Research On Oil Well Target Detection Method In High Resolution Satellite Remote Sensing Images

As one of the major oil consuming countries in the world,the import volume of crude oil in China is513 million tons in 2021,and the degree of foreign dependence is as high as 72%.It might can provide data support for China’s energy security strategy for mastering the global oil field extraction status under the situation of serious dependence on oil energy imports.And the oil recovery well network composed of oil wells is an important representation of the oilfield exploitation status.Therefore,it has important theoretical and application values for China’s energy security in monitoring the number of oil wells in global oil fields.In addition,oil extraction activities cause serious damage to natural resources such as soil,air and water,and facilities such as oil extraction wells encroach on a large amount of land resources.Therefore,monitoring the density of oil wells in oil fields is of great significance for environmental protection.With the development of optical remote sensing imaging technology and deep learning-based target detection technology,automatic detection of oil wells from high-resolution satellite remote sensing images has become possible.However,the current studies on target detection of oil and gas elements are focused on facilities such as oil tanks and well sites,and no studies on target detection for oil wells have been done so far in the academic field.This thesis firstly constructs a dataset for oil well target detection in high-resolution remote sensing images-NEPU-OWOD V2.0 to this end.A semi-automatic annotation model with manual and model collaboration is designed since it is labor-intensive to annotate oil well targets in high-resolution satellite remote sensing images.Improve model accuracy with manually labeled samples,reduce labeling costs with model prediction samples,and promote the spiral improvement of models and data.Then the mainstream target detection algorithms are compared and analyzed on the constructed well dataset,and the Faster RCNN model with the highest detection performance is optimized and improved.The main improvement strategies include adding an edge detection module to the Faster RCNN algorithm to suppress the background noise,which effectively improves the detection accuracy of the Faster RCNN;using the proposed Cluster RPN sub-network instead of the original RPN sub-network to improve the accuracy of the Faster RCNN for the detection of densely arranged clumped wells,thus enhancing the overall detection of the model performance.This thesis has completed a prototype system for oil well target detection from high-resolution remote sensing images based on the above analysis.The functions of data management and model training are demonstrated.It is proved that the prototype system of oil well target detection with high resolution remote sensing images can support the research of remote sensing oil well target detection with good application value.