The Temporal Dynamics of Slums Employing a CNN-Based Change Detection Approach
The Temporal Dynamics of Slums Employing a CNN-Based Change Detection Approach
复制标题
DOI：
10.3390/rs11232844
复制
发表时间：
2019-12-01
期刊：
影响因子：
5
通讯作者：
Persello, Claudio
中科院分区：
文献类型：
--
作者：
Liu, Ruoyun;Kuffer, Monika;Persello, Claudio
沿着快速城市化，贫民窟的增长和持续存在是一个全球性挑战。虽然遥感图像越来越多地用于制作贫民窟地图，但只有少数研究分析了其时间动态。本研究探讨了全卷积网络（FCNs）的潜力，以分析在印度班加罗尔使用甚高分辨率（VHR）图像的临时贫民窟的小集群的时间动态。该研究开发了两种基于FCNs的方法。第一种方法使用分类后变化检测，第二种方法训练FCN直接对贫民窟的动态进行分类。对于这两种方法，比较了3 x 3内核和5 x 5内核网络的性能。虽然个别年份的分类结果显示出平均88.4%的相对较高的F1分数（3x 3内核），但变化准确度较低。分类后的结果获得了53.8%的F1分数，变化检测网络获得了53.7%的F1分数。根据轨迹误差矩阵（TEM），后分类结果得分较高的整体准确性，但较低的变化轨迹的准确性差异比变化检测网络。虽然这两种方法在准确性方面没有显着差异，但变化检测网络的噪声较小。在我们的研究区域内，贫民窟的面积总体上略有下降;贫民窟的年增长率（2012年至2016年）为7173米（2），而年下降率为8390米（2）。然而，这些数字掩盖了更大的空间动态。有趣的是，贫民窟消失的地区通常会变成绿色地区，而不是建成区。拟议的变化检测网络提供了一个强大的地图的位置变化的准确边界的置信度较低。这显示了FCN在检测VHR图像中贫民窟动态方面的潜力。
Along with rapid urbanization, the growth and persistence of slums is a global challenge. While remote sensing imagery is increasingly used for producing slum maps, only a few studies have analyzed their temporal dynamics. This study explores the potential of fully convolutional networks (FCNs) to analyze the temporal dynamics of small clusters of temporary slums using very high resolution (VHR) imagery in Bangalore, India. The study develops two approaches based on FCNs. The first approach uses a post-classification change detection, and the second trains FCNs to directly classify the dynamics of slums. For both approaches, the performances of 3 x 3 kernels and 5 x 5 kernels of the networks were compared. While classification results of individual years exhibit a relatively high F1-score (3 x 3 kernel) of 88.4% on average, the change accuracies are lower. The post-classification results obtained an F1-score of 53.8% and the change-detection networks obtained an F1-score of 53.7%. According to the trajectory error matrix (TEM), the post-classification results scored higher for the overall accuracy but lower for the accuracy difference of change trajectories than the change-detection networks. Although the two methods did not have significant differences in terms of accuracy, the change-detection network was less noisy. Within our study area, the areas of slums show a small overall decrease; the annual growth of slums (between 2012 and 2016) was 7173 m(2), in contrast to an annual decline of 8390 m(2). However, these numbers hid the spatial dynamics, which were much larger. Interestingly, areas where slums disappeared commonly changed into green areas, not into built-up areas. The proposed change-detection network provides a robust map of the locations of changes with lower confidence about the exact boundaries. This shows the potential of FCNs for detecting the dynamics of slums in VHR imagery.