Field in-situ visible-near infrared spectroscopy (VIS-NIR) can effectively improve the detection efficiency of soil properties, but due to the influence of in-situ soil moisture factors, the prediction accuracy of soil property models is difficult to reach expectations. How to effectively remove the influence of soil moisture on the spectrum prediction of other soil properties is a difficult problem for using field in-situ spectroscopy to predict soil properties with high precision. It is also a breakthrough for soil spectroscopy technology to shift from indoor to field. The effective solution to this problem can eliminate the process of soil sample collection and indoor pretreatment, and achieve field in-situ spectroscopy of soil properties. In this study, a total of 115 surface soil samples of 0~20 cm were collected by using the grid sampling method in the cotton field of the 12th group in the Alar reclamation area in southern Xinjiang. And use the SR-3500 portable ground object spectrometer to collect 231 sample points of field in-situ spectral data. The soil samples are air-dried, ground and sieved, and then their indoor spectrum and organic matter content are measured. 115 soil samples were divided into two subsets, i.e. the conversion subset (69 samples) and the prediction set (46 samples), with the Kennard-Stone algorithm. The external parameter orthogonalization (EPO), spectral direct standardization (DS) and piece-wise spectral direct standardization (PDS) were performed on spectra and the spectra with other three pretreatments including first-order differential of reflectance (R'), logarithm of reflectance (LOG(R)) and inverse reflectance (1/R), respectively. The random forest (RF) model is used to construct different combination models and evaluate the accuracy. The results showed: (1) The higher the soil organic matter content, the lower the soil spectral reflectance. The in-situ spectral reflectance of soil in the field is lower than the indoor spectral reflectance of soil; (2) The correlation between indoor spectral reflectance and soil organic matter content is greater than that of field in-situ spectra. The correlation between indoor spectra and soil organic matter content is significantly improved after first-order differential transformation. (3) The prediction accuracy of the soil indoor spectral reflectance model (R²=0.86, RPD=2.08, RMSE=1.55 g·kg^-1, MAPE=0.14) is higher than the field in-situ spectral reflectance model (R²=0.71, RPD=1.49, RMSE=2.17 g·kg^-1, MAPE=0.20). Among the moisture removal algorithm models, the EPO first-order differential model has the best moisture removal effect, with the coefficient of determination R² increased from 0.71 to 0.83, RPD increased from 1.49 to 2.04, RMSE decreased from 2.17 to 1.58 g·kg^-1, and MAPE decreased from 0.20 to 0.14. This study achieved the removal of the influence of soil moisture factors, improved the accuracy of field in-situ spectroscopy prediction of soil organic matter, and provided important reference value for the prediction of large-scale soil organic matter and the evaluation of soil fertility in southern Xinjiang cotton fields.