Abstract
Reduction in chemical usage for crop management due to the environmental and health issues is a key area in achieving sustainable agricultural practices. One area in which this can be achieved is through the development of intelligent spraying systems which can identify the target for example crop disease or weeds allowing for precise spraying reducing chemical usage. Artificial intelligence and computer vision has the potential to be applied for the precise detection and classification of crops. In this paper, a study is presented that uses instance segmentation for the task of leaf and rust disease detection in apple orchards using Mask R-CNN. Three different Mask R-CNN network backbones (ResNet-50, MobileNetV3-Large and MobileNetV3-Large-Mobile) are trained and evaluated for the tasks of object detection, segmentation and disease detection. Segmentation masks on a subset of the Plant Pathology Challenge 2020 database are annotated by the authors, and these are used for the training and evaluation of the proposed Mask R-CNN based models. The study highlights that a Mask R-CNN model with a ResNet-50 backbone provides good accuracy for the task, particularly in the detection of very small rust disease objects on the leaves.
Highlights
Precise Smart Spraying in SustainableThe application of chemicals such as fungicides, pesticides and herbicides via spraying in agriculture, with the aim to control plant disease, pests and weeds and ensure good crop yields, is common
To evaluate the capability of the proposed system for the tasks of leaf and rust disease identification, experiments were conducted for object detection and segmentation precision and disease detection
Three different feature extraction backbones to Mask R-CNN were trained as described in Section and evaluated, and these were ResNet-50, Large MobileNetV3-Large and MobileNetV3-Large320, respectively
Summary
The application of chemicals such as fungicides, pesticides and herbicides via spraying in agriculture, with the aim to control plant disease, pests and weeds and ensure good crop yields, is common. While promising results have been shown in the classification of crop disease from leaf images [12], there are some limitation regarding the datasets in which each image consists of single leaf with a contrasting background, this is not comparable to real-world conditions where leaves are bunched together, with other foliage as the background This issue was a motivation for the introduction of the Plant Pathology Challenge 2020 dataset [11]. An initial study is presented with the aim is to address the challenge of providing pixel level classification on images that depict real-life scenarios (i.e., multiple leaves per image with different disease states), which has the potential to provide accurate instructions to a smart spraying system capable of reducing chemical usage. A Mask R-CNN based instance segmentation system for precise leaf and rust identification on apple trees from RGB images; Manually annotated segmentation maps for a subset of the Plant Pathology Challenge.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
