Building detection using a dense attention network from LiDAR and image data

Accurate building mapping using remote sensing data is challenging because of the complexity of building structures, particularly in populated cities. LiDAR data are widely used for building extraction because they provide height information, which can help distinguish buildings from other tall objects. However, tall trees and bridges in the vicinity of buildings can limit the application of LiDAR data, particularly in urban areas. Combining LiDAR and orthoimages can help in such situations, because orthoimages can provide information on the physical properties of objects, such as reflectance characteristics. One efficient way to combine these two data sources is to use convolutional neural networks (CNN). This study proposes a CNN architecture based on dense attention blocks for building detection in southern Toronto and Massachusetts. The stacking of information from multiple previous layers was inspired by dense attention networks (DANs). DAN blocks consist of batch normalization, convolution, dropout, and average pooling layers to extract high- and low-level features. Compared with two other widely used deep learning techniques, VGG16 and Resnet50, the proposed method has a simpler architecture and converges faster with higher accuracy. In addition, a comparison with the two other state-of-the-art deep learning methods, including U-net and ResUnet, showed that our proposed technique could achieve a higher F1-score, of 0.71, compared with 0.42 for U-net and 0.49 for ResUnet.