To address the challenges posed by highly repetitive structures, unstructured terrain, and pronounced sensor disturbances in agricultural environments, this paper proposes a simultaneous localization and mapping framework that integrates explicit spatial feature modeling with time-varying noise estimation. First, a curved-voxel-based intensity–geometry joint probability model is constructed to transform conventional point features into spatial features capable of effectively capturing local structural patterns. These features are then aligned by minimizing the joint probability distance using the normal distributions transform, thereby enhancing feature discriminability and registration stability in highly repetitive scenes. Second, a maximum a posteriori-based recursive noise estimator is designed. By employing a dual sub-filter architecture, the proposed estimator enables joint modeling and online optimization of the noise parameters associated with LiDAR odometry and the inertial measurement unit, thereby improving the system’s adaptability to terrain-induced perturbations and sensor uncertainties. Experimental results demonstrate that the proposed method achieves a mean relative translation error of 1.42% and a mean relative rotation error of 1.53°/100 m in agricultural scenarios. In addition, the average error in row spacing estimation derived from the reconstructed map is constrained within 0.071 m. Compared with baseline methods, the proposed system exhibits significant advantages in both pose estimation accuracy and agronomic structural perception capability.