ABSTRACT

Informal settlements are now becoming an unavoidable and growing phenomenon in the global south, whose locations are inappropriate in official statistics and maps. The distinctive spectral signature of informal settlements in high-resolution optical images causes difficulties in finding automatic approaches. This paper investigates the Multi-Shape Multi-Size Morphological Profile-Guided filter (MShMSiMP-GF) approach for detecting the informal settlements from high-resolution optical images. The development of the proposed approach was inspired by the Superpixel-based Guided Filter (SGF) approach, which uses superpixels to construct the guidance image. Though the superpixel based guidance image extracts detailed contextual information (scale, size), it fails to model geometrical information (shapes, structures) in an image. In order to incorporate the geometrical information, the MShMSiMP-GF approach generates multi-shape multi-size morphological profiles with increasing radius of guided filters. Quantitative and qualitative results of the proposed approach are investigated by four different images of Madurai city, India acquired by Kompsat-2 and WorldView-2 sensors. From the classified maps, it is observed that the proposed MShMSiMP-GF approach achieves an overall accuracy of 91.37%, 90.19%, 93.46% and 99.36% for the subsets 1, 2, 3 and 4, respectively.

摘要

非正式定居点现在正在成为全球南方不可避免且日益严重的现象，其位置在官方统计数据和地图中并不合适。高分辨率光学图像中非正式住区的独特光谱特征导致难以找到自动方法。本文研究了用于从高分辨率光学图像中检测非正式住区的多形状多尺寸形态学轮廓引导滤波器 (MShMSiMP-GF) 方法。所提出方法的发展受到基于超像素的引导滤波器（SGF）方法的启发，该方法使用超像素来构建引导图像。尽管基于超像素的引导图像提取了详细的上下文信息（比例、大小），但它无法对图像中的几何信息（形状、结构）进行建模。为了合并几何信息，MShMSiMP-GF 方法生成多形状多尺寸形态轮廓，引导滤波器的半径增加。通过 Kompsat-2 和 WorldView-2 传感器获取的印度马杜赖市的四张不同图像，研究了所提出方法的定量和定性结果。从分类地图中可以看出，所提出的 MShMSiMP-GF 方法对子集 1、2、3 和 4 的总体准确率分别为 91.37%、90.19%、93.46% 和 99.36%。印度被 Kompsat-2 和 WorldView-2 传感器收购。从分类地图中可以看出，所提出的 MShMSiMP-GF 方法对子集 1、2、3 和 4 的总体准确率分别为 91.37%、90.19%、93.46% 和 99.36%。印度被 Kompsat-2 和 WorldView-2 传感器收购。从分类地图中可以看出，所提出的 MShMSiMP-GF 方法对子集 1、2、3 和 4 的总体准确率分别为 91.37%、90.19%、93.46% 和 99.36%。