A deep learning approach for detection and location small portions of water in aerial images acquired by drones

Resumo

Drones have been used to automatically identify objects and scenarios (normally water tanks, buckets, plant pots, and other containers contained in open-air trash) that characterize potential breeding sites of mosquito, such as Aedes aegypt, from the acquired images. However, despite knowing that water stagnation is an essential condition for mosquito breeding sites, computer vision systems proposed in the literature for automatic image analysis do not include the detection of water in suspicious objects and scenarios, which constitutes a technical limitation for the effective use of drones in vector monitoring and control actions. In this work, a method based on deep learning is proposed to detect and locate small portions of water in multispectral images acquired by drones. To carry out the experiments, we composed a database of multispectral images acquired from simulated scenarios containing small containers with and without water in a controlled environment. The high rates obtained in terms of the mAP50 metric (above 90%) in computational experiments using a convolutional neural network YOLOv8 to segment the images confirm the potential of proposed approach to increase the technical viability of existing computer vision systems, making them more effective in combating mosquito breeding sites, bringing important contributions to the public health area.