Artificial intelligence-powered multiclass deep learning model for detection of aflatoxin-related defects in Ugandan groundnuts

Abstract

Aflatoxin contamination poses a persistent challenge to groundnut value chains in sub-Saharan Africa, where conventional laboratory-based detection methods are costly, time-consuming, and often inaccessible to smallholder farmers. This study presents an artificial intelligence–powered multiclass deep learning framework for image-based detection of aflatoxin-related defects in groundnuts. A curated dataset of 2252 groundnut kernel images was compiled and categorized into four classes: Healthy, Moldy, pest-infested, and physiological disorder. The dataset was partitioned into training, validation, and test sets, with targeted data augmentation applied to address class imbalance. The proposed model employs an Inception-ResNet-V2 architecture with transfer learning, class-weighted categorical cross-entropy loss, and optimized hyperparameters to enhance multiclass discrimination. Model performance was evaluated using accuracy, class-wise precision, recall, F1-score, and receiver operating characteristic analysis. The model achieved an overall classification accuracy of 99.29% on the independent test set, with class-specific AUC values of 1.00 (Moldy), 0.98 (Healthy), 0.97 (Pest-Infested), and 0.99 (Physiological Disorder). These results demonstrate strong generalization and robust differentiation of visually similar defect classes. The findings indicate that multiclass deep learning can effectively support early-stage screening of aflatoxin-associated defects in groundnuts, providing a scalable and low-cost complementary tool to conventional aflatoxin testing methods. Publicly Available Content Database