A CONVOLUTIONAL NEURAL NETWORK-BASED APPROACH FOR DETECTION OF OBJECTS AND SCENARIOS SUSPECTED OF BEING POSSIBLE BREEDING SITES OF AEDES AEGYPTI MOSQUITO

Resumo

In remote sensing field, the use of satellite images to detect small objects becomes impractical
due to the low spatial resolution of such images. With the use of unmanned aerial vehicles (UAVs),
popularly known as drones, it is possible to acquire aerial images with high spatial and temporal
resolutions that allow the detection of small objects on the earth's surface and the perception of changes
in a certain region in a short period of time. However, the detection of certain objects in the images
acquired by UAVs has been a great challenge due to the amount of details present in these images,
especially those acquired in urban areas. This work investigates the automatic detection of objects and
scenarios in aerial images acquired by UAVs. The proposed approach aims to detect target objects
(typical containers for water storage) and scenarios suspected to be possible breeding sites of the Aedes
aegypti mosquito. For the detection of the objects and scenarios, two convolutional neural network
(CNN) architectures from the YOLOv3 framework were employed. In the experiments conducted, we
considered 142 images acquired in peripheral regions of the city of São Paulo/Brazil containing water
tanks; 25 images of real scenarios, obtained by Google image search, containing garbage and old tires
and 35 images containing scenarios simulating the main mosquito breeding sites (old tires, gutters,
among others). In the experiments performed, in which the proposed approach was evaluated using the
mean average precision (mAP), the rates of 0.95 and 0.97, respectively, were obtained for the detection
of objects and scenarios, indicating that the proposed approach is a good alternative to solve the
investigated problem.