Abstract:
Early detection of rice leaf folder (Cnaphalocrocis medinalis) is essential for mitigating
yield losses in Sri Lankan paddy fields, where rice is a staple crop under constant threat from
pest infestations. This study assesses the operational readiness of AI-based pest detection
systems and the field imaging constraints that impact their deployment in real-world
agricultural environments. Field data were collected in the Trincomalee district, capturing
diverse imaging challenges such as leaf overlap, motion blur, glare from standing water, and
limited camera access due to field conditions. A manually annotated dataset was constructed
and used to train and evaluate various AI models. Seven categories of AI architectures,
including deep CNNs, object detection networks, lightweight and hybrid models, were
assessed for accuracy, efficiency, and deployability. Models such as DenseNet201 (93.87%),
MobileNetV2 (89.20%), and YOLOv8 (88.90%) demonstrated high performance while
meeting the requirements for field use in resource-limited settings. However, the study
emphasizes that model accuracy alone is insufficient; interpretability, transparency, and trust
are crucial for adoption by farmers and agricultural officers. The integration of explainable
AI (XAI), domain adaptation using local datasets, and hardware compatibility were
identified as key enablers for field deployment. This research underscores the significance
of addressing real-world imaging limitations and developing localized, low-cost, and
scalable AI tools for early pest detection. Future efforts will focus on expanding datasets
across agro-ecological zones, improving drone and smartphone-based imaging, and co-
developing mobile applications with stakeholders to promote practical, real-time decision-
making in integrated pest management.
