Research On Detection Of Oil Pollution In Soil Of Oilfield And Its Spatial Distribution Using Multi-source Remote Sensing Data

Oil,often referred to as the"blood"of the modern industry,is the most important energy for human society,and is highly related to national economic development and defense security.With the expansion of oil exploitation,oil leakage events during exploitation,refining,storage,and transportation occur frequently;the vast majority of them occur on land.With a vast amount of oil production and consumption,China stands as the world’s largest developing nation.Since 2000,China’s annual crude oil output has been maintained at about 190 million tons.In the early stage,oil pollution has been caused by the backwardness of production conditions and environmental protection technology,of which soil oil pollution needs to be paid attention.Given that soil is the core element of the ecosystem and plays an extremely important role in the ecological environment.Oil pollution in soil will seriously threaten our lives.The traditional detection method for oil pollution in soil needs to carry out considerable on-site sampling and indoor testing,which is time consuming and labor intensive and is easily affected by the spatial distribution of sampling points.In addition,the detection results are not meeting the required standards in terms of timeliness.However,with the advancements made in machine learning theory,the semantic segmentation algorithm for high-resolution remote sensing images based on deep learning can extract ground objects accurately.Hyperspectral satellite can image the land surface repeatedly with a wide coverage and a short revisit period.The combination of deep learning algorithm and hyperspectral imaging technology makes it possible to detect oil pollution in soil quickly and accurately.In this paper,a remote sensing detection method for oil pollution in soil was proposed.The method used an advanced deep learning algorithm,which was trained using soil samples from the oil exploitation area,and combined with multi-source remote sensing data,to detect oil pollution in soil.The spatial distribution of oil pollution in soil in the study area was then deeply discussed using spatial statistical analysis methods.The primary research contents and conclusions of this study are outlined below:(1)Based on the hyperspectral data and petroleum hydrocarbon content of soil samples,a new algorithm for hyperspectral characteristic band selection and content estimation of soil petroleum hydrocarbon was proposed.The algorithm combined the grouping search algorithm with the point-by-point search algorithm.The features obtained by feature extraction methods have poor interpretability and missing physical information.Meanwhile,feature selection methods are difficult to implement because of the massive combinations of bands.Considering the issues outlined above,a new hyperspectral characteristic band selection and content estimation algorithm for soil petroleum hydrocarbon was proposed.It combined the advantage that the grouping search algorithm can minimize the time required of computation with the advantage that the point-by-point search algorithm can determine the importance of each band.The content of soil petroleum hydrocarbon was then estimated using 17 selected characteristic bands.The RMSE and R~2 of the estimation result were 3.52 and 0.90,respectively,which proved the effectiveness of the characteristic band selection and content estimation model applied in this study.Moreover,this method provided a new idea for the selection of characteristic band and content estimation of other substances in soil.(2)A suitable backbone feature extraction network for bare soil area extraction was selected on the basis of the bare soil dataset made for the study area.In accordance with the land cover status,the land cover interpretation marktable of the study area was established.The bare soil dataset was created using GF-2 images following the remote sensing image properties of bare soil areas.For the selection of the backbone feature extraction network,the loss value and m Io U of Res Net50 were 0.124 and 81.67%,respectively;the loss value and m Io U of Mobile Net V2 were 0.338 and 69.31%,respectively;the loss value and m Io U of Xception were 0.123 and 82.01%,respectively.The performance of Res Net50 was similar to that of Xception,but the parameters and size of Xception were nearly 1.5 times those of Res Net50.After comprehensive comparison,Res Net50 was selected as the backbone feature extraction network in this study.(3)On the basis of the ideas of sliding prediction,expansion prediction,and Deep Labv3+algorithm,a deep learning algorithm for semantic segmentation of large-scale remote sensing images was proposed.The performance of existing semantic segmentation algorithms is limited by the characteristics of large-scale remote sensing images,such as wide imaging range,high data volume,and complex background.In this study,the Deep Labv3+algorithm was improved,and sliding prediction was used to enhance the processing ability for large-scale remote sensing images.At the same time,the expansion prediction method was used to improve the accuracy of bare soil extraction.The improved Deep Labv3+was applied to measure the bare soil in the study area(about 1037.6km~2).A total of 28537 bare soil vectors were extracted,with an area of 19017528.17m~2.This algorithm provided a feasible solution for semantic segmentation of large-scale remote sensing images.(4)The remote sensing information of oil pollution in soil in the study area was extracted on the basis of hyperspectral images.The spatial distribution of oil pollution and the oil production factors were analyzed using spatial statistical methods.Through combining with the hyperspectral data,estimation algorithm of soil petroleum hydrocarbon content,and the bare soil vectors,this study obtained the information regarding the distribution of oil pollution in soil.The oil production factors,such as oil production equipment and station yards in the study area,were extracted manually.A total of 10933 production wells,2363989.18m roads,and 10102 various production station yards were interpreted,with a total area of 15250520.48m~2.The spatial relationship between oil production factors and oil pollution in soil was analyzed using spatial statistical methods,such as Kernel Density,Moran’s I,and Directional Distribution.The results showed that the risk of oil leakage at processing stations in the study area should be highlighted.A positive correlation between the degree of oil pollution in soil and the density of oil wells and roads in the study area was determined.Plants also exerted a certain mitigation effect on oil pollution in soil.