Optimal Models for Plant Disease and Pest Detection Using UAV Image

• Published in: Nature environment and pollution technology 2022-12, Vol.21 (4), p.1609-1617
• Main Authors: Liang, Dashuang, Liu, Wenping, Zhao, Yugang
• Format: Article
• Language: English
• Subjects: Accuracy; Affine transformations; anchor-free detector, deep learning, unmanned aerial vehicle, automated machine learning, pest detection; Automation; Consumption; Datasets; Deep learning; Forestry; Least squares; Machine learning; Medical imaging; Methods; Model accuracy; Object recognition; Pests; Plant diseases; Remote sensing; Resource consumption; Training; Unmanned aerial vehicles
• ISSN: 2395-3454; 0972-6268; 2395-3454
• DOI: 10.46488/NEPT.2022.v21i04.013

The use of deep learning methods to detect plant diseases and pests based on UAV images is an important application of remote sensing technology in modern forestry. This paper uses a CenterNet-based object detection method to construct models for plant disease and pest detection. The accuracy of the models is influenced by parameter alpha, which is used to control the affine transformation in the preprocessing of CenterNet. First, different alphas are sampled for training and testing. Next, the least square method is used to fit the curve between alpha and accuracy measured by mAP (mean average precision). Finally, the equation of the curve is fitted as mAP = -0.22 * alpha2 + 0.32 * alpha + 0.42. In comparison, an automated machine learning (AutoML) method is also conducted to automatically search for the best model. The experiments are done with 5,281 images as the training dataset, 1,319 images as the verification dataset, and 3,842 images as the test dataset. The results show that the best alpha value obtained by the least square method is 0.733, and the accuracy of the corresponding model is 0.536 in mAP@[.5, .95]. In contrast, the accuracy of the AutoML method model is higher with the model accuracy of 0.545 in mAP@[.5, .95]. However, the training time and training resource consumption of the AutoML method are about 3 times that of the least square method. Therefore, in practice, a trade-off should be made according to the accuracy requirements, resource consumption, and task urgency.