Edge-Compatible Deep Learning Models for Detection of Pest Outbreaks in Viticulture

The direct effect of global warming on viticulture is already apparent, with unexpected pests and diseases as one the most concerning consequences. Deploying sticky traps grape plantations to attract key insects has been backbone conventional pest management programs. However, they are time-consuming processes for winegrowers, conducted through visual inspection via manual identification counting insects. Additionally, winegrowers usually lack taxonomy expertise accurate species identification. This paper explores usage deep learning edge identify quantify counts automatically. Different mobile devices were used acquire a dataset yellow delta traps, consisting 168 images 8966 manually annotated by experienced specialists. Five different models suitable run locally selected, trained, benchmarked detect five insect species. Model-centric, data-centric, deployment-centric strategies explored improve fine-tune considered models, where tested low-end high-end devices. SSD ResNet50 model proved be architecture deployment devices, accuracies per class ranging from 82% 99%, F1 score 58% 84%, inference speeds trap image 19.4 s 62.7 smartphones, respectively. These results demonstrate potential approach proposed integrated into mobile-based solution vineyard monitoring providing automated detection vector