Accurate prediction of land leveller operating performance and dynamic updating of a virtual terrain map based on multimodal data fusion

For the closed-loop operation of intelligent land levellers involving simultaneous operation, updating, and replanning, this study aimed to achieve stable prediction of levelling performance under complex working conditions, such as load fluctuations and actuator lag, and to use these predictions for the dynamic updating of a virtual terrain map, thereby reducing terrain-perception drift and planning failures caused by cumulative errors. To this end, levelling performance was characterised by three indicators, namely levelling rate, surface flatness, and maximum elevation difference, while input features were constructed from process information including blade height, soil load, and operating speed. To mitigate time-varying deviations caused by hydraulic actuation lag and nonlinearity, a Fuzzy PID adaptive compensation module was incorporated into the prediction pipeline. In addition, an improved particle swarm optimisation (IPSO) strategy was introduced to optimise the parameters of the support vector regression (SVR) model, thereby improving prediction stability and generalisation. On this basis, a prediction-driven incremental virtual terrain map updating method was proposed, and the consistency between virtual updating and actual terrain improvement was evaluated using global indicators, local residual distribution, and change trends. The results showed that the model achieved R 2 values of 0.983, 0.979, and 0.973, RMSEs of 4.26%, 4.74 cm, and 5.28 cm, and MAEs of 3.76%, 4.27 cm, and 4.82 cm, respectively, outperforming conventional machine-learning algorithms. Compared with the IPSO-SVR model, the R 2 for levelling rate increased by 30.89%, while the RMSE and MAE decreased by 73.62% and 75.46%, respectively. Furthermore, the ablation results showed that adaptive IPSO optimisation reduced dependence on empirical parameter settings and lowered the risk of local optima, while Fuzzy PID compensation alleviated actuation-chain deviations and was crucial to long-term reliability. The model also demonstrated good predictive capability and robustness across different path segments. Overall, this study provides a feasible route for improving the long-term stability and planning reliability of intelligent land-levelling operations.