Given that Sentinel-2 (S2) multispectral images provide extensive spatial information and that ground-based hyperspectral data capture refined spectral characteristics, their integration can enhance both the comprehensiveness and precision of surface information acquisition. This study seeks to leverage these data sources to develop an optimized estimation model for accurately monitoring large-scale soil organic carbon (SOC) content, thereby addressing current limitations in multi-source data fusion research. In this study, using mathematical transformation and discrete wavelet transform to process the ground hyperspectral data in the delta oasis of the Weigan and Kuqa rivers in Xinjiang, China, in combination with the S2 multispectral image, machine learning algorithms were employed to construct estimation models of SOC content for total variables and characteristic variables, and spatial inversion of SOC content in the oases was carried out. We found that the spectral transformation of R-DWT-H9 can significantly enhance the correlation between spectral data and SOC content (p < 0.001). The estimation accuracy of the models constructed based on the feature variables selected by SPA and IRIV was generally higher than that of the total variable models. The IRIV-RFR model had the highest estimation accuracy and stable estimation capability. The values of R-2 for the training and validation sets were 0.66 and 0.64, respectively. The RMSE values were < 1.5 g center dot kg(-1), and the values of RPD were > 1.4. In the interior of the oasis, the SOC content was mainly deficient (61.35%) or relatively deficient (8.17%), while on the periphery of the oasis, it was extremely deficient (30.48%). Combine ground hyperspectral data and S2 images to construct an inversion model for SOC content, thereby providing a reference for accurately evaluating soil fertility in arid oasis regions.