SPINACHAI: Multiclass Spinach Leaf Disease Classification and Explainable AI Using Efficientnetb0 and Grad-CAM

Abstract

Spinach (Spinacia oleracea L.) is one of the major globally grown leafy vegetables and is highly susceptible to several fungal diseases, most no-tably Downy Mildew (Peronospora effusa, previously classified as P. fari-nosa f. sp. spinaciae). Prompt and accurate identification of the disease can contribute significantly to minimising yield losses and can serve preci-sion agriculture applications. We propose an automated multiclass spinach leaf disease classification framework, SpinachAI, based on transfer learn-ing with EfficientNetB0 and Grad-CAM explainable AI (XAI). The system classifies spinach leaf images into three classes, namely Healthy, Downy Mildew, and Invalid input. We build and prepare a spinach leaf image dataset comprising 458 labelled images acquired from agricultural image repositories and field-collected samples. To address the small size of the training dataset and improve generalisation, image augmentation and two-stage transfer learning are incorporated. The experiments show good clas-sification performance, with an overall accuracy of 94.98%, a weighted F1-score of 95.20%, a Matthews Correlation Coefficient (MCC) of 0.9132, and an AUC-ROC score of 0.9890. Grad-CAM visualisation is shown to in-crease the interpretability of the model by highlighting the regions on which the model relies, thereby increasing user confidence in the model output. The complete framework is deployed as a real-time Streamlit application that can be used for agricultural crop diagnosis and smart agriculture.