Research On Extraction Method Of Mechanical Damaged Surface Based On GF-2 Remote Sensing Image

Mechanical engineering,such as mining,engineering construction,soil and rock filling,has caused serious damage to the surface of the earth and a large number of mechanical damaged surfaces,which not only damage the ecological environment,but also easily lead to soil erosion,landslides and other natural disasters.It is very important for ecological environment protection,natural resource management,urban planning and so on to acquire mechanical damage information in time with the help of remote sensing technology.However,there is no report on the method of mechanical damage surface extraction.When the actual situation of the surface is complex,it is easy to "over segmentation" and "under segmentation" phenomenon to extract the object based on the single scale segmentation patch.At the same time,due to the different generation time,the different degree of damage and the different degree of weathering and other reasons,the spectral characteristics are different.The situation of "the same material and different spectrum" and "the same spectrum foreign matter" is serious,and it is easy to be confused with other features,which greatly restricts the extraction accuracy of the mechanical damaged surface.How to extract the mechanical damaged surface efficiently and pertinently is an urgent problem.In view of the above problems,GF-2 image was used as the data source to explore a method suitable for the extraction of mechanical damaged surface,and the method was verified in the area with dense mechanical damaged surface in Tanglangchuan basin.Firstly,based on the principle of multi-scale segmentation,combined with RMAs index and maximum area method,the optimal segmentation parameters of various features in the study area were obtained.Secondly,the spectrum and shape characteristics of all kinds of ground objects were analyzed.The ground objects that are easy to be distinguished by mechanical damaged surface belong to the non-target layer,and the buildings,bare land and mechanical damaged surface that are seriously confused belong to the target layer.According to the difference of spectral diversity,the objects in the target layer were divided into several sub categories,and a multi-level classification system wase constructed.In the non-target layer,the rules were mainly constructed by combining the spectral and shape features of the objects,and the membership classification method was used for classification and extraction.In addition to the spectral and shape features,texture features were also added to the objects in the target layer,and relief F algorithm was used to optimize the features,and the mechanical damaged surface was further extracted by establishing the extraction rules of the objects.Finally,in order to verify the superiority of the classification method,the multi-level object-oriented classification method without texture feature,the single level object-oriented classification method with texture feature and the classification method based on pixel were used to extract the mechanical damaged surface,and the precision of the extraction was compared and analyzed.The results show that:(1)Combining spectral diversity to build a multi-level classification system,not only effectively avoid the problem of "over segmentation" and "under segmentation",but also effectively make up for the defect of "the same thing and different spectrum" in the image.Compared with the single level classification method,the user accuracy of the mechanical damaged surface extracted by the multi-level classification method is improved by 10%,the overall accuracy is improved by 12%,and the kappa coefficient is increased by 0.15%.(2)The texture feature is added to the extraction rules of mechanical damaged surface,which effectively solves the problem of serious confusion of mechanical damaged surface,bare land and buildings in the target layer.Compared with the multi-level classification method without texture features,the accuracy of the mechanical damaged surface extracted by this method is improved by 20%,the overall accuracy is improved by 6%,and the kappa is improved by 0.07%.(3)The object-based classification method improves the extraction accuracy of mechanical damaged surface,while the pixel based supervised classification method has obvious salt and pepper phenomenon.Compared with the pixel based supervised classification method,the user accuracy of the mechanical damaged surface extracted by this method is 23% higher,the overall classification accuracy is 19% higher,and the kappa coefficient is 0.23 higher.(4)In this paper,the multi-level classification method with texture features is proposed.The user accuracy of mechanical damaged surface is 89%,the overall accuracy and kappa coefficient are 88% and 0.86 respectively.Furthermore,the multi-level classification method with texture features can effectively improve the extraction accuracy of mechanical damaged surface.To sum up,based on GF-2 image,aiming at the problem of mechanical damaged surface recognition and extraction,this paper constructs a multi-level classification method with texture features.After comparison and verification,this method can effectively and accurately identify mechanical damaged surface information.It has an important application value for the extraction of the ground feature information in the remote sensing image.