Building Footprint Recognition through Satellite Imagery Identification by a Convolutional Neural Network

We implemented a deep learning model training to detect building footprints automatically for the metropolitan area in Taiwan. Building a database is one of the main components to develop a natural hazard risk model for Taiwan. Due to the limitations of privacy policies, databases from government sources are usually not supposed to be published. Alternatively, we produce building footprint data through Mask R-CNN, a convolutional neural network (CNN) widely used in image segmentation to predict individual objects. To detect building footprints based on the Mask R-CNN, we proposed three procedures: image preprocessing (obtain pan-sharpening multispectral remote sensing image and calculate a normalized vegetation index from red and near infrared wave bands) to increase data information; model training to produce a general model; and post processing to normalize detection results. We implemented ca. 9,000 building footprints in Taichung City from the National Land Surveying and Mapping Center as our training data. The Mask R-CNN model showed good performance in excluding vegetation around buildings, and obtained a 0.78 average precision score over IoU (intersection over union) >0.5. To validate the applicability, we applied this model to another image and got recall and precision of 0.75 and 0.65, respectively. Our deep learning model provides a more efficient way to build our building information database, shedding light on a more comprehensive natural hazard risk assessment.