结合模糊度和形态学指数约束的深度学习建筑物提取

基于高分辨率遥感影像的建筑物提取一直是研究的热点问题,深度学习的深层次特征提取方法,非常适合高分辨率影像中建筑物的提取,但使用深度学习提取建筑物时,大多以改变网络结构为主进行算法优化,很少与其他方法结合。本文研究在改进深度学习网络结构的基础上,结合影像模糊度约束增强、形态学建筑指数约束增强等方法,对建筑物提取方法进行更全面更有针对性的改进。本文主要改进内容为：① 提出PwDeepLab网络,该网络基于DeepLab v3+网络结构,在特征融合方式和损失函数等方面进行了改进。② 提出模糊度约束方法,在固定影像块大小的情况下,通过影像模糊度约束对影像进行上采样增强。③ 提出形态学指数约束方法,通过形态学建筑物指数（MBI）约束范围拉伸增强的方法,在较少改变原始影像特征的情况下,突出建筑信息。本文在Massachusetts数据集和武汉大学的Satellite Dataset Ⅱ(East Asia) 数据集上进行验证, 2个数据集的主要建筑类型存在较大区别。本文提出的方法在2个数据集上的精度相对于DeepLab v3+分别提高了10.9%和3.8%,相对于U-Net分别提高了10.0%和9.6%。实验结果表明本文提出的方法对建筑物提取效果有明显提升,且具有很好的鲁棒性和通用性。

Building Extraction by Deep Learning Method Combined with Ambiguity and Morphological Index Constraints

Extraction of buildings from high-resolution remote sensing images has been a hot topic. It is important to quickly and accurately extract the location and outline of buildings from high-resolution remote sensing images for earthquake disaster assessment, urban and rural planning management, smart city construction, and other fields. However, it is challenging to extract buildings accurately from high-resolution remote sensing images due to the complexity of ground features and the occlusion phenomenon. The classic building extraction algorithms usually have incomplete results with many wrong detections and missed detections. The deep-level feature extraction method of deep learning is very suitable for the extraction of buildings from high-resolution images. However, for the existing researches using deep learning algorithms, most algorithms are optimized by changing the network structure and are rarely combined with other methods. The pertinence of building extraction is not strong. This paper not only studies the influence of the internal structure of the convolutional neural network in deep learning on the extraction results, but also studies the combination of ambiguity, Morphological Building Index (MBI) with deep learning. The main improvements in this paper are as follows: (1) Based on the DeepLab v3+ network structure, we propose the PwDeepLab network which improves feature fusion methods and loss functions; (2) We propose a blur degree constraint method. We define a new blur degree formula to evaluate the image blur degree. In the case of a fixed image block size, the image is upsampled and enhanced by the image blur degree constraint; (3) We propose a morphological index constraint method. The pixels of the original image are stretched and enhanced where the Morphological Building Index (MBI) is above the threshold. Therefore, the building information can be highlighted with less changes to the original image characteristics. Our method is verified using the Massachusetts dataset and the satellite dataset II (East Asia) of Wuhan University. The main building types of the two datasets are quite different. As a result, the accuracy of the proposed method on the two datasets increases by 10.9% and 3.8%, respectively, compared with DeepLab v3+, and increases by 10.0% and 9.6%, respectively, compared with U-Net. The higher accuracy reflects the superiority and robustness of our method. Moreover, the extracted results match the real labels very well in details. The experimental results show that the method proposed in this paper can significantly improve the building extraction results.