster rapid assessment model and carrying out the application demonstration. To summary the whole research contents of the paper about disaster target automatic classification technology, it takes earthquake disaster assessment as an example application in the last. This paper builds the model of earthquake disaster rapid assessment, and experiments with post-disaster UAV(unmanned aerial vehicle) images of Ludian earthquake in Yunnan province. The assessment results are accurate, included damaged building proportion, damaged building area and seismic intensity. It can indicate the validity of this model.