Feature Selection Random Forest Model for Crop Disease Detection in Greenhouse Vegetable Images

Abstract

The disease-detection system model for greenhouse vegetables must be very accurate, simple to understand, and compatible with a low-cost monitoring apparatus. The majority of deep learning techniques necessitate large-scale annotated images and provide less insight into the visual symptoms that influenced the classification choice. Make a selection of explainable features. This study used a Random Forest algorithm to detect crop diseases in photos of greenhouse vegetables. Using a method, extract lesion color, texture, form, boundary, and local contrast features from illness candidate locations. Then, use a relevance-redundancy-importance selection strategy to create a small set of symptom-sensitive features. The chosen features are used to train a Random Forest classifier, and category-level illness difference is interpreted using feature significance analysis. Experimentally evaluate the computational efficiency, illumination robustness, confusion among related diseases, feature reduction ratio, and detection accuracy. With a general accuracy of 94.3%, a macro-F1 score of 93.6%, a feature reduction rate of 61.8%, and an average inference speed of 6.4 ms per image, all of the metrics in the suggested model have produced excellent results. As a result, the selected symptom descriptors will perform well in categorization and offer more comprehensible features at the representation level.