Hyperspectral Remote Sensing Based Models For Soil Moisture And Salinity Prediction

As one of the three major irrigation areas in China,Hetao Irrigation District in Inner Mongolia faced with an extremely serious soil salinization problem.Salinization will result in soil degradation,crop yield reduction and ecological destruction,which seriously restricts the agriculture development in this area.The water and salt monitoring of an irrigation area is the basis of the status and law of soil water and salt transportation analysis and prediction.It is quite important to formulate reasonable soil salinization improvement measures.Hyperspectral remote sensing technology can acquire volume nano-scale spectral information of ground objects.High spectral resolution reflects the subtle characteristics of ground objects spectrum.It can quantitatively analyze and retrieve water and salt information based on diagnostic spectral absorption characteristics.In this paper,the sand trench irrigation area of Hetao Irrigation Area is taken as the research area.Soil samples with different salinization degrees are collected.The soil moisture content,total salt and eight main water-soluble base ions(K⁺,Ca²⁺,Na⁺,Mg²⁺,Cl^-,SO₄^2-,HCO₃^-,CO₃^2-)are obtained by analysis.With the measured soil hyperspectral data,the inversion models of soil moisture content,total salt and eight major ions were established by using different characteristic spectral screening methods and mathematical modeling methods,and the models were evaluated comprehensively according to different assessing indexes.The main conclusions are as follows:?1?In the soil moisture content inversion,it is found that the soil spectral reflectance curves waveforms under different water content are basically similar.It has verified that the reflectivity decreases first and then increase with the water content increasing.It is found that the water absorption valleys of reflectivity were obvious near 1450 and 1950 nm,and was unapparent near 500 and 2200 nm.In model inversion,stepwise regression SR model and ridge regression RR model based on reciprocal logarithmic LR index have the highest determinant coefficient,which indicates that LR is the optimal spectral index for water content inversion.In addition,ridge regression,a biased estimation,can screen less but high precision feature band.The LR-RR model based on LR index,with the highest R_c² and R_p² of0.975 and 0.979,RMSE of 0.012 and RPD of 5.89 respectively,performs best in fitting and predicting.The results showed that after ridge regression screening,the collinearity has been heavily weakened,and the model has a strong robustness.?2?With the feature bands extracted by three different methods?VIP,SR and GC?,PLSR and SVR models of partial least squares regression for different salt-based ions were established.The results showed that PLSR model and SVR model have performs well in most ions modeling and predicting,but the best band screening methods for different base ions varies.Among them,VIP method screens more characteristic bands?34.5%⁴2.5%of the whole band?,and the model accuracy is the highest;SR method screens the least bands?1.5%⁴%of the whole band?,and the model accuracy is the lowest;the number of bands screened by GC method varies greatly?7%⁵5%of the total band?,and the accuracy of the model is general.The best spectral inversion prediction results of different water-soluble saline ions in soil are quite different,among which Ca²⁺,Na⁺,Cl^-,Mg²⁺and SO₄^2-ion models have the best prediction results,with RPD values of 3.97,3.15,2.98,2.75 and 2.75,respectively.The prediction results of other ions are general or fail to invert.?3?Comparing the soil spectral curves under different total salt content,it was found that the soil spectral characteristics tended to be consistent in morphology.With the increase of salt content,the soil spectral reflectance did not increase significantly.After one-dimensional correlation analysis of spectral data pretreated by fractional-order derivative?FOD?,it was found that FOD could clear the change trend of spectral reflectance correlation and soil salinity information.After the spectral differentiation of most ions was processed,the number of bands that pass the significance test shows a trend of increasing first and then decreasing,but the corresponding order was not the same one when the number of bands was the largest.Among the different machine learning models,the extreme learning machine?ELM?has the best inversion effect.Among them,the models of total salt,Ca²⁺,Na⁺,Cl^-,Mg²⁺and SO₄^2-ion have the best predictive effects,and their RPD values are2.631,2.328,2.869,3.264,3.054 and 2.259,respectively.The predictive effects of other ions are general or cannot be inverted.