Aiming at the demand for rapid detection of heavy metals in soil,a selection for characteristic wavelength variables was performed based on model population analysis method,and a method of X ray fluorescence spectroscopy was proposed for the detection of contents of heavy metals in farmland soils.X ray spectrum values of 91 configured soil samples were collected,and used to establish a soil heavy detection model.The characteristic wavelength variables were extracted by the multi feature series method.Firstly,the interval combination optimization algorithm(ICO) was used for rough selection of the wavelength.Secondly,the competitive adaptive reweighted sampling(CARS) was adopted to remove those irrelevant variables in the interval wavelength.Finally,the successive projections algorithm(SPA) was performed for wavelength reduction.The ICO-CARS-SPA algorithm with multiple features was used to select the feature variables in the X ray fluorescence spectrum to obtain 5(26,25,29,39,33) characteristic wavelength points.Based on this,a partial least squares(PLS) detection model for the contents of five heavy metals,i.e.Cu,Zn,As,Pb,Cr in soil was developed,which was compared with other traditional characteristic wavelength selection methods.Results showed that the variables selected by the ICO-CARS-SPA algorithm were combined with partial least squares(PLS) to have the best modeling effect.The determination coefficients of Cu,Zn,As,Pb,Cr were 0.993 3,0.992 6,0.995 6,0.993 2,0.988 6，respectively.The root mean square errors of five metals were 6.938 5,23698 4,3.632 6,8.510 6 and 14.764 5，and the mean relative biases of the prediction sets were 0.255 1,0.065 0,0.102 5,0.241 4,0.104 7,respectively.The ICO-CARS-SPA algorithm based on X-ray fluorescence spectroscopy combined with multi-feature series strategy could eliminate invalid wavelengths,thus increasing contribution of effective information,simplifying detection model complexity,and providing a theoretical support for selecting an appropriate feature band extraction method for the prediction model of soil heavy metal content.