Non-road tires have typical characteristics such as large structural size, harsh and changeable working conditions, and obvious load fluctuations. Its vertical load significantly affects the longitudinal, vertical and lateral dynamic characteristics of the vehicle. Aiming at the problem of difficulty in obtaining the vertical load of non-road tires and the insufficient accuracy of traditional physical model deductions, a vertical load inversion algorithm was proposed based on strain information and machine learning technology. Taking the R-1 herringbone pattern non-road tire as the research object, a tire strain information collection system consisting of a large-range flexible strain sensor, low-power data collection and wireless transmission module was designed. Using parameters such as tire pressure, speed, load as variables, a variety of typical working condition tests were carried out on the drum test bench, and the strain change pattern of the tire contact point was analyzed. On this basis, a deep neural network model for tire-oriented vertical load estimates was built. The algorithm parameter optimization was carried out based on the AdamW optimizer and grid search method. The test results showed that the deep neural network model based on AdamW optimizer showed a high accuracy on the prediction of the non-road tire vertical load prediction. Under the trial conditions, the maximum average relative error was reduced from 4.10% to 0.30%. Test results for the generalization capacity of models showed that the average naturalization of deep neural network models was reduced by 55.91% compared with the SVR model, and the generalization performance was superior. Studies showed that the deep neural network model proposed based on the AdamW optimizer had accurate reaction to the non-road tire vertical load. This method provided the basis for the acquisition of reliable key parameters of tire mechanics for the dynamic control system of non-road vehicles.