, published on May 17, also investigates the use of OpenStreetMap for semantic labeling of satellite image. My best individual model simply uses OpenStreetMap layers and multispectral layers as the input of the deep neural network simultaneously (as described in Figure1).įigure 1: Best individual model with using OpenStreetMap and Pan-sharpened Multispectral data. In addition, I found the use of OpenStreetMap data is effective for predicting the building footprint. My final submission is the averaging ensemble from individually trained three U-Net models. I applied a modified U-Net model, one of deep neural network model for image segmentation. The polygon of building footprint proposed by the algorithm is considered as a true positive if its IOU (Intersection over Union, Jaccard index) score is higher than 0.5. The algorithm is evaluated based on F-score. SpaceNet Challenge Round2 asks its participants to submit an algorithm that inputs satellite images (of Las Vegas, Paris, Shanghai and Khartoum) and outputs polygons of building footprints. It consists of an online repository of freely available satellite imagery, co-registered map layers to train algorithms, and public challenges that aim to accelerate innovation in machine leanring. SpaceNet is a collaboration between DigitalGlobe (a commercial vendor of space imagery and geospatial content), CosmiQ Works (a division of In-Q-Tel Lab) and NVIDIA (the world leading company in visual computing technologies). For training a deep neural network model, the computational time on p2.xlarge (Tesla K80) is two times longer than my personal graphic card (GeForce GTX 1080).Adding OpenStreetMap layers into the input of U-Net model significantly improves F-score.
0 Comments
Leave a Reply. |