DETECTION AND IDENTIFICATION OF INSECT PESTS ON CROP LEAVES USING A SMALL-OBJECT-OPTIMISED YOLO DEEP LEARNING MODEL

Abstract

Insect pests are a major cause of crop loss worldwide, and the timely, accurate detection of an infestation is the foundation of integrated pest management and of targeted, rather than calendar-based, pesticide application. Manual scouting for pests is slow, requires entomological expertise, and is especially error-prone for the very small insects — aphids, thrips, whiteflies, and mites — that cause much of the damage. This paper presents an automatic system for detecting and identifying insect pests on crop leaves using a single-stage YOLO object detector adapted for very small targets. To improve recall on tiny insects that occupy only a few dozen pixels, the detector is augmented with an additional high-resolution (P2) detection head, attention modules in the neck, and a tiling strategy that preserves the native resolution of small objects. The system localises each pest with a bounding box, classifies it into one of six common species, counts the pests per leaf, and derives an infestation-level alert. Trained and evaluated on a leaf-pest image dataset spanning six species, the proposed model achieves a mean Average Precision (mAP@0.5) of 0.89, improving on a baseline YOLO detector by five percentage points, with per-species AP ranging from 0.81 for the smallest pests (spider mites) to 0.93 for the largest (armyworm). The system runs in real time and provides an accurate, scalable basis for early-warning pest monitoring and site-specific pesticide application.