Journal of Geo-information Science ›› 2018, Vol. 20 ›› Issue (9): 1306-1315.doi: 10.12082/dqxxkx.2018.170289

Previous Articles     Next Articles

Supervised Dense Rural Residential Extraction from High-resolution Remote Sensing Images Based on Automatically Augmentation of Training Samples

LU Chen1,2(), YANG Xiaomei1,*(), WANG Zhihua1   

  1. 1. State Key Laboratory of Resources and Environment Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2017-12-27 Revised:2018-06-20 Online:2018-09-25 Published:2018-09-25
  • Contact: YANG Xiaomei;
  • Supported by:
    National Key Research and Development Program of China, No.2016YFC1402003;Natural Science Foundation of China, No.41671436, 41421001


Compared with the unsupervised methods of residential areas extraction, the supervised methods are of relatively higher accuracy. However, the supervised methods rely on large amounts of training samples, and manually labeling residential areas is tedious and time-consuming, limiting their applications in the era of remote sensing big data. In order to improve application performances of the extraction methods based on supervised classification, it is necessary to overcome the disadvantage that training samples need to be manually labeled with high accuracy. Dense rural residential areas composed of single-family building blocks are the predominant type of rural residential areas in North China Plain. In this paper, we set extraction targets to be the dense rural residential areas, and propose a novel extraction method for high-resolution remote sensing images. The proposed method utilizes a supervised classifier but only one positive sample labeled manually is required. Firstly, four exclusion rules are designed based on the features of rural residential areas in high-resolution remote sensing images. According to the exclusion rules, all of the image blocks are classified into two categories of residential areas or non-residential areas. After the coarse classification, a certain number of the negative samples are randomly selected from the image blocks belonging to the category of non-residential areas. Then one positive sample is labeled manually, and more positive samples are collected from the image blocks in the neighborhood of the only one positive sample by performing the nearest neighbor classifier. At last, the K-Nearest Neighbor classifier is adopted to pick up image blocks which are closer to more positive samples in the feature space. The classifier filters image blocks belonging to the category of non-residential areas from the coarse extraction result, then the final extraction result is obtained. Experimental results conducted on the test images confirm that the proposed approach is both efficient and robust to images with different backgrounds.

Key words: rural residential area composed of single-family building blocks, extraction methods, automatically augmenting training samples, high-resolution remote sensing images, exclusion rules