Spider Mites Detection in Wheat Field Based on an Improved RetinaNet

Abstract

As a daily staple food of more than one third of the world’s population, wheat is one of the main food crops in the world. The increase in wheat production will help meet the current global food security needs. In the process of wheat growth, diseases and insect pests have great influence on the yield, which leads to a significant decline. Wheat spider mites are the most harmful to wheat because they are too small to be found. Therefore, how to use deep learning to identify small pests is a hot spot in modern intelligent agriculture research. In this paper, we propose an improved RetinaNet model and train it on our own dataset of wheat spider mites. Firstly, the wheat spider mites dataset is expanded from 1959 to 9215 by using two different angles and image segmentation methods. Secondly, the wheat spider mite feature detection head is added to improve the identification of small targets. Thirdly, the feature pyramid in FPN is further optimized, and the high-resolution feature maps are fully utilized to fuse the regression information of shallow feature maps and the semantic information of deep feature maps. Finally, the anchor generation strategy is optimized according to the amount of mites. Experimental results on the newly established wheat mite dataset validated our proposed model, yielding 81.7% mAP, which is superior to other advanced object detection methods in detecting wheat spider mites.

Keywords

• wheat spider mites
• improved RetinaNet
• object detection
• image processing