Dynamic Covariance Estimation in EKF via Deep Learning for Agricultural Vehicle Localization

Accurate full-state estimation of a rigid body is essential in many navigation and control tasks, but may be challenging with low-cost sensors. This work addresses the problem using only Inertial Measurement Unit (IMU) and Global Positioning System (GPS) data. We propose a hybrid architecture based on the Extended Kalman Filter (EKF), enhanced by a neural network that adaptively estimates process and measurement noise covariances. This Neural Enhanced Extended Kalman Filter (NEEKF) improves robustness under time-varying and nonlinear noise conditions. Standard EKFs assume fixed covariances (Q, R), limiting adaptability in real-world scenarios - e.g., autonomous tractors on irregular terrain. To overcome this limitation, we train neural models to estimate Qk and Rk from internal filter signals such as innovations and uncertainty metrics. The approach is validated in a high-fidelity Unity simulation replicating realistic disturbances including rain, dropouts, and soil-induced vibrations. Results show that adaptive noise modeling significantly enhances estimation accuracy. Among the tested architectures, feedforward networks showed the best trade-off between performance and simplicity, while Gated Recurrent Units (GRUs) and Transformers offered improvements in dynamic and temporally correlated conditions.