OEC-RNN: Object-Oriented Delineation of Rooftops With Edges and Corners Using the Recurrent Neural Network From the Aerial Images

It is an important task to automatically and accurately map rooftops from very high resolution remote sensing images since buildings are closely related human activity. Two typical technologies often utilized accomplish the task, i.e., semantic segmentation instance segmentation. The independently allocate a label (e.g., “building” or not) each pixel, resulting in blob-like segments. On contrary, one might model boundary of rooftop as polygon improve shape by encouraging vertices adhere rooftop’s boundary. Following this line work, we present multitask learning approach predict corners sequent way using attention learned where boundaries given image region. simulates process manual delineation rooftops’ outline image, which can produce accurate with sharp straight lines between them. Specifically, proposed method consists three components, object detection, pixel-by-pixel classification both edges corners, manner convolutional recurrent neural network (RNN). called object-oriented, (OEC)-RNN article. Three datasets employed validate performance OEC-RNN, compared state-of-the-art methods for experimental results show that OEC-RNN achieves best terms overlay, adherence, vertex location ground-truth predicted polygons.