Urban building extraction based on information fusion-oriented deep encoder-decoder network in remote sensing imagery

Abstract

The building extraction technology in remote sensing imagery has been a research hotspot. Building extraction in remote sensing imagery plays an important role in land planning, disaster assessment, digital city construction, etc. Although many scholars have explored many methods, it is difficult to realize high-precision automatic extraction due to the problems in high-resolution remote sensing images, such as the same object with different spectrum, the same spectrum with different object, noise shadow and ground object occlusion. Therefore, this paper proposes an urban building extraction based on information fusion-oriented deep encoder-decoder network. First, the deep encoder-decoder network is adopted to extract the shallow semantic features of building objects. Second, a polynomial kernel is used to describe the middle feature map of deep network to improve the identification ability for fuzzy features. Third, the shallow features and high-order features are fused and sent to the end of the encoder-decoder network to obtain the building segmentation results. Finally, we conduct abundant experiments on public data sets, the recall rate, accuracy rate, and F1-Score are greatly improved. The overall F1-score increases by about 4%. Compared with other state-of-the-art building extraction network structures, the proposed network is better to segment the building target from the background.