Fine-Grained Building Extraction With Multispectral Remote Sensing Imagery Using the Deep Model

Extracting buildings from remote sensing imagery can serve many industries. Single-category building extraction, however, has been unable to meet the actual needs of sectors, such as disaster prevention and mitigation. In this study, we create a large-scale high-quality multispectral dataset (BFE-Set) for fine-grained extraction of buildings. BFE-Set uses the GaoFen-2 multispectral satellite as the data source and divides buildings into three fine-grained categories according to their structure types: steel and reinforced concrete structure, masonry structure, and block stone structure. It includes 20 countries, where disasters often occur in China, and contains nearly 260000 building instances. We propose the BFE-Net, a fully convolutional neural network that performs better on the BFE-Set for fine-grained extraction. The styles of buildings in different regions are quite different. Through the combined use of batch normalization (BN) and instance normalization (IN), BFE-Net enhances the model's style invariance. BFE-Net extracts feature from visible light and near-infrared data through a parallel feature extraction subnetwork and uses an attention mechanism to fuse the two modal features. This takes full advantage of the characteristic complementary effects of visible light and near infrared. In addition, knowledge distillation (KD) is used to fully mine the interclass relationships of fine-grained buildings, making the model more generalized in fine-grained classification. The experimental results show that BFE-Net is superior to other building extraction methods, and the fine-grained extraction effect is encouraging, which can provide a valuable reference. BFE-Set and BFE-Net will play a greater role in industries, such as disaster response and smart cities.