Insect interaction analysis based on object detection and CNN

Abstract

Direct observation to study biodiversity can be time consuming, however, other methods often provide indirect measurements and are possibly biased. To solve these problems, images can be a useful tool and ecologists have started to rely more and more on images as a source of data and on automated image analysis. However, the existing methods mostly perform image classification. In this paper we present an efficient method based on object detection to access deeper information the content of an image. Using high resolution images, we built a pipeline to slice the original images, perform detections and later refine these observations. We illustrate the interest of this pipeline by using it on-field images taken in agroforestery banana-coffee systems to study invertebrate communities around the banana pests Cosmopolites sodidus and Metamasius sp. and the interactions between the different animals within this community. Experimental results show that our pipeline reaches 87.8% F1score and allows us to successfully detect and identify 23 species and ant castes. These 23 species are divided into 7 superclasses, but the ant super-class, that shows more individuals and interactions is described more precisely. We are then able to study the interaction network between different species of this community and identify major predators of banana pests within this ecosystem.

A Systematic Literature Review on Insect Detection in Images

“…In Lyu et al [25], the pictures of the desired insects were taken in the laboratory. While Yang et al [26] used Zhongwei Kechuang industrial cameras installed in field, other authors [27]- [29] used digital mobile cameras. Currently, Du et al [30] use high-resolution UAV, and Bjerge et al [31] use previously constructed portable computer versions, while Ard et al [32] use a Scoutbox along with a camera, and Rustia et al [33] use a wireless image monitoring system.…”

Deep learning object detection to estimate the nectar sugar mass of flowering vegetation

“…Later Xia et al (2018) obtained a detection accuracy of 0.76 for their CNN of sea cucumber species on model-unseen internet images, and that of Tresson et al (2019) detected and identified individuals and castes of 23 ant species to a detection accuracy of 0.88.…”

Hierarchical Classification of Very Small Objects: Application to the Detection of Arthropod Species

“…One solution is to slice the large input image and perform detection on slices separately before merging all results. This method can be used for satellite imagery analysis [16] or insect detection for example [17].…”

“…With large input images, detection might require a preprocessing phase. For this step we propose to apply an approach based on the method proposed by Tresson et al [17]. The pre-processing step is summarized in Algorithm 1.…”