LONG-DISTANCE CABBAGE DAMAGE AND PEST DETECTION METHOD USING YOLO11

Abstract

Context. To ensure sustainable yield, plant health must be constantly monitored with timely measures applied to prevent disease spread. Traditional approaches rely on manual inspection of plants, while neural networks require large amounts of annotated data to train. Both manual inspection and data annotation require expert knowledge and are time-consuming. Close-up photos of leaves are often used for training as they are easier to collect from the Internet. However, this complicates disease spread estimation at a scale. Cabbage is one of the plants widely grown in Ukraine, but existing research focusing on cabbage health monitoring is limited.
Objective. The goal of this work is to build a neural-network-based cabbage disease and pest detection system, which can be trained in on a small number of training images. At inference the system should detect pests on plant images at a distance of a whole plant.
Method. Given that existing plant disease datasets, such as IP102 and PlantDoc mostly contain close-up images of diseased plants, the networks trained on such datasets suffer from lack of generalization to images at a distance. To select the best object detection model, state-of-the-art object detection architectures, namely YOLO 8, 9, 10, 11, and RT-DETR have been analyzed in the work. To increase detection distance multi-image loss is proposed to improve hyperparameter search using Tree-Structured Parzen Estimators. Also, to improve detection quality, a novel cabbage disease dataset has been collected in Dnipro region, Ukraine. The new classes include crucifier flea beetle (widespread pest in Dnipro region) and damaged leaf. When the pest is not visible, but leaf damage is taken, determining specific pest might not be possible. Therefore, we introduce additional damaged leaf class, that captures generic plant damage. This also enables tracking of plant healing rate, when measures to stop pest spread have been taken. We combine collected images with the larger IP102 dataset to increase the number of pests covered to form new Cabbage+IP102 dataset.
Results. 1) Tree-Structured Parzen Estimators search on the multi-image loss has improved the YOLO 11 M performance from 0.3642 to 0.3892 mAP50-95 on images taken at a distance. 2) Collected dataset has enabled detection of cabbage plant health problems at a distance, including cases when the pest is currently not visible, but the damage is present.
Conclusions. In this work, the cabbage pest and damaged leaf YOLO 11 M detection system has been presented. The detector
architecture has been selected as the best-found during analysis on 2 datasets. The developed system requires only 7 annotated cabbage images to be trained and to perform pest and damaged leaf detection on high resolution images (2016x2016) of whole cabbage plants. The final model can be used to monitor cabbage health problems, damage, and rate of healing using images taken at a distance.