| Oil is one of the most important strategic resources in the world.The production as well as transportation by tankers and pipelines is extremely active in oceans.However,oil spills caused by drilling platform accidents,capsized tankers and broken pipelines pose risks to the marine navigation environment,causing long-term impacts and serious threats to marine ecology.After an oil spill accident,remote sensing information are used to estimate the amount of oil,track and predict the trend of oil diffusion,which is of great significance for maritime traffic emergency command and oil cleaning.Hyperspectral remote sensing technology is a commonly used and effective method to obtain quantitative information.However,due to the spatial resolution,the information received by the spectrometer is a mixed spectrum of constituent materials in the instantaneous field of view,affectting further application.Therefore,hyperspectral unmixing is a fundamental work for information applications,which is a key point for subsequent target detection,recognition and classification.Marine oil spill is a typical water-containing scene,with frequent and obvious nonlinear effects in seawater.The recognition and extraction of hyperspectral remote sensing information need to be carried out on the basis of nonlinearly mixed spectra.This thesis focuses on hyperspectral images of oil spills at sea,introducing a nonlinear spectral mixing model applicable for oil spill scenarios at sea.Based on the model,an unmixing method is designed for oil abundance estimation,distribution,and thickness.deriving identification of oil distribution and thickness preliminary.The main innovative work of this thesis is as follows:（1）According to the type of intimate mixed oil and water,trigonometric functions are introduced to the classic polynomial model for multi-layer mixture,to characterize the interaction of light between elements at different water depths.A novel polynomial and sine model is proposed to provide a more accurate and detailed description of nonlinear effects in such scenarios.Compared to classic polynomial models,the details of nonlinear interactions are better expressed and quantified,and the reconstruction accuracy is improved for both synthetic and real datasets（reconstruction error in real images is reduced by at least 5.21%）.Both the polynomial and trigonometric parts of the model play important roles in characterizing nonlinearities,with statistically linear dependence areas covering more than 90%and 30%,respectively in real oil spill images.（2）An abundance inversion method is proposed based on the above model.In order to reflect the correlation between the contribution of nonlinear components to the overall mixed spectrum and material content,the nonlinear components of the model were added to abundance inversion,and a norm-based term-wise calculation method was designed to avoid small values being covered by large values.The experiments of synthetic and real images show that the method can achieve smaller abundance reconstruction errors,and the estimated abundance distribution is clearer in main spill and more accurate in details at the thin oil film.（3）In response to extremely complex oil spill scenarios,especially those with low oil content caused by long-term diffusion and collected in ice-covered sea area,this paper proposes a normalized method on the basis of the polynomial and sine model to avoid the overfitting problems.Wavelet transform is performed on the endmembers and collected spectral signatures before abundance inversion.It can effectively avoid the underestimation problem of low oil concentration caused by classical models,and has significant advantages in logarithmic abundance mean square error.However,due to the bias of highlighting low energy information,the error ratio with magnitude difference is relatively large when the concentration is too low.To solve this problem,wavelet packet is used to remove the bias towards low energy,which weights the coefficients derived from wavelet transform inversion in each frequency band according to their energy distribution.The improved method further enhances the accuracy of abundance estimation with real values at magnitude orders from10-1 to10-5.Even in the lowest order,it can still maintain or approach the true magnitude.In real oil spill images,the oil spill diffusion can be accurately displayed when the oil concentration is low.In the simulated oil spill image containing varing magnitude orders of oil,the logarithmic mean square error of the overall abundance estimation is 0.5057.This inversion method can avoid misjudgement for already cleanning-up and cutting off cleaning plans caused by underestmatation of oil.It can be used for long-term monitoring oil diffusion and flux changes,and provides an effective solution for quantitative unmixing of oil spills in the ice-covered sea areas with typically compound mixture types of multi-layer mixing and intimate mixing.（4）In current research concerning oil thickness estimation,hyperspectral information in hundreds of bands are redundant,and the interaction of light in oil and water is not considered.In this article,for oil spill with a certain thickness,the reflection spectrum is regarded as a mixed spectrum,generated by the nonlinear interaction between water and oil,so that oil thickness indices are proposed by use of the proportional relationship between linear and nonlinear effects in the mixed spectrum,and processing flow is designed from them.The relative thickness distribution derived in the real oil spill images is more in line with the real situation,especially for the thickness changes of thicker oil films.The proposed method can provide a rapid inversion in oil spill accidents respond,which is a supplement to previous research methods.In order to provide accurate information on oil volume and diffusion for response in oil spill accidents,this thesis deals with unmixing and preliminarily identifying based on hyperspectral remote sensing data.A nonlinear spectral mixing model and its corresponding abundance inversion method suitable have been proposed,applicable to marine oil spill scenarios;for extremely complex effects in scenarios,especially those with low oil content,normalization method together with wavelet transform and wavelet packet methods are used to avoid the underestimation problem,and achieve accurate inversion of different magnitudes of oil content;finally,the relative oil thickness is quickly and effectively distinguished using the method of nonlinear effect proportion. |
