Mapping fire blight cankers and autumn blooming in pear trees using Faster R-CNN

Fire blight disease causes significant losses in pear orchards. Fire blight infection is accompanied by visual symptoms that can easily be recognized by a grower or adviser, but such visual inspection is time consuming. The present work focused on the use of convolutional neural networks (CNNs) to identify one type of visual symptom (cankers on the main trunk of dormant trees) as well as autumn blooming, which in Israel plays an important role in the epidemiology of Erwinia amylovora-the bacterium responsible for fire blight. Images of dormant trees were acquired with a tripod-mounted DSLR camera while, for autumn blooming detection, the images were acquired using a small unmanned aerial vehicle flying a few meters above the trees. In both cases, several Faster R-CNNs were trained and tested with several datasets acquired at various locations and over several years. Overall, the CNNs for canker detection achieved precision and recall rates that exceeded 90% while, for autumn blooming detection, the precision and recall rates exceeded 80% in all but one case. These trained CNNs were used to analyze automatically geo-references images, hence generating infection/blooming maps. Such maps could be one of the information layers used by growers for managing the orchard, for instance to determine whether winter sanitation is needed and/or if it was carried out properly, or to decide when and where costly manual removal of autumn flowers or copper application is required.

Automated summary

This study utilizes convolutional neural networks to identify visual symptoms of fire blight disease in pear trees and to detect autumn blooming. The trained CNNs are used to analyze images and generate infection/blooming maps, which can assist growers in managing orchards and reducing labor and economic costs.