Predicting soil organic matter (SOM) content based on images has the advantages of convenience and low cost. Interfered by objective factors such as soil type and moisture, there is still a gap between the detection accuracy of image prediction SOM content and traditional methods, which limits the promotion and popularization of related technologies. In order to improve the accuracy of image prediction of SOM content, a N_DenseNet multi-scale pooling module was added to DenseNet169 to improve the performance of the model by obtaining more dimensional features, and combined the development of SOM real-time detection APP on the Android side to realize the timely transmission of server and mobile phone data through intranet projection. Based on 350 soil samples from Youyi County, Heilongjiang Province, Changping District, Beijing City and Tai’an City, Shandong Province, high-definition images, R-band, red-edged band and near-infrared band images of in situ soil were obtained through mobile phones and multispectral drones to enrich data information, and image samples of soil samples under different moisture gradients were taken through indoor stress to alleviate the impact of moisture on the image. Compared with different deep learning models, the N_DenseNet trained based on multispectral image data performed the best, the overall performance was better than that of DenseNet169, the test set R2 was 0.833, RMSE was 3.943g/kg, and R2 was improved by 0.011 compared with the visible light data, which proved that the addition of R-band and red-edged and near-infrared images to the training process helped to improve the performance of the model, which proved the feasibility of the method. The mobile phone APP was connected to the background data to realize real-time data transmission, and realized the real-time detection of SOM content of soil samples in the field, and the model predicted R2 as 0.805 and the detection time was 2.8s after field verification, which met the needs of SOM content detection in the field and provided an idea for real-time detection of SOM content.