A New Approach for ‘Local Climate Zones’ Classification Using Remote Sensing Images Combining Neighborhood and Sub-Pixel Information

Abstract

The local climate zones (LCZ) divide urban landscapes into 17 distinct classes based on land cover, urban structures, and building materials. In the original World Urban Database and Access Portal Tools (WUDAPT) workflow, random forest (RF) machine learning classifies LCZ at a resolution of 100 m using resampled Landsat-8 images. This study classified LCZs for three mega cities: Madrid in Spain, Rome in Italy, and Chicago in USA using more advanced techniques such as convolutional neural networks (CNN). A total of 6 schemes (S1-S6) were tested for LCZ classification. S1 uses the standard WUDAPT RF classifier. In S2-S3, neighborhood information of single pixels such as mean, maximum, minimum, median, and standard deviation in a moving window is added to the S1 variables extending the S1 feature set. S2 uses a 3x3 kernel while S3 uses a 7x7 kernel. The novel idea of this study is using 10×10 sub-pixel segments in one LCZ pixel of 100 m resolution. The original 30 m Landsat bands were bilinearly resampled to 10 m resolution, so there could be a total of 100 sub pixels in each 100 m pixel. S4 adopts a CNN classifier to LCZ classification. The 100 pixels of 10 m resolution fed into the CNN as an input image, which was allocated to one LCZ class. S5 is also a sub-pixel based approach, but uses the RF classifier. For each spectral band, five statistical features (i.e., mean, maximum, minimum, median, and standard deviation) of the 100 sub pixels in each 100 m resolution pixel were added to the original S1 feature set. S6, the last scheme, uses combining neighborhood and sub-pixel information by the RF classifier. With the feature set of S1, the features of neighborhood information in S2 or S3 and sub-pixel information in S5 were combined. The overall accuracies (OA) of S1 were 74% for Madrid, 63% for Rome and 73% for Chicago, respectively. Compared to the standard approach S1, OAs of S2 and S3 increased for all three cities. Between S4 and S5, both using sub-pixel information, the RF classifier (S5) showed a higher accuracy than the CNN classifier (S4). The final scheme (S6) combining neighborhood and sub-pixel information, showed considerably increased OA (84% for Madrid, 76% for Rome and 88% for Chicago, respectively). The suggested combination approach of neighborhood and sub-pixel information might be effective for LCZ applications in other heterogeneous cities.