Identifying temporary water bodies from drone images at real-time using deep-learning techniques

The detection and control of larval habitats—in particular, temporary water bodies—can play an important role in fighting against mosquito-transmitted diseases like malaria. In this paper, we explore the feasibility of identifying temporary water bodies from images taken by surveillance drones at real-time using the YOLOv4 algorithm. For our experiments, we use a dataset consisting of 782 drone-captured images, which includes 1522 temporary water bodies of different sizes and shapes, including puddles (i.e., small accumulations of water on a surface). Although still preliminary, our results are encouraging: the YOLOv4 trained models obtained a general score of 71.47(6.42)% mAP, and the processing time per image is 24.87 milliseconds, on average. In terms of size, the models can identify puddle objects as small as 23x23 cm on images taken from 120 m over surface-level.