Abstract: Tires are one of the most important components to carry and drive agricultural machinery, such as tractors. At the same time, the tires can also contact with the ground and continuously interact during the operation of agricultural vehicles. The working state of tires can directly dominate the various performance of agricultural machinery. Agricultural tires are often confined to large load fluctuations, special pattern shapes, harsh working environments, and tire body vibrations. However, it is also difficult to accurately acquire the tire vertical loads in practical operations. Particularly, the vertical load can have a serious impact on the performance of agricultural machinery. The key influencing factors of the vertical load can be used to evaluate and optimize the efficiency and stability during operation. In this study, a systematic estimation of the vertical load was performed on the off-road tires using a bidirectional strain of wheel rims. The vertical loads of the agricultural tires were obtained for the high estimation accuracy of the improved models. Taking the off-road tire as the research object, the radial and tangential strains of the wheel rim were determined under vertical load. A wheel rim strain acquisition system was designed using high-precision pin axis sensors. A test bench of tire drum was used to conduct the multiple working tests. The stress-strain curve of the wheel rim was obtained during tire rolling. A series of operations were conducted on the data denoising, period partitioning, data filtering, and period feature extraction on the strain curve. A dataset was then established for the tangential and radial strains of the wheel rim during tire rolling. An estimation model of the tire vertical load was constructed using a self-attention convolutional neural network (AT-CNN). The periodic tangential and radial strain curves were also taken as the inputs. The model consisted of two single-load prediction networks and a coupled network. Among them, both outputs were predicted by the vertical loads. The results show that the mean absolute error (MAE) of the AT-CNN estimation model was 95.68 N, and the root mean square error (RMSE) was 100.54 N. The MAE of the deep neural network (DNN) model was 557.35 N, and the RMSE was 588.07 N. As such, the MAE and RMSE of the AT-CNN were reduced by 82.83% and 82.90%, respectively, compared with the DNN network. The t-distributed stochastic neighbor embedding (t-SNE) was applied to reduce the dimensionality of features in the network. The feature of the load prediction network was extracted and then visualized after optimization. The visualization results show that there were closer similar features with the same strain after the input data was processed by the AT-CNN. There were more outstanding features after classification. A 10-fold cross-test was carried out to verify the generalization of the AT-CNN. The average MAE of DNN was 647.32 N, and the average RMSE was 650.16 N in ten experiments. While the average MAE of the AT-CNN was 98.73 N, and the average RMSE was 102.33 N. Therefore, the AT-CNN model shared a higher prediction accuracy, compared with the DNN model. There were the more concentrated distributions of the MAE and RMSE. The AT-CNN model shared higher prediction accuracy and generalization. The self-attention and feature fusion modules were verified to significantly improve the estimation accuracy of the network and the overall performance of the AT-CNN using ablation experiments. Real-time and accurate estimation was achieved in the vertical load of the off-road tires using a bidirectional strain of the wheel rim.