| In recent years,marine oil spill incidents in China occur frequently,causing serious pollution and great harm,which has seriously endangered China’s social and economic development and marine ecological environment security.Therefore,effective emergency monitoring technology is urgently needed for accurate monitoring.However,due to the characteristics of sudden,high dynamic,low recognition and wide range of drift and diffusion,especially in the complex marine environment,the ability of emergency monitoring is greatly limited.The scope of oil spills is related to the spatial distribution and quantity of oil spills,which is an important reference for determining the source of oil spills and on-site disposal;the types of oil spills involve traceability punishment and decontamination treatment,which is an important basis for oil spill pollution disposal.Passive and active optical remote sensing are important means of oil spill monitoring,in which passive optical remote sensing has the advantages of hyperspectral and high spatial resolution,and is most widely used in oil spill monitoring.At present,the problem of using optical remote sensing to monitor oil spill on the sea surface is that it is easily affected by clouds and sea surface sunglints,and it is difficult to identify effectively different light oils.In addition,most of the optical remote sensing monitoring of oil spill on the sea is carried out based on a single dimension of spectral or polarization or infrared data,and the detection and identification accuracy is difficult to meet the needs of marine oil spills accurate monitoring.In conclusion,with the support of related projects,this paper designs an outdoor oil spill experiment,carries out the extraction and analysis of hyperspectral,polarization and infrared multi-dimensional optical features of oil spill on the sea,and introduces new classification algorithms such as integrated learning and deep learning to build a multi-dimensional optical remote sensing accurate detection and identification model of oil spill,which can provide important technical support for marine emergency monitoring and oil spill monitoring business of relevant business departments.The main research contents and conclusions are summarized as follows.(1)Under the support of related projects,the outfield multi-dimensional optical observation experiments of oil spill were carried out,and the multi-dimensional optical data of five typical oils(crude oil,fuel oil,palm oil,diesel oil and gasoline)and sea water were obtained,including ground object spectrum,multi angle visible polarization,long wave infrared polarization and airborne hyperspectral data.The spectral characteristics of five oils at different solar altitude angles were analyzed.The spectral separability analysis of oil-water,heavy oil and light oil,light oils were carried out,and the variation of visible polarization information of different oils with the zenith angle and sun-camera relative azimuth angle was discussed.The results show that: 1)The continuum removal method can effectively highlight the absorption and reflection characteristics,and the separable spectral ranges of the three light oils are expanded,mainly concentrated in the ultraviolet and blue-green bands;2)When the azimuth is fixed,the visible degree of polarization of the five oils are approximately symmetrical at the zenith angle of-50° to 0° and 0° to 50°,and the minimum value is at the zenith angle of 0°,that is,when the camera is vertically downward,and the maximum value is at the range of 40° to 50°.For a zenith angle,the maximum value of visible degree of polarization of five oil products is located at 180° of relative azimuth,and the minimum value is located at 0° of relative azimuth,that is,the direction of solar mirror reflection;3)The degree of long wave infrared polarization increases with the increase of oil film thickness of crude oil and fuel oil,and the intensity contrast of infrared polarization image of heavy oil and sea water increases with the increase of zenith angle.(2)Aiming at the problem that the optical remote sensing detection of oil spill on the sea surface is vulnerable to the interference of sea surface sunglints,considering that the sunglint belongs to high-frequency component and is concentrated on a certain scale image,the combination of multi-scale information obtained by wavelet transform and DCNN(Deep Convolutional Neural Network)algorithm for mining deep features can suppress the sunglint to a certain extent.In this paper,a hyperspectral oil spill detection model based on multi-scale feature DCNN is constructed.The input of the model includes visible,near infrared and short wave infrared spectral dimensions,and also involves the spatial dimensions of different image scales.The ability of single classifier to mine sample features is limited.Considering that deep learning and shallow learning algorithms have their own characteristics in the generalization ability of sample learning,decision fusion based on fuzzy membership degree can inherit the advantages of both.In this paper,a hyperspectral remote sensing detection method of oil spill based on deep learning and shallow learning decision fusion is developed.Based on the airborne AISA+ hyperspectral data of Penglai 19-3 oil spill in 2011 and airborne AVIRIS hyperspectral data Gulf of Mexico oil spill in 2010,the method application experiments were carried out.The experimental results show that: 1)Compared with SVM(Support Vector Machine),RF(Random Forest),DBN(Deep Belief Network)algorithms,the oil spill detection accuracy of multi-scale DCNN model can be improved by 1%～5%,in which the algorithm accuracy of original scale combined with level-one scale is the highest,reaching 87.51%;2)The decision fusion model based on fuzzy membership degree can further improve the oil spill detection accuracy by 2%～10% compared with single classifiers;3)The improvement of accuracy is mainly reflected in the reduction of missing and wrong points in the sunglint area of the original image,which proves that the multi-scale DCNN model and decision fusion method based on fuzzy membership degree are suitable for the accurate detection of oil spill under the sunglint.(3)Aiming at the problem that it is difficult to effectively identify different light oils,this paper develops a RF(Random Forest)integrated learning oil identification method based on spectral/infrared polarization features and a hyperspectral RPnet(Random Patch Network)deep learning model based on multi feature fusion.The method is applied to the multi-dimensional optical data such as land-based visible polarization,long wave infrared polarization and airborne hyperspectral under the same oil spill scene.At the same time,this paper constructs a SVM oil identification model of spectral separability of typical oils.Using the ASD(Analytical Spectral Devices)full spectrum of five typical oil products and sea water,the spectral separability analysis of oil-water,heavy-light oil and light oils is carried out by using continuum removal,spectral standard deviation method and factor analysis method,and then the model experiment is carried out by using spectral separability characteristics.The experimental results show that: 1)Compared with the single dimension visible intensity image,the visible polarization multi-dimensional image can effectively identify gasoline,diesel and palm oil,but it is not sensitive to the same kind of oil film with different thickness;2)Compared with the single dimension long wave infrared intensity image,the recognition rate of light oils of the long wave infrared polarization multi-dimensional image is evenly improved 7.77%,which proves that the random forest ensemble learning oil recognition method based on multi-dimensional features can effectively recognize different light oils,and can also recognize different thickness of crude oil and fuel oil;3)The best characteristic spectral ranges of five oils recognition is 360～540 nm,560～600 nm,610～630 nm and640～660 nm,and the oil identification accuracy is 11% higher than that of the full spectrum,reaching 90.74%. |
PDF Full Text Request