基于高分辨率遥感影像的农村房屋自动化识别

Automatic identification of rural houses based on high-resolution remote sensing images

关键词:  高分辨率遥感影像  农村房屋  支持向量机集成方法  自动化识别农业

Key words:  High-resolution remote sensing imagery  rural houses  SVM ensemble approach  automated identification

摘要:

【目的】 为了改善遥感影像分类算法对“同谱异物”的农村房屋与乡村道路的区分能力，提高房屋识别的空间平滑性与自动检测精度，为后续的农村宅基地遥感监测提供技术支撑。【方法】 文章基于光谱-空间核函数集成支持向量机（光谱-空间核SVM）算法框架，发展了一种适于高分辨率遥感影像的农村房屋自动化识别方法。首先，对高分辨率遥感影像进行空间分割以及影像的空间邻域关系进行建模。其次，获取分割图斑内像元灰度值的均值作为空间特征，以像元灰度值作为光谱特征，利用光谱-空间核SVM算法进行房屋预提取。单独提取影像中的道路、裸地等区域，并制作掩膜对房屋预提取结果进行修正。最后，通过众数滤波，对房屋识别结果进行空间平滑处理，抑制噪声。【结果】 该试验得到的Kappa系数、总体精度和F1分数分别为0.75、98.0%和0.76；而基于像元光谱的常规识别方法得到的上述指标分别为0.40、91.8%和0.44。【结论】 该文提出的方法能有效抑制将道路、裸地识别为房屋，且具有良好的空间平滑性能。

Abstract:

[Purpose] In the context of the incoming remote sensing monitoring of rural homestead，this study is aimed at improving algorithms of remote sensing images classification to relieve ambiguities between rural houses and roads，which have similar spectral characteristics，and to improve the smoothness and accuracy of automatic identification of rural houses.[Method] Based on the spectral-spatial support vector machine（SVM）ensemble framework，an automated method of rural house identification using remotely sensed images of high spatial resolution（high-resolution）was developed in this study. Specifically，the presented method mainly includes the following four steps. Firstly，spatial neighborhood relationships within the applied images were modeled by image segmentation，using the algorithm of entropy rate superpixel segmentation. Secondly，the spectral-spatial features of each patch is extracted. The average of grayscale values of all pixels with a patch is used as its spatial signatures，and the grayscale values of the current pixel is used as spectral signatures. And then，a preliminary extraction of houses is performed based on the spectral-spatial features of each patch，using the spectral-spatial kernel SVM algorithm. The spectral-spatial features is composed by an ensemble approach that uses a weighted liner combination of kernel functions to realize the comprehensive utilization of multiple features，for identifying houses and avoiding potential misclassification. Thirdly，using the aforementioned approach，roads and bare ground are separately extracted，and made to be a mask to correct the results of houses extraction，for further eliminating misclassification. At last，a plural filtering method is designed to spatially smooth the results of houses identification and suppress noises. This is realized by iterating over the whole classification map with a gliding window. The number of pixels of "house" and "other" within the gliding window are counted separately. The class with more number of pixels was considered to be the class of the center pixel of the gliding window. An experiment was undertook in Dahua Town，Meishan City，Sichuan Province of China. The experimental area presets a typical landscape of rural area. A high spatial resolution multispectral image collected with a unmanned aerial vehicle（UAV）were applied in this experiment.[Result] The Kappa coefficient，overall accuracy and F1 score obtained by using the aforementioned method are0.75，98.0% and 0.76 separately. Comparatively，the same indicators are 0.40，91.8% and 0.44，obtained using conventional method that based merely on spectra of individual pixels. In addition，according to the resulting map of house identification，The proposed approach improves the ability to distinguish between geo-objects with similar spectra，such as rural houses，cement roads，and arid bare soil，and also enhances the spatial smoothness of classification results.[Conclusion] The result confirms the effectiveness of the automated rural house recognition strategy based on high-resolution remote sensing images proposed in this study，and therefore，it is considered to have the potential to provide important support on vital data and methods for relevant decision-making departments in the future digital management tasks of rural homesteads.