The more frequent human activities with the modernization of the society intensify the soil heavy metal pollution.When the content of heavy metal elements in the soil exceeds its risk screening value,there may be risks to human health.Therefore,screening out the soil with the risk of heavy metal pollution is an important part of soil pollution control.The spectral data of 59 national standard soil samples were obtained by X-ray fluorescence (XRF) spectroscopy,and then pre-processed by wavelet soft threshold denoising and iterative discrete wavelet transform background deduction.Moreover,the competing adaptive reweighted sampling (CARS) algorithm was applied to screen the heavy metals in the soil.Finally,the screened results wereinput to the one-dimensional convolutional neural network (1D-CNN) model to predict whether soil samples were at risk of heavy metal contamination.The results showed that the number of feature channels sampled by the CARS algorithm was significantly reduced from 2048 to 37,53,37 and 45 for Ni,Cu,As and Pb respectively,which is 1.81%~2.59% of the original number of channels.Compared with the no screening (i. e.original data) and successive projections algorithm (SPA),the accuracy of the CARS-1D-CNN model can reach 96.67%,93.22%,91.67% and 88.33%, respectively in determining whether the soil samples are at risk of contamination with Ni,Cu,As and Pb.Based on CARS screening,1D-CNN has a significant advantage over traditional partial least squares regression (PLSR) methods regarding predictive accuracy.Therefore,the CARS combined with the 1D-CNN method proposed in this paper improves the model prediction accuracy while reducing its computing complexity,which is a good theoretical guidance for soil heavy metal elemental contamination risk screening.