Leveraging Lesion Segmentation Masks to Validate CNN Focus in Apple Disease Classification using Explainable AI

Abstract

This paper presents a procedure to explain the focus of the Convolutional Neural Networks (CNNs) for classifying apple diseases. The goal of this work is to promote more transparency and trust in CNN-based diagnostic tools by using Explainable AI (XAI) methods -here Grad-CAM (Gradient-weighted Class Activation Mapping) in the agricultural setting. The main concept in the proposed pipeline is to use apple leaf images as well as the manually created lesion segmentation masks. A pre-trained CNN is used for disease classification, where the last two-weighted layers are employed to extract the significantly enriched features, and then Grad-CAM is used to output the heatmap to highlight the informative parts for the decision. One of the main contributions of this study is to quantitatively compare these Grad-CAM heatmaps with the ground truth labels (lesion masks) in terms of Intersection over Union (IoU) score. This test gives us a way to quantitatively evaluate if the CNN is learning from real disease symptoms. By making decisions about the model dependent on pathological features, this approach intends to provide the application of CNNs for apple disease classification with much valuable confidence and reliability.

Plant phenotyping relevance

リンゴ葉の病徴（病斑）を画像から扱い、Grad-CAMの病斑適合性をIoUで定量検証する手法研究であり、疾病状態の推定・検証が中心である。

Code and data availability

The paper describes apple leaf images, manually created lesion segmentation masks, and Python analysis code (Grad-CAM/IoU pipeline), but no block contains any public dataset deposit, repository URL, or code availability statement. The data appears to be a small sample/synthetic demonstration, and no authors' public URL

No evidence-backed public reproduction asset is currently recorded.