Consequently, it is possible to improve F-score by eliminating small objects. I noticed precision on small objects is not good compared with other objects. After training the model with 70 percent of the data, the trained model is evaluated on the remaining 30 percent and chose the best one. To avoid overfitting, early stopping with Jaccard coefficient is applied for training. To develop the individual U-net model, I split the training data into two parts: 70 percent for training and the remaining 30 percent for validation. L-shaped and concave buildings are detected successfully.įinal Approach: Averaging Ensemble of three U-Net based models Parameter optimization Later I found that using 8-bands multispectral data improves the performance.įigure 2.1: Another example output in Vegas. In early stage of the contest, I only used RGB 3 channels. By contrast, my U-Net based model can detect L-shaped and concave buildings successfully. Lee Cohn, a data scientist at CosmiQ Works, described his result of applying the Multi-task Network Cascades (MNC) and MNC struggles with L-shaped and concave buildings. Another example output of L-shaped and concave buildings in Vegas is shown in Figure 2.1.

Most building footprints are successfully detected with high intersection area over union (> 0.8). Figure 2 shows an example output by my solution with U-Net models.

It can be trained end-to-end from few images and outperform the prior best method on the ISBI cell tracking challenge 2015. U-Net is one of the most successful and popular convolutional neural network architecture for medical image segmentation. My model is based on a variant of fully convolutional neural network, U-Net, which is developed by. I solved the problem as a semantic segmentation task in computer vision. I individually investigated and evaluated the similar approach on Spacenet Challenge dataset (my first submission with the approach is on May 13, a little earlier than their publication). , 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.

Spacenet 2