An empirical study of CNN models and XAI for grape leaf disease detection based on data split ratios

Abstract

The efficiency of deep learning (DL) models in detecting plant diseases is frequently evaluated on controlled datasets, with little attention paid to data-partitioning techniques that impact generalization in real-world scenarios. This study examines the effects of three data split ratios (80:10:10, 70:15:15, and 50:25:25) on the generalizability of four convolutional neural network (CNN) architectures, DenseNet121, ResNet50, InceptionV3, and VGG16, for grape leaf disease classification across four classes of Black Rot, Esca (Black Measles), Leaf Blight, and Healthy. In addition to the standard performance metrics, we include external validation using an independent Kaggle dataset (9,027 images) to assess true generalizability, complementing the primary PlantVillage dataset (3,251 images). Results indicate that split ratio significantly affects model performance, with DenseNet121 showing outstanding stability across all splits and superior external test accuracy (up to 99.6%), while InceptionV3 demonstrates extreme instability, and ResNet50 shows noticeable variation. Explainable artificial intelligence (XAI) via gradient-weighted class activation mapping (Grad-CAM) reveals that DenseNet121 and ResNet50 focus on pathologically significant leaf regions rather than spurious background correlations. This study offers a framework for assessing DL models in agriculture, highlighting that no single split ratio universally optimizes performance and emphasizing external validation for deploying trustworthy solutions.