Models For Predicting The Li Content In Salt Lake Based On Remote Sensing

As the world’s important emerging industry and High-tech industry resources,Li resources are known as the"energy upstart"and widely used in batteries,chemical industry,medicine,astronautics,nuclear energy and other fields.Its value and importance far surpasses other precious metals.With the industrial structure adjustment and economic transformation,China is increasingly paying attention to high-tech industries and the demand for Li resources is increasing year by year.China is a large country of Li resource reserves.Because of the technical level of utilization,it is impossible to extract high-purity Li,so Li resources are still mainly imported from abroad.According to the global Li resource storage,more than 78.3%of Li resources are stored in salt lake brine,so the salt lake has very important development and utilization value.With the rapid development of remote sensing technology,the rapid and low-cost prediction of Li resource content in salt lakes by remote sensing technology has received more and more attention.This paper selects the Arizaro Salt Lake as the study area in Salta Provin ce of Argentina and uses Sentinel-2A as the data source.Combined with the measured Li content of the salt lake brine and correlation analysis method,thr ee Li content remote sensing prediction models are established namely PCA re gression model,BP neural network model and RF model to predict Li content in salt lakes,the main results are as follows:（1）Li content remote sensing prediction model is established by using th e spectral reflectance of salt lake and the transformed spectral variables and th e measured values of the salt lake as content.Firstly,in order to improve the correlation between the original spectral reflectance and Li content,the principa l component analysis of Li index shows that the correlation coefficient between the reflectance of lnPC1 component image and Li content is increased to-0.852.so the PCA regression model is established by using the spectral variables of lnPC1.Secondly,Original spectral reflectance undergoes reciprocal logarithmi c transformation?ln?（1?R）,logarithmic transformation ln R,first order differentia l transformation R^′,second order differential transformation R^′′,first-order diffe rential transformation of the reciprocal logarithm ln?（1?R）′.The BP neural netw ork model with ln?（PC1）、ln?（PC3）、ln（1?R₆）′、R₆^′combinations as input varia bles is established.Finally,RF model established with the same spectral variabl es ln?（PC1）、ln?（PC3）、ln（1?R₆）′、R₆^′.（2）By comparing the decision coefficients R² and RMSE of the three m odels,it is found that R² of PCA regression model is 0.725,RMSE is 29.07,R² of BP neural network model is 0.731,RMSE is 15.719,R² of RF model is0.771,RMSE is 12.822.Eight test points are used to test the prediction accur acy of the three models.It is found that the relative error of the RF model pr ediction result is relatively low,the minimum relative error is 0.052,and the maximum relative error is 0.213.Therefore the sequence of Li content predicti on ability by three models is RF model>BP neural network model>PCA regr ession model（3）The optimal RF model is used to predict the Li content of the whol e salt lake.It is concluded that the Li content is not uniformly distributed in t he whole salt lake,but mainly distributed in the north,southeast and southwest of the salt lake.It is proved that the RF model can effectively predict the Li content in the Arizaro salt lake.