An Automatic Extraction Architecture of Urban Green Space Based on DeepLabv3plus Semantic Segmentation Model

Abstract

Urban green space plays an important role in maintaining the balance of ecological environment and the sustainable development of the city. Extracting urban green space from remote sensing imagery can provide fast and accurate reference for urban planning and management. Deep learning semantic segmentation is a new exploration for image processing including remote sensing (RS) images these years. This paper describes a multilevel architecture which targets urban green space extraction from GF-2 imagery. The pillar of the architecture is semantic segmentation model (DeepLabv3plus) that is used for satellite imagery classification. In this paper, we take 19687 256 pixels × 256 pixels slices from Gaofen-2(GF-2) satellite images of three different cities and ground truth as training samples, and data of another city in Hebei province is taken for model verification. Then, five comparative methods are carried out for monitoring urban green space distribution, including ML (Maximum likelihood), SVM (Support Vector Machine), Object-oriented method, FCN and U-Net. The accuracy indexes of each method are obtained by calculating the difference between the extraction result and the ground truth. The results shows that the architecture with DeepLabv3plus outperforms the other five methods allowing us to better extract urban green space, in particular eliminating interference from farmland pixels. The architecture allows us to reach the target extraction accuracy of 89.46%. This paper is also an exploration of applying artificial intelligence technology to remote sensing image processing.