Knowledge correction and ε-insensitive criterion-leveraged zero-order TSK fuzzy system for rice leaf disease diagnosis

Abstract

The advent of complex application scenarios introduces new challenges for diagnosing rice leaf diseases using machine learning methods. Two critical requirements are identified: 1) The model must exhibit high interpretability to mitigate the adverse effects of incorrect diagnoses; and 2) practical applications often suffer from insufficient samples and noise in rice leaf disease datasets, which requires the model to have strong generalization ability and robustness. However, existing methods still have certain limitations in practical scenarios due to a lack of comprehensive consideration of interpretability, generalization ability, and robustness. To address this issue, this article proposes a novel knowledge correction and $\epsilon$-insensitive criterion-leveraged zero-order TSK fuzzy system (0-TSK-FS), named KE-0-TSK-FS. The KE-0-TSK-FS method is developed with 0-TSK-FS as the baseline, enhancing the generalization ability of the model by introducing the knowledge correction method and its iterative learning strategy to extract more information from limited samples. In addition, the objective function based on the $\epsilon$-insensitive criterion makes KE-0-TSK-FS exhibit robustness when the samples contain noise. On three rice leaf disease datasets and six real-world non-rice leaf disease datasets, experiments were conducted on three metrics, namely accuracy, GM, and rule complexity. The experimental results show that the KE-0-TSK-FS method outperforms other comparative algorithms in terms of generalization ability, interpretability, and robustness in the diagnosis of rice leaf diseases under insufficient samples and noise situations, and its average accuracy on rice leaf disease datasets is nearly 3% higher than that of other comparative algorithms.