Compared with the traditional detection methods, hyperspectral technology has the characteristics of rapid, accurate and low cost in estimating soil heavy metal arsenic content and can dynamically monitor heavy metal arsenic pollution of oasis soils in arid regions. Based on the collection of soil samples from the cultivated layer of the delta oasis of Weigan-Kuqa river in Xinjiang, soil spectral data and heavy metal arsenic content were obtained. Through the four wavelet basis functions bior1.3, db4, gaus4 and mexh, the original spectral reflectance of the soil was subjected to continuous wavelet transformation. The transformed spectral data was correlated with the heavy metal arsenic so that the selected sensitive wavelet coefficients were taken as independent variables, using partial least square regression, support vector machine regression, BP neural network and random forest regression methods to perform hyperspectral inversion of heavy metal arsenic content. The results showed that: (1) The spectral decomposition effect of the four wavelet basis functions at scales 3 to 8 was obviously better than that of other scales, especially the continuous wavelet transform at scales 4 to 6, effectively improved the correlation between the spectral reflectance with soil heavy metal arsenic, and the number of wavelet coefficients passing the significance test increased significantly (p<0.01), and there had a strong correlation in the vicinity of 400~700 nm in visible light and 1 100~1 700 and 2 200~2 400 nm in near-infrared. (2) By comparing the ability of the four wavelet basis functions to identify effective information in the spectral data, it was believed that the wavelet basis functions bior1.3 and mesh were better than db4 and gaus4. Among them, bior1.3 had the best spectral decomposition effect, and gaus4 was relatively weak. Through the spectral transformation of the 5th scale of bior1.3, the number of bands significantly related to soil heavy metal arsenic was the largest, which was 507 (p<0.01). (3) Comparing the inversion results of the four modeling methods, it was found that the SVMR, BPNN and RFR models had stronger estimation capabilities than the PLSR model, and the estimation accuracy of the model was high. After comprehensively analyzing each model’s stability and estimation accuracy, it was concluded that the bior1.3-25-RFR model could be used as the best estimation model for the heavy metal arsenic in the study area. The R2 of the training set and the validation set of the model were 0.893 and 0.639 respectively, the RMSE were 1.075 and 1.651 mg·kg-1, and the RPD were 2.89 and 1.64 respectively, indicating that the model had a better estimation effect and powerful stability. Using appropriate wavelet basis functions to carry out continuous wavelet transform can reduce the white noise in hyperspectral soil data, excavate the effective information in soil spectral data, and provide a strong technical guarantee for accurate estimation of soil heavy metal arsenic content.