Soil organic matter (SOM) is one of the key chemical properties for evaluating soil fertility. Traditionally, it was measured by wet chemistry analyses, which are time-consuming, expensive, and require complicated sample treatment procedures. Currently, a variety of agricultural sensors have been applied to determine soil properties rapidly. Many studies have been conducted on the use of single sensor (vis-NIR, mid-IR and LIBS) to evaluate soil attributes. However, sometimes the prediction of soil properties with single sensors is less robust for soil samples from different sampling sites due to the complex nature of soils. This study integrates visible near infrared (vis-NIR) spectroscopy, mid infrared (mid-IR) spectroscopy and laser-induced breakdown spectroscopy (LIBS) to achieve rapid measurement of SOM. In order to reduce data redundancy, we selected characteristic bands firstly by using genetic algorithm-partial least-squares regression (GA-PLSR) and uninformative variable elimination (UVE) algorithm. We compared the two characteristic bands selection methods and found the better method for the following steps. And then, we built the prediction models from three aspects: predictions using partial least-squares regression (PLSR) based on single sensor data; predictions using PLSR based on outer-product analysis (OPA), which means combining the different sensors data by OPA to a new dataset and build model; predictions using Bayesian Model Averaging (BMA) based on sensors prediction results fusion, which means combining the different sensors prediction results as a new dataset to build model. And the aims of this paper are to: 1) Compare the prediction ability of the three proximal soil sensors, i.e. vis-NIR spectrometers, mid-IR spectrometers and LIBS. 2) Study the prediction ability of multi-sensors fusion from two aspects: sensors data fusion and sensors prediction results fusion. 3) Compare the prediction ability of single sensor, sensors data fusion and sensors prediction results fusion.