Optimizing machine learning models for tractor surface recognition: Performance and model size reduction for embedded systems

Abstract

This research applied surface recognition in autonomous tractors is essential for efficient navigation across diverse terrains of asphalt, gravel, and soil by using machine learning models and processing data acquisition from the BNO055 IMU sensor. In addition, this aims to evaluate and optimize various machine learning models, including Logistic Regression, K-Nearest Neighbors (KNN), SVC, Decision Tree, Random Forest, Gradient Boosting, AdaBoost, and XGBoost, for surface classification, with a focus on reducing model size without sacrificing classification accuracy. The research applies model pruning techniques to optimize these models for TinyML environments. The results demonstrate that XGBoost and Random Forest achieve high classification accuracy but have large model sizes, which can be mitigated through pruning, significantly reducing their size while maintaining performance. This study provides valuable insights into the trade-offs between model size and accuracy, contributing to the development of more efficient models for embedded systems in autonomous tractors. The findings highlight the potential of model pruning in enabling real-time surface recognition in resource-constrained environments, offering a scalable solution for deployment in agricultural machinery. The best results demonstrate 91 percent accuracy for XGBoost; Random Forest, on the other hand, has a very large trimmed model size (19,603 KB) and more efficient operations than other machine learning models for surface classification.