Soybean Disease Detection via Interpretable Hybrid CNN-GNN: Integrating MobileNetV2 and GraphSAGE with Cross-Modal Attention

MD ABRAR JAHIN 1,2, SHAHRIAR SOUDEEP 3, M. F. MRIDHA 3, MD. JAKIR HOSSEN 4, NILANJAN DEY 5 · 2025

Read Soybean Disease Detection via Interpretable Hybrid CNN-GNN: Integrating MobileNetV2 and GraphSAGE with Cross-Modal Attention by MD ABRAR JAHIN 1,2, SHAHRIAR SOUDEEP 3, M. F. MRIDHA 3, MD. JAKIR HOSSEN 4, NILANJAN DEY 5 (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-2503-01284.txt.

Introduction

Md Abrar Jahin and co-authors build a system to spot soybean leaf diseases in work from 2025. Many leaf diseases look alike, so photo tools often confuse them. Farmers need a tool that also shows why it chose a label.

The system pairs two networks in sequence. A compact image network pulls local signs of damage from each leaf photo. A graph network then links like photos to weigh shared patterns. Heat maps mark the leaf areas that drove each call, so users can check the logic. On a set of ten diseases, the system scored near 97 percent correct. The clear visuals may help agronomists trust and debug the tool.

Catalog record

CREATORMD ABRAR JAHIN 1,2, SHAHRIAR SOUDEEP 3, M. F. MRIDHA 3, MD. JAKIR HOSSEN 4, NILANJAN DEY 5
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-01284

Browse the full library · About the Tower of Babel

Full text

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

Soybean Disease Detection via Interpretable Hybrid CNN-GNN: Integrating MobileNetV2 and GraphSAGE with Cross-Modal Attention MD ABRAR JAHIN 1,2, SHAHRIAR SOUDEEP 3, M. F. MRIDHA 3, MD. JAKIR HOSSEN 4, NILANJAN DEY 5 Abstract Soybean leaf disease detection is critical for agricultural productivity but faces challenges due to visually similar symptoms and limited interpretability of conventional methods. While convolutional neural networks (CNNs) excel in spatial feature extraction, they often neglect inter-image relational dependencies, leading to misclassifications. This paper proposes an interpretable hybrid sequential CNN-Graph Neural Network (GNN) framework that synergizes MobileNetV2 for localized feature extraction and GraphSAGE for relational modeling. The framework constructs a graph where nodes represent leaf images, with edges defined by cosine similarity-based adjacency matrices and adaptive neighborhood sampling. This design captures fine-grained lesion features and global symptom patterns, addressing inter-class similarity challenges. Cross-modal interpretability is achieved via Grad-CAM and Eigen-CAM visualizations, generating heatmaps to highlight disease-influential regions. Evaluated on a dataset of ten soybean leaf diseases, the model achieves 97.16% accuracy, surpassing standalone CNNs ( ≤ \leq 95.04%) and traditional machine learning models ( ≤ \leq 77.05%). Ablation studies validate the superiority of the sequential architecture over parallel or single-model configurations. With only 2.3 million parameters, the lightweight MobileNetV2-GraphSAGE combination ensures computational efficiency, enabling real-time deployment in resource-constrained environments. The proposed approach bridges the gap between accurate classification and practical applicability, offering a robust, interpretable tool for agricultural diagnostics while advancing CNN-GNN integration in plant pathology research. † † address: Physics and Biology Unit, Okinawa Institute of Science and Technology Graduate University, Okinawa 904-0412, Japan (e-mail: [email protected]) † † address: Thomas Lord Department of Computer Science, Viterbi School of Engineering, University of Southern California, Los Angeles, CA 90007, USA (e-mail: [email protected]) † † address: Department of Computer Science, American International University-Bangladesh, Dhaka 1229, Bangladesh (e-mail: [email protected], [email protected]) † † address: Department of Robotics and Automation, Multimedia University, Melaka, Malaysia (e-mail: [email protected]) † † address: Department of Computer Science & Engineering, Techno International New Town, Kolkata 700156, India (e-mail: [email protected]) † † corresponding: Corresponding author: Md. Jakir Hossen(e-mail:[email protected]) and M. F. Mridha (e-mail: [email protected]). I Introduction Soybean ( Glycine max ) is one of the most significant crops worldwide, providing essential nutrients and oil for both human consumption and animal feed. However, various diseases, including soybean rust, Septoria brown spot, and frog eye leaf spot, often threaten its production. These diseases severely affect the quality and yield of soybean crops, leading to substantial economic losses for farmers. Traditional methods of disease detection, which are primarily based on manual inspection, are time-consuming, labor-intensive, and subjective, making them unsuitable for large-scale, automated applications. With the advent of machine learning and deep learning, significant progress has been made in automating plant disease detection, particularly through convolutional neural networks (CNNs). CNNs have demonstrated strong performance in image classification tasks by automatically learning spatial features from raw images, eliminating the need for manual feature extraction. Numerous studies have highlighted their effectiveness in soybean leaf disease classification [ 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 ] . However, most existing approaches — whether CNNs or transfer learning techniques are used [ 10 , 11 ] — focus on extracting features from individual images, overlooking critical relational information between images. This becomes particularly problematic when diseases present visually similar symptoms triggered by different factors, such as nutrient deficiencies, pest damage, or environmental stress, often leading to misclassifications. Moreover, these conventional models offer limited explainability, providing little insight into which leaf regions drive predictions and reducing interpretability and trust among agricultural experts. To address these limitations, graph neural networks (GNNs) have emerged as a complementary approach capable of modeling relational dependencies between samples. GNNs are particularly well suited for cases where relationships between images, such as symptom similarity or shared environmental conditions, provide valuable diagnostic cues [ 12 , 13 ] . By treating images as nodes and defining edges on the basis of pairwise similarities, GNNs aggregate information from neighboring images, enabling context-aware classification incorporating local features and global relational patterns. However, GNNs alone lack the ability to extract fine-grained spatial features directly from raw images — a key strength of CNNs. Therefore, combining CNNs and GNNs into a hybrid framework offers a synergistic advantage: CNNs capture localized spatial features within individual images, whereas GNNs enrich these representations with relational context across images. This hybrid approach is particularly valuable for soybean leaf disease classification, where local lesion characteristics and broader symptom similarity across fields, varieties, and conditions are essential for accurate and interpretable diagnosis. To address these gaps, we propose an interpretable hybrid sequential CNN-GNN architecture that sequentially combines MobileNetV2 for efficient spatial feature extraction and graph sample and aggregation (GraphSAGE), a GNN architecture, for relational dependency modeling between soybean leaf images. By constructing a similarity graph where nodes represent leaf images and edges encode pairwise feature similarity, GraphSage [ 14 ] aggregates information from neighboring nodes, enriching the feature representations with a relational context. This fusion of local spatial learning and global relational learning enhances classification accuracy while ensuring computational efficiency, making the model suitable for real-time field deployment. Additionally, we incorporate Grad-CAM and Eigen-CAM visualizations to provide interpretable heatmaps that highlight the specific leaf regions influencing each classification decision, bridging the gap between model predictions and expert validation. To the best of our knowledge, this is the first interpretable CNN-GNN hybrid framework applied to soybean leaf disease detection, addressing critical gaps in relational modeling, model transparency, and computational efficiency in plant disease classification research. We make the following key contributions in this work: 1. Sequential CNN-GNN Architecture: We propose a novel pipeline combining a pretrained MobileNetV2 for local feature extraction and a GraphSAGE model for global relational reasoning, enhancing our model’s ability to capture fine-grained disease symptoms and inter-symptom dependencies. 2. Graph Construction with Node Fusion and Adaptive Sampling: We introduce a domain-specific graph construction where each image is represented as a node with embeddings that fuse spatial and semantic features, whereas adaptive neighborhood sampling ensures robust classification even with similar symptoms or background noise. 3. Cross-Modal Interpretability: We employ Grad-CAM and Eigen-CAM for both CNN and graph-level feature attribution, providing clear insights into which image regions and relational cues contributed to our model’s decision and enhancing transparency in disease diagnosis. The remainder of this paper is organized as follows: Section II reviews related work in plant disease detection via CNNs and GNNs. Section III describes the proposed methodology, including the model architecture. Section IV presents the data preprocessing and experimental setup. Section V discusses the results and compares the performance of the proposed model with other baseline models, and Section VI concludes the paper with suggestions for future research. II Literature Review Soybean disease identification has become a key research focus in smart agriculture, with machine learning and deep learning techniques significantly improving classification accuracy. Early methods relied on traditional image processing and handcrafted features, such as K-means clustering and SVMs [ 15 ] or Gabor filters with ANNs [ 16 ] , but these approaches struggled to generalize across diverse symptoms and complex backgrounds. Recent agricultural image classification research has increasingly adopted deep learning, which has demonstrated strong performance across various

Read the complete text