Convolutional neural network for honeybee density estimation

Abstract

Honeybees (Apis mellifera L.) perform an important service to the ecosystem as they function as significant pollinators for plants. Over the past few decades, honeybees have suffered a progressive decline. The ability to observe and control the activity and density dynamics of honeybees in their hives in an automated way can allow to stimulate their actions and improve the overall efficiency of the hive. In this paper we present a novel honeybee observation method that is primarily based on honeybee density estimation using a convolutional neural network. First, specially designed stationary robots were positioned inside an arena designed for young honeybees. Three robots were used, each equipped with six infrared sensors for honeybee detection. The hardware and software setup that was used during the raw data collection process is described. Using the collected data from experiments with different numbers of honeybees we tested different convolutional neural networks to evaluate the relation between the network parameters and the estimation accuracy. To obtain better results, the numbers of honeybees were grouped into four different categories. It is shown that the most influential parameters are the number of epochs and the feature map size. By using the correct parameters it is possible to obtain 100 % accuracy during network training process and 86 % accuracy during evaluation process.