融合高度特征的高分遥感影像多尺度城市建筑类型分类

城市区域建筑类型信息在城市功能区识别、城市环境变量反演等应用领域具有重要作用。本文提出一种融合高分辨率遥感影像高度特征的多尺度城市建筑类型分类方法。首先利用语义分割模型识别高分辨影像中建筑和阴影对象;然后借助建筑对象及其阴影信息在卫星成像时的几何关系估算建筑高度;最后基于多尺度图像分析思想,提取一系列表征建筑对象的高度、空间结构、几何等多尺度特征,利用机器学习方法进行建筑类型分类,并进一步分析不同粒度的建筑类型分析单元对分类结果的影响。选取福州市主城区国产高分二号高分辨率影像进行实验验证。结果表明：① 基于所提方法的建筑类型分类总体精度达到82.98%, kappa系数为0.77,分类精度优于本文中未加入高度信息的分类方法和单一尺度分类方法;② 引入高度特征有效提高了中低层居民楼和高层商住两用建筑类型的分类精度,较未加入高度特征的分类结果,总体精度提高了11.28%;③ 融合多个尺度的图像特征可有效减少粘连建筑误分为密集型建筑的情况,较单一尺度分类方法,总体精度提高了2.77%。在精细的数字表面模型数据缺失下,利用高分辨影像阴影信息可为建筑物高度估计提供一种有效的策略,提高城市建筑类型分类精度。此外,融合多粒度图像特征可提升城市区域复杂建筑类型的表征能力,进而提高分类精度。

Integrating Height Features for Multi-scale Urban Building Type Classification from High-Resolution Remote Sensing Images

Urban building type information is crucial to many urban applications such as the identification of urban functional areas and estimation of urban environmental variables. This paper presents a new method to extract urban building types using multi-scale features and integrating height features derived from high resolution remote sensing images. We first conduct an image semantic segmentation to extract building and shadow objects from remote sensing images, and then estimate the height of buildings based upon the directional relationship of a building object and its shadow information. Following multi-scale image analysis concept, we extract a series of multi-scale features regarding the height, geometry, and spatial structure of building objects. Last, we use a machine learning method based upon random forest to classify building types. We also analyze the impact of different spatial units of building types on classification results. Experiments were conducted in Fuzhou, Fujian province, China, using a Chinese GF-2 satellite images acquired on February 18, 2020. Our results show that: (1) The overall accuracy of building type classification combined with multi-scale features reached 82.98%, and the kappa coefficient was 0.77, which was better than other conventional methods, namely a Multi-scale Classification Without Height Features (MC_NH), a Single-scale Classification Incorporating Height Features (SC), and a Single-scale Classification Without Height Features (SC_NH) in this paper; (2) The classification accuracy of middle-low residential buildings and high-rise commercial and residential buildings was improved by adding height features. Compared with classification results without using height features, the overall accuracy was improved by 11.28%; (3) The fusion of image features at multiple scales can reduce the misclassification of adjacent buildings into dense buildings. Compared with a single-scale classification method, the proposed method improved overall accuracy by 2.77%. We conclude that the use of high-resolution remote sensing images provides an effective strategy to estimate building heights based upon shadow information and improves the classification accuracy of urban building types, particularly when detailed digital surface model data are absent. In addition, the fusion of multi-scale image features can improve the characterization of complex building types in urban areas and the subsequent classification accuracy accordingly. Nevertheless, we also observed that the results of classified building types were affected by the initial extraction of building information from high resolution remote sensing images, implying that a further improvement of building type classification can be done by improving the extraction methods, e.g., using a more advanced semantic segmentation model.