Real-time estimation of soil pulverization ratio after tillage using a camera mounted on agricultural machinery

• A real-time soil pulverization ratio estimation system was developed for tractor use. • The system presented both software and hardware configurations for practice. • The estimation accuracy was a root mean square error of 16.1% across three field types. • Cross-validation accuracy using a deep neural network model improved to 11.4%. • Generalization remains limited due to insufficient diversity in training data. Accurate measurement of the soil pulverization ratio is essential for automated tillage operations. We developed and implemented a real-time system for estimating the pulverization ratio using a camera mounted on agricultural machinery. Then, we verified the accuracy of the system under field conditions. First, we evaluated a conventional method that classifies soil clods based on brightness. The brightness threshold was calibrated to match the pulverization ratio, which was measured directly in a field. Real-time measurements were conducted in three regions of Japan: Akita, Saga, and Ibaraki. The estimation accuracy, expressed as the root mean square error (RMSE), was 10.3% for training data and 16.1% for testing data. Using the 131 data obtained from these experiments, we then evaluated the estimation performance of a convolutional neural network (CNN) model. The results were as follows: (1) When cross-validated by stratifying the data by region, the CNN achieved 14.5% RMSE with a 95% confidence interval (CI) of 11.8 ∼ 16.0, demonstrating improved performance through multi-region training. (2) When data were randomly divided for cross-validation, the CNN further improved to 11.4% RMSE with a 95% CI of 10.8 ∼ 11.8, indicating that a larger, more diverse dataset enhances prediction accuracy. Furthermore, the computational efficiency of the system was substantially improved. Using the CNN reduced the required image resolution from 1392 to just 100 pixels without compromising estimation performance.