Transformer-based architectures for early detection of storage pests in maize grain silos: A review

Abstract

Post-harvest losses due to insect infestation represent a critical threat to global food security, with maize grain silos facing particular vulnerability. Traditional pest monitoring methods-manual sampling, electronic noses, and acoustic detection-suffer from fundamental limitations including labor intensity, environmental instability, and inability to penetrate deep grain piles. This review examines the emerging application of Transformerbased artificial intelligence architectures for early detection of storage pests in maize grain silos. We synthesize recent advances in three interconnected domains: sensor systems for in-situ pest detection, time-series datasets capturing pest activity patterns, and Transformer models adapted for pest counting and species identification. The Pest Manager framework, representing the first systematic deployment of Transformers for invisible grain pile monitoring, achieves 99.2% counting accuracy and 86.9% species identification accuracy across five major storage pests. We compare this approach with alternative deep learning methods including convolutional neural networks, hybrid architectures, and lightweight models for resourceconstrained deployment. Key challenges addressed include device-induced variability, fine-grained species discrimination, and real-time processing requirements. The review concludes with research priorities for integrating Transformer-based detection into comprehensive smart storage management systems, with particular emphasis on maize silo applications in developing economies where storage losses exceed 20% annually.