Abstract: Soil organic matter (SOM) can enhance the soil structure and fertility in the nutrient cycling under carbon sequestration. However, conventional chemical analysis can suffer from some drawbacks, including the lengthy processing times, operational complexity, high costs, and the destruction of soil samples. It is often required for the rapid, and large-scale monitoring in modern precision agriculture. In this study, a portable device was developed for the accurate, rapid and non-destructive detection of SOM content in agricultural field environments using multispectral technology. The 237 soil samples were collected from Ningxia and Shaanxi Province in China. Visible-near infrared (Vis-NIR) spectral data was obtained for the reference values of the SOM content. Spectral preprocessing was implemented to remove the outliers. Four algorithms of the feature wavelength-namely moving window PLS (MWPLS), iterative random forests (iRF), variable dimension particle swarm optimization using combined moving window (VDPSO-CMW), and moving window smoothing on the ensemble of competitive adaptive reweighted sampling (MWS-ECARS)-were employed to conjunct with the variable importance in Projection (VIP). Seven characteristic wavelengths of the SOM were successfully identified: 420, 530, 600, 630, 855, 900, and 1 345 nm. Thereby the spectral dimensionality and complexity of the dataset were significantly reduced after optimization. Furthermore, a correction wavelength at 1 450 nm was also selected for the soil moisture adjustment, in order to avoid the interference of the soil moisture on the spectral signals. The hardware and software systems of the device were designed, according to the characteristic wavelengths and the ESP32 embedded platform. Four modules were integrated in the device, including a main control, a multispectral acquisition, a power management and a human-machine interaction. Multispectral data was acquired using narrow-band LED light sources. The characteristic wavelengths were integrated with the photodiodes. Subsequently, the device was also used to collect the soil multispectral data. The dry soil, global and stratified modelling were constructed using three machine learning algorithms-partial least squares (PLS), multilayer perceptron (MLP), and support vector machine (SVM). The results indicated that the stratified modelling was yielded the best performance. The independent SOM prediction models were established for each moisture gradient. The soil moisture content was firstly estimated during actual prediction. The moisture gradient was determined as the sub-model to predict the SOM content. Thereby the accurate SOM estimation was realized under the varying soil moisture conditions. A PLS-MLP stratified regression model was constructed to acquire multispectral data using the prototype device. The PLS algorithm was used for the moisture gradient classification, and then applied the MLP algorithm to estimate the SOM content. The better performance of the model was achieved in a coefficient of determination (R²) of 0.84, a root mean square error (RMSE) of 3.93 g/kg, and a relative prediction deviation (RPD) of 2.50. The model was embedded into the device for testing. The correlation coefficient between the estimated and laboratory-measured values reached 0.82, with an RMSE of 5.27 g/kg, an RPD of 1.64, a standard deviation for the repeated tests of less than 0.4 g/kg, and a single detection time under 9 s. Therefore, the rapid and accurate estimation of the SOM content can provide the strong application potential to the rapid assessment of the soil fertility.