A low-cost TinyML model for Mosquito Detection in Resource-Constrained Environments

Yearly, more than 200 million malaria cases are recorded worldwide. Most of these cases are witnessed in less developed countries as the environments are not well-maintained, which forms breeding places for mosquitoes. Female mosquito-anopheles is responsible for malaria infection, dengue, chikungunya, and zika. Developing countries struggle to fight diseases; malaria still claims more than 400,000 lives annually. One current way to keep away anopheles mosquitoes is using commercially available electric liquid mosquito repellents, which can adversely affect the human body when used for extended periods. Furthermore, energy and sprays are wasted as they constantly work even without the presence of anopheles mosquitoes. We propose a low-cost IoT-based TinyML model that intelligently discharges the mosquito repellent when an anopheles mosquito is in the room. First, we prove the concept by exploring two lightweight deep learners with a 1D Convolution Neural Network (1D-CNN) and 2D Convolution Neural Network (2D-CNN) to classify raw sounds from mosquito wingbeats. We adopted a Leaky ReLU in building the 1D-CNN to speed up training and improve classification performance. Furthermore, we adopted batch normalization to avoid degradation and vanishing gradient problems. We implemented the experiments in an Edge impulse platform. Each of the CNN models recorded stable classification performance during the proof of concept study, while the 1D-CNN took less time and computing resources in training, validation, and testing. As we aimed to propose a low-cost solution, we evaluated the performance of the 1D-CNN-based prototype in the actual deployment by playing mosquito wingbeat sounds on a laptop which we placed next to it in intervals of 0.5, 1.0, 1.5, 2.0, 2.5, and 3 meters. The model showed promising results across distances and thus could be used to chase away mosquitoes in a room of small to medium size.