Simulating the architecture of a termite incipient nest using a convolutional neural network

Subterranean termites form colonies containing thousands of individuals, and maintain these colonies by consuming wood and other materials containing cellulose. In this consumption process, they cause serious damage to wooden structures. Information on the population size of termites is an important factor in developing strategies aimed at controlling termites. In this study, we provide a reasonable possibility of estimating the population of an incipient nest dug by a colony that has not yet discovered any food source. We build an agent-based model to simulate termite tunnel patterns in which the behavior of simulated termites (agents) is governed by simple rules based on empirical data. The simulated termites do not communicate with each other using pheromones. They move towards the ends of tunnels, excavate when their progress in that direction is blocked, and transport the excavated soil. Using simulations, we determine termite tunnel patterns according to three parameters: the number of simulated termites (N), the passing probability of two encountering termites (P), and the distance moved by termites to deposit soil parcels during tunneling activity (D). We train a convolutional neural network (CNN) using 80% of the tunnel patterns and apply the CNN to the remaining patterns to estimate the value of N. The application results show that the validation accuracy is approximately 41% and the training accuracy of the CNN is approximately 51%. Although the validation accuracy is not high, the estimation failures occur near the correct N values.