Supervised Machine Learning for Bridge Detection in Digital Elevation Models

Ryan Carlson and Andy Danner

One application of Geographic Information System (GIS) data is the hydrological analysis of gridded Digital Elevation Models (DEMs). Using Light Detection and Ranging (LIDAR) technology, large, hi-resolution elevation data can be collected. In addition to terrain, the data might include smaller features like trees, buildings, and bridges. The goal this summer was to automate detection of bridges, which impede watershed analysis, and to alter the terrain to yield hydrologically correct DEMs. A first attempt revealed that a set of simple filters were able to detect bridges, but also detected many other non-bridge features. A machine learning algorithm called AdaBoost was used to combine these filters. It takes features describing hand-labeled positive and negative training examples as input. The algorithm combines these features into a powerful bridge classifier. This classifier is then used on unlabeled test data and outputs a grid that assigns to each likely bridge location a numerical score. Using a training set of about 400 hand-tagged positive and negative examples, the classifier correctly classified bridges over 95% of the time. Given more training data and features, results should improve. The next steps are to add more hand-tagged data, add more features, and condition the terrain to remove these obstructions to allow for correct watershed mapping.