Early prediction of bumblebee flight task using machine learning

Abstract This work demonstrates the development of a neural network algorithm able to determine the function of a bee’s flight within six measurements ( ≈ 18 s with current radar technology) of its relative position on leaving a nest. Engineering advancements have created technology to track individual insects, unlocking research possibilities to investigate how bumblebees react to their environment in more detail. This includes how they discover and make use of resources. The development of an intelligent algorithm would allow for the automated monitoring of resource use and nest health. An imbalance of bee flight tasks may indicate a shortage of resources or over-reliance on a plant that may soon stop flowering. Recent developments using drones to track insects can benefit from an intelligent target acquisition system given limited drone battery life. Such knowledge will also benefit the tracking itself by allowing for customised flight parameters to match target flight patterns. Data captured by these tracking techniques are taxing to parse manually using human expertise. Artificial intelligence can produce meaningful knowledge faster with equal precision. In this work, a comparison between a neural network (NN), random forest (RF), and support vector machine (SVM) is provided to distinguish the best model for the task by comparing cross entropy loss and accuracy across the dataset, showing improved results as time goes on. In situations where the radar lost sight of the target, a purpose-built filter was created to mitigate signal losses. The generated model provides results with a peak accuracy of 92%. This model, combined with the filter, create an opportunity to monitor the number of bees leaving the nest for each flight task with smaller, cheaper, and stationary receiver solutions with shorter ranges by removing the need to track a bee for its entire flight to ascertain its errand.