Comparative Experiments for Generation of Banana Black Sigatoka Disease Stages as Labels for Multiclass Classification Tasks

Abstract

Early and accurate diagnosis of Black Sigatoka (BSD), a fungal disease affecting banana production, is crucial for minimizing crop losses. An empirical understanding of the disease stages is important in developing predictors with adequate recommendations to farmers. This paper explores the use of K-means clustering algorithms to classify BSD stages in banana leaf images without relying on manual labelling. We evaluate the effectiveness of various image features, including “infected area”, “colour histograms”, “statistical features”, and “texture features”. The results indicate that using solely “infected area” achieved a moderate cluster separation, as observed from a Silhouette score of 0.5374 compared to others whose Silhouette scores were 0.3299 and 0.1366 for “statistical feature” and “colour histogram feature” respectively. Combining features, particularly, infected-area with texture or statistics, offered a promising balance between cluster separation and within-cluster variation as their Silhouette scores ranged between 0.32 and 0.47. Further investigation is needed to confirm the robustness of combining all features.
This research lays the groundwork for developing automated BSD classification systems to aid farmers in early disease detection and improved crop management. Automated classification using machine learning algorithms can significantly reduce the time and effort required for disease monitoring. Additionally, the integration of multiple image features could enhance the accuracy and reliability of the classification system. Future work will focus on validating these findings with larger datasets and exploring advanced machine learning techniques to further improve classification performance. Ultimately, the implementation of such systems could lead to better-informed decision-making in banana cultivation, reducing the impact of BSD on global banana production.