Object-oriented Classification Of Mine Features

With the continuous development of remote sensing technology,people's requirements for remote sensing image classification are becoming higher and higher in practical production and application.At present,in the mine geophysical interpretation work of the Aviation Geophysical Remote Sensing Center,it still relies on the personnel with rich mine interpretation experience and professional geological knowledge to carry out human-computer interactive interpretation.This method of work is very easy to produce fatigue and very inefficient.In this paper,based on high resolution remote sensing image and object-oriented method,mine features are automatically classified.On the basis of radiation correction,geometric correction,registration and image fusion for high-resolution images,a classification system of eight types of terrain,including buildings,transit sites,bare land,stopes,mountains,waters,vegetation and roads,is established.The main contents of this paper are as follows:(1)Using multi-scale segmentation technology,the image is divided into image object units corresponding to the actual object categories.The main parameters affecting the segmentation results are scale parameter,spectral factor,shape factor,smoothness factor and compactness factor.According to the analysis and comparison of the segmentation results under different segmentation scales,the optimal segmentation parameters suitable for the classification of mine features are finally obtained.When the segmentation scale is 300,the shape factor is 0.3 and the compactness is 0.7,all kinds of objects can be segmented to the maximum extent.(2)After segmentation,the image object unit has some characteristic parameters,including gray mean,brightness mean,length-width ratio,shape parameter,maximum difference measure and normalized vegetation index.A part of image object entity units with prominent features are selected as samples,and the object features in the sample space are reconstructed to establish classification rules for different objects,and the whole image object units are classified according to the nearest neighbor idea.The brightness value of stope,building,road and bare land is higher.The transit site is generally near the stope.The texture characteristics of mountain body are complex.The normalized index of vegetation is between 0.1 and 0.25.The water body boundary is smooth,and the shape of road and building is regular.(3)Combining with relevant reference data,the final classification results are qualitatively analyzed.The classification results of stope,transit site and building are basically consistent with the actual situation.The obfuscation matrix is used to evaluate the quantitative accuracy of the classification results of eight types of terrain objects.The producer accuracy and user accuracy of stope reached 88.5% and 88.98% respectively.The producer accuracy and user accuracy of transit stope reached 83.51% and 84.40% respectively.The overall classification accuracy reached 83.94% and the Kappa coefficient was 0.8046.