Abstract: The complex terrain, vibrations, and other factors in off-road environments can cause sensor degradation and significantly affect the performance of visual-inertial odometry (VIO). To address this issue, we propose an adaptive fusion algorithm combining the global navigation satellite system (GNSS) and VIO based on a multidimensional scoring model. First, during the initialization, we construct an extended state variable window to estimate the coordinate transformation and apply the interpolation algorithm to reducing the impact of time synchronization errors. Second, we develop a scoring model incorporating the motion, rotation, and consistency dimensions to assess the reliability of the sensor data in real time. Based on the scoring results, we adaptively adjust the noise covariance matrix of the sensors to optimize the data fusion effect and mitigate the effect of sensor degradation. Finally, we conduct experiments in an agricultural scenario and three different campus scenarios. The results reveal that the proposed algorithm effectively addresses sensor degradation and significantly improves the positioning accuracy of the VIO system (with an average improvement of 79%). Compared with GSI-SLAM, a tightly coupled GNSS-VIO algorithm designed for arable farming, the proposed algorithm improves the positioning accuracy by 37%.