Explainable AI-Enhanced Deep Learning for Pumpkin Leaf Disease Detection: A Comparative Analysis of CNN Architectures

Md. Arafat Alam Khandaker 1 , Ziyan Shirin Raha 1 , Shifat Islam 2 , Tashreef Muhammad 2 · 2025

Read Explainable AI-Enhanced Deep Learning for Pumpkin Leaf Disease Detection: A Comparative Analysis of CNN Architectures by Md. Arafat Alam Khandaker 1 , Ziyan Shirin Raha 1 , Shifat Islam 2 , Tashreef Muhammad 2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-05449.txt.

Introduction

Pumpkin leaves announce their diseases in spots and mildew, but spotting them in the field is slow and error-prone. Khandaker and co-authors trained image networks to name five states from photographs: downy mildew, powdery mildew, mosaic disease, bacterial leaf spot, and healthy. Their dataset held two thousand high-resolution images gathered from several farms so the models would meet real variety.

Seven established network designs were compared, and ResNet50 won, reaching 90.5 percent accuracy with matching precision, recall, and F1 scores. To keep the system honest, the team applied explainable AI tools that paint heat maps over each leaf, showing which regions drove the call. The authors argue that an accurate model whose reasoning a farmer can inspect could bring earlier, surer treatment to pumpkin fields.

Catalog record

CREATORMd. Arafat Alam Khandaker 1 , Ziyan Shirin Raha 1 , Shifat Islam 2 , Tashreef Muhammad 2
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-05449

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Explainable AI-Enhanced Deep Learning for Pumpkin Leaf Disease Detection: A Comparative Analysis of CNN Architectures Md. Arafat Alam Khandaker 1 , Ziyan Shirin Raha 1 , Shifat Islam 2 , Tashreef Muhammad 2 Abstract Pumpkin leaf diseases are significant threats to agricultural productivity, requiring a timely and precise diagnosis for effective management. Traditional identification methods are laborious and susceptible to human error, emphasizing the necessity for automated solutions. This study employs on the “Pumpkin Leaf Disease Dataset”, that comprises of 2,000 high-resolution images separated into five categories. Downy mildew, powdery mildew, mosaic disease, bacterial leaf spot, and healthy leaves. The dataset was rigorously assembled from several agricultural fields to ensure a strong representation for model training. We explored many proficient deep learning architectures, including DenseNet201, DenseNet121, DenseNet169, Xception, ResNet50, ResNet101 and InceptionResNetV2, and observed that ResNet50 performed most effectively, with an accuracy of 90.5% and comparable precision, recall, and F1-Score. We used Explainable AI (XAI) approaches like Grad-CAM, Grad-CAM++, Score-CAM, and Layer-CAM to provide meaningful representations of model decision-making processes, which improved understanding and trust in automated disease diagnostics. These findings demonstrate ResNet50’s potential to revolutionize pumpkin leaf disease detection, allowing for earlier and more accurate treatments. [email protected], [email protected], [email protected], [email protected] I Introduction Pumpkin leaf diseases including downy mildew, powdery mildew, mosaic disease, and bacterial leaf spot are significant threats to the production of agriculture, culminating with substantial crop losses. These diseases not only diminish yields but also have an impact on crop quality, making effective disease control vital for farmers. Early and precise detection of these diseases is critical for reducing their impact, but traditional procedures based on manual inspection are frequently time-consuming, labour-intensive, and prone to human error. As a result, these problems highlight the importance of efficient, automated, and accurate diagnostic approaches that can be used at scale in agricultural contexts. Deep learning has transformed areas like healthcare and automotive by automating complicated operations [ 1 ] , but its applicability in agriculture is still underexplored. One of the most significant challenges to its widespread adoption in agriculture is a lack of transparency in how AI models make verdicts, which is critical for creating assurance and ensuring practical utility in real-world circumstances. Explainable AI (XAI) [ 2 ] provides an appealing approach for making AI models more transparent and intelligible to users. XAI approaches such as visual heatmaps assist in emphasizing the essential elements of images that influence a model’s prediction, providing insights into decision-making processes. This is especially crucial in agriculture, where openness and confidence in AI-driven decision-making may help farmers and agricultural experts make sound decisions based on trustworthy AI insights. The inclusion of AI increases its potential for clinical diagnosis and research [ 3 ] . This formidable combination has the potential to broaden its applications across a variety of fields, including potato disease detection [ 4 ] , detection of plant leaf disease [ 5 ] , plant disease recognition [ 6 ] and automatic detection of fruit disease [ 7 ] . Standard convolutional neural networks (CNN) models, that include DenseNet [ 8 ] , ResNet [ 14 ] , Xception [ 15 ] , and InceptionResNetV2 [ 16 ] , perform well in detecting plant diseases, but have poor interpretability. This study aims to address this issue by using cutting-edge XAI approaches such as Grad-CAM [ 17 ] , Grad-CAM++ [ 18 ] , Score-CAM [ 19 ] , and Layer-CAM [ 20 ] to produce visual explanations of model predictions. These techniques not only improve transparency but also make agricultural applications more practical by highlighting the areas of leaf images that are most relevant to disease classification. The research paper’s contributions are summarized below: • We evaluated seven pre-trained convolutional neural network (CNN) models (ResNet50, ResNet101, DenseNet121, DenseNet169, DenseNet201, Xception) and InceptionResNetV2) for automated pumpkin leaf classification. • Explainable AI techniques including GradCam, GradCam++, ScoreCam, and LayerCam are being utilized to improve transparency. • This research expands the gap in Explainable Artificial Intelligence (XAI) technique for the pumpkin classification challenge by being the first to use XAI techniques in this setting and providing novel insights into model interpretability. Our study improves the field of AI in agriculture by incorporating XAI into the diagnostic procedure, which leads to high-performing and interpretable models that are essential to effective disease management decision-making. In this study, we have already explained the introduction in Section I . A thorough related works review will follow it in Section II . Then in different subsections, we discuss the conducted work’s methodology in Section III . Then we describe our results in Section IV . Afterwards that, we are presented our limitations and future work in section V followed by a conclusion in Section VI to summarize our conducted study. II Related Works We have reviewed several studies on pumpkin disease detection. However, none of them have incorporated Explainable AI (XAI) techniques into their methodologies. This section offers a concise review of the existing research consisting of Machine Learning and Deep Learning Models. It highlights the relevant methods and their respective contributions to the field. Mosaddek Ali et al. [ 3 ] demonstrated the effectiveness of traditional machine learning models, such as SVM, KNN, Decision Tree, and Naive Bayes, with CNNs for detecting pumpkin leaf diseases. in agriculture, employing a dataset of 2000 pumpkin leaf pictures to reach a high validation accuracy of 87.67%. This work demonstrates CNN’s capacity to extract significant features from picture data, which is critical for accurate disease identification. Moreover, Yousef Methkal et al. [ 11 ] developed an ACO-CNN model that improves classic CNN architecture with Ant Colony Optimization. This approach not only obtained an outstanding 99.98% accuracy, but also maintained good precision, recall, and F1-scores, demonstrating the advantages of hybrid models in complicated classification problems. Furthermore, Naeem et al. [ 6 ] developed DeepPlantNet, a 28-layer deep learning network designed for plant disease classification. This model obtained accuracies of 98.49% and 99.85% for its eight- and three-class classification tasks, respectively, demonstrating advances in deep learning for agriculture. highlighting the adaptability of CNNs. However, Asad et al. [ 7 ] utilized a CNN model to identify and categorize common citrus diseases, achieving a test accuracy of 94.55%. This indicates CNNs’ broad usefulness, not just in leaf disease detection, but also in citrus disease management, which presents distinct obstacles. Besides, Sk Mahmudul et al. [ 9 ] investigated different CNN architectures on the PlantVillage dataset and found that EfficientNetB0 was the most efficient. This demonstrates the efficacy of advanced CNN models in plant disease detection and how architectural differences might affect detection accuracy. Additionally, Eman Abdullah et al. [ 5 ] leveraged the PlantVillage dataset to implement the YOLOv4 algorithm for real-time disease identification and classification. The model’s excellent accuracy of 99.99% highlights its versatility in operating contexts where speed and accuracy are critical. Besides, Mobeen et al. [ 10 ] proposed an efficient CNN-based technique adapted for mobile devices, utilizing MobileNetV3Large. The study focused on class imbalance and resource constraints, demonstrating that stepwise transfer learning can successfully improve model performance while remaining resource-economical. Whereas, Touhidul et al. [ 13 ] compared the performance and computational efficiency of a bespoke CNN, LDDTA, to nine

Read the complete text