Remote Sensing Information Extraction For Elements Of Surface Mine Based On Bag-of-Words Model

In China,the irrational exploration of mineral resources has caused various eco-environmental problems.Scholars home and abroad increasingly focus on how to exploit and utilize mineral resources properly and implementation of dynamic monitor management to mine.Now,obtaining mine distribution and surface mine production situation relies mainly on the spot investigation or artificial visual interpretation using high-resolution remote sensing images,but this needs enormous manpower and material and financial resources.This paper select mining area in HuaYuan County of Hunan Province as study area,and choose IKONOS data and remote sensing images from Google Earth to establish multiple Bag-of-Words(BOW)model to express image features in study area.Through different remote sensing image classification methods,we finally realized the recognition,classification and extraction of surface mine elements.The main conclusions are as follows:(1)We use BOW model and Spatial Pyramid Matchkernel Based BOW(SPM_BOW)model to classify 120 remote sensing images respectively.When using BOW model,99 images have been correctly classified,and the classified accuracy is 82.51%.While using SPM_BOW model,109 images have been correctly classified,and the classified accuracy is 90.83%.It is clear that SPM_BOW model is quite better than BOW model.(2)The overall classification accuracy and Kappa coefficient of Probabilistic Latent Semantic Analysis(PLSA)based on Multi-Scale BOW(MS_BOW)model are 90.50%and 0.88 respectively.While the overall classification accuracy and Kappa coefficient of object-oriented classification method based on the region of interest(ROI)are 86.25%and 0.83 respectively.Although two methods can both meet the classification requirements,PLSA based on MS_BOW model is much perfect.(3)PLSA classification methods based on MS_BOW can correctly classify mine heaps,tailing ponds and cistern.The number and location of these features are also content with the actual situation,and the accuracy is all above 80%.The main errors exist in mine road classification,in which case mine roads are mistakenly classified as mine heaps.Object-oriented Classification method based on ROI can classify tailings ponds and cistern precisely,however,the classification accuracy of mine heaps and mine road identification is slightly lower,and there are more false classifications,especially the recognition of mine roads,mine construction and residential buildings.(4)BOW model use the histogram to express the characteristics of remote sensing image,which includes a certain special semantic information.Compared with traditional method,BOW model can express mine elements more precisely and effectively,and it can also largely improve the classification accuracy.However,SPM_BOW model is superior to the classic BOW model,as the former one possesses special characteristics.SPM_BOW model can express the global and local features of mine images,and make up for the lack of special semantic information.(5)Compared with traditional Object-oriented Classification method,MS_BOW model is based on the principle of multi-scale segmentation,and contains multi-scale image information Using PLSA method to classify high resolution images can solve the problems about polysemy and synonymy to a great extent.Our results show that MS_BOW model is much more suitable to extract surface mine elements from high resolution images.