Detecção de embarcações em imagens de sensores orbitais ópticos por meio de um método baseado em características contextuais e comparação de contorno

Abstract

Due to the economic importance of fishing such as oil and gas reserves in marine subsoil and the use of sea routes for trade, tourism and recreation, the Brazilian maritime territory is one of the main resources of the country. Therefore, it is necessary to monitor and protect it. Remote sensing is used to monitor and get data from earth resources over large areas because it is not necessary any direct contact with the investigated region. This dissertation presents a method to detect vessels in images from orbital optical receivers that uses contextual characteristics, related to spectral, texture and shape, such as compactness and diameter, and contour shape comparison characteristics of targets detected in image by applying edge detection filters. These characteristics represent descriptors that are submitted to a neural artificial network classification process. The method was evaluated by precision, recall and F-measure and it was applied to RapidEye images with 5 meters spatial resolution. The average method result was 94,94% and the average method recall was 100%. It means that all medium to large sized vessels in images were detected. Precision failures were caused by false positives related to small and medium islands. The method was able to detect vessels beside each other or docked.