Hyperspectral Remote Sensing Inversion Of Soil Salinity Based On Spatial-spectral Constrained Unmixing

Soil salinization has seriously restricted the efficient land use and agricultural development in China.Efficient and high-precision monitoring of soil salinization is an important prerequisite for land improvement and sustainable development planning.The traditional soil salinity monitoring method is realized by field sampling and chemical analysis,which is time-consuming and labor-consuming,and it is difficult to meet the dynamic and large-scale monitoring of soil salinization.Remote sensing technology has the characteristics of easy access,large scale and strong timeliness,which makes up for the shortcomings of traditional methods.With the development of remote sensing technology,hyperspectral remote sensing has become a new method for quantitative estimation of soil salinity due to its high spectral resolution,multiple bands,integrated spectrum and large amount of information.However,the underlying surface of salinized area in the real scene is affected by vegetation cover,and the hyperspectral remote sensing image often exists in the form of mixed pixels,which has become the bottleneck of quantitative monitoring of soil salinity by hyperspectral remote sensing.In this paper,the southern region of the Yellow River Delta is selected as the experimental area.In order to obtain the mixed spectral and hyperspectral remote sensing pixel data set of laboratory simulation.We set up different gradient groups of soil salt content and vegetation coverage to simulate the laboratory mixed scene and measure the field vegetation coverage(FVC)in situ.Combined with the soil salt test data,we carried out the research on the influence of mixed spectrum on soil salt estimation,the thinking and research on the coupling of mixed decomposition model and remote sensing quantitative model.The spectral Euclidean distance measure and heat kernel weighting are introduced,and the"unique variation spectrum"constraint and"smooth space"constraint are designed.Finally,the regional soil salinity estimation and mapping application verification were carried out.The main research contents and conclusions are as follows:(1)Based on the data of simulated mixed spectrum and hyperspectral remote sensing mixed pixels,the spectral characteristics of mixed spectrum and its influence on soil salt estimation were clarified.Firstly,based on the field and indoor work such as soil sample collection,in-situ measurement of vegetation cover,indoor spectral measurement,physical and chemical properties analysis.Then,the laboratory mixed simulation experiment and hyperspectral image mixed pixel acquisition experiment are carried out,and the quantitative evaluation is carried out through spectral analysis and statistical methods.With the increase of FVC in the mixed spectrum,the absorption characteristics of vegetation will be gradually strengthened,and the absorption characteristics under the influence of soil salt will be covered,thus reducing the accuracy of soil salt estimation.For the laboratory simulation spectrum,the group with FVC above 42.28%no longer has the effectiveness of quantitative estimation of soil salt content,and the estimation accuracy of PLSR model is at a low level(R²<0.62,RMSE>7.41 g·kg^-1,RPD<1.4).For satellite hyperspectral,vegetation mixing has a greater impact on salt assessment.When 53%FVC and above groups,the spectrum shows typical vegetation spectral characteristics,and has little similarity with bare soil spectrum.The accuracy of PLSR model for soil salt estimation constructed by satellite hyperspectral remote sensing mixed pixels is not ideal(R²<0.45,RMSE>4.15 g·kg^-1,RPD<1.2).Different spectral preprocessing has different weak effects on the modeling of simulated mixed spectrum and satellite mixed pixel,but it can not be used as an effective means to eliminate the influence of mixed spectrum on soil salt estimation.(2)We propose an unsupervised unmixing based deep extraction method of soil salt hyperspectral information to solve the problem of fusion of hyperspectral unmixing and quantitative remote sensing.Based on non negative matrix factorization(NMF),the proposed method takes the advantage of unsupervised.The corresponding pixel near the sample point is taken as the input unit of spectral unmixing.The soil spectra of all sample points were obtained iteratively and used as spectral variables of remote sensing quantitative model.PLSR,BPNN,SVR and RF were used to verify the effectiveness of the spectra.The results showed that the deep purification of dynamic soil salt spectral information could greatly improve the accuracy of hyperspectral soil salt assessment.However,the mixed spectrum has high additional noise,and the decomposition effect of NMF is uncertain,which leads to high instability of soil salt estimation results.(3)We propose a joint Spearman correlation coefficient and VIP value analysis strategy to explain the impact of mixing spectrum on soil salt estimation,and the reason why unsupervised unmixing can improve soil salt estimation.The mechanism is as follows:with the increase of mixed noise,the important wavelengths related to salt are gradually covered and a lot of uncertain noise wavelengths are added;unsupervised hyperspectral unmixing can effectively eliminate the spectral interference of vegetation,and successfully extract the soil spectrum and include the important wavelength features of soil salt;but when the mixed pixel contains 63.55%or more FVC,unsupervised hyperspectral unmixing is not effective.However,when the mixed pixels contain 63.55%or more FVC,unsupervised hyperspectral unmixing will lose the ability to purify soil salt spectrum.(4)We propose a spatial spectral constrained unmixing method for soil salt estimation to improve the robustness of the unsupervised unmixing based deep extraction method of soil salt Hyperspectral Information.Based on the field investigation,the conditions of"unique spectral variation constraint(UV)"and"spatial smoothing constraint(SA)"are designed,and the updating rules of endmember matrix and abundance matrix are derived.A non negative matrix factorization(uvsnmf)method with spatial spectral constraint is proposed.The unique spectral variation constraint(UV)uses spectral Euclidean distance measure to constrain the similarity of vegetation spectrum,so that the soil spectrum will not be over decomposed and lose the spectral response characteristics of salt,and the error of vegetation endmember caused by vegetation variation can be considered.The spatial smoothing constraint(SA)uses the spectral similarity principle of near pixels weighted by heat kernel to limit the error accumulation of abundance matrix.The results show that the unique spectral constraint and smooth spatial constraint can better purify the spectrum and improve the stability of soil salt estimation.The joint constraint uvsmf algorithm has the best spectral constraint effect in the complex vegetation mixed scene,which can maintain the overall estimation accuracy at a high level,and improve the robustness of unsupervised unmixing in soil salt assessment.(5)Based on UVSNMF,the hyperspectral remote sensing soil salt estimation and regional mapping were studied,and the application of uvsmf in soil salt assessment was tested and explored.In the construction of soil salt estimation model,the direct inversion accuracy of hyperspectral images in some vegetation covered areas is low,which can not meet the requirements of soil salt estimation.In some vegetation covered areas,soil spectra were unmixed and purified by NMF to improve the estimation accuracy of the model,but the RPD was between 1.37 and 1.53,which indicated that the model of soil spectra decomposed by NMF was unstable.After uvsmf decomposition,the spectral model not only has higher accuracy(R²>0.63,RMSE<3.99 g·kg^-1,RPD>1.79),but also has higher robustness.In addition,compared with Kriging interpolation,NMF and UVSNMF have the same spatial structure and more spatial details.UVSNMF can eliminate the abnormal value of soil salinity after considering the constraints of spatial smoothing and spectral variation,which is more consistent with the law of soil salinity distribution than NMF.