Urban building blocks mapping through a generalizable trained classifier in Google Earth Engine platform

Over the past three decades, with the rapid development of spatial-based satellite imagery, remote sensing technology has found a special place in various applications of urban management. Production of status maps of urban structures, the study of energy loss status, identification of thermal islands, monitoring of urban vegetation, and assessment of air pollution are just a few examples of areas related to urban management that remote sensing technology is the basis for indirect measurement of the related quantities. Maps of urban structures such as building blocks are commonly used in crisis management, urban design, and urban development studies.
 Materials

In this study, the production of urban building block maps using Sentinel 1 and 2 satellite images has been conducted. Normalized Difference Vegetation Index (NDVI) and Normalized Difference Building Index ( NDBI ) for three consecutive months, the slope feature derived from the 30-meter Shuttle Radar Topographic Mission (SRTM)Digital Elevation Model of the study area, along with two Vertical – Vertical (VV) and Vertical - Horizontal ( VH ) polarization in both ascending and descending orbits, form the set of input features.

The proposed method of this paper relies on the use of a generalizable trained classifier. Initially, the classifier is trained in 2015 using training samples obtained from a new rigorous refining process using different remote sensing and spatial products. This rigorous refining process uses a reference urban map of 2015. In the first step, the corresponding areas related to the ways and roads are removed using the OpenStreetMap data layer. Areas suspected of vegetation with NDVI greater than 0.2 are then discarded. Also, due to the high backscattering of buildings in Synthetic Aperture Radar images, areas with a value less than the average backscattering coefficient of the remaining areas are eliminated. Finally, the residual map is refined using the Mahalanabis distance and the Otsu automatic thresholding method. The trained classifier is then used to generate a map of building blocks at similar time intervals for the three target years (2018, 2019, and 2020). Due to the diversity of texture and density of building blocks in the metropolis of Tehran, the proposed method has been evaluated in this area. Due to the concentration of political, welfare, and social facilities, Tehran has experienced more unplanned and irregular expansion and urbanization than other cities in Iran, which has lead to changes in buildings and constructions. Also, due to the availability of free satellite images and various online processing operations, the Google Earth Engine platform has been used in this study. The performance of three different classifiers including Random Forest (RF), Minimum Mahalanabis Distance (MD), and Support Vector Machines (SVM) are examined in this process. In order to evaluate the proposed method, reference samples obtained from visual interpretation of high-resolution satellite images (Google Earth) in all three target years have been used.

The performance of the aforementioned classifiers has been investigated using 3 different criteria: overall accuracy, user accuracy, and F-score of building blocks. The RF method with an overall accuracy of over 93% in all three target years has shown the best performance. The SVM method ranks second with an accuracy of about 91% every three years. However, the MD method with an overall accuracy below 85% in all three target years has not performed well.

The results show better performance of the RF method in all three target years with an overall accuracy of over 93%. It should be noted that the MD classifier with higher user accuracy than other methods, has shown better performance in detecting the class of building blocks. However, the RF method is the best classifier in terms of the user accuracy of the background class. The effect of using two VV and VH polarization and also the slope derived from the SRTM Model in the input feature set on the final accuracy of classification was also investigated. According to the results, the simultaneous use of these three features produces more accurate results in both target classes. However, the results show that the use of VV polarization increases the final classification accuracy compared to VH polarization. The presence of slope feature along with both polarizations has also increased the classification accuracy of each class, especially the background class. However, the exclusion of both VV and VH features from the input feature set has resulted in a more than 10% reduction in overall classification accuracy.

Based on calculated overall accuracies which are above 80% in the majority of investigated cases, two different results can be concluded. First, the trained classifier has shown good temporal generalization and has achieved acceptable accuracy in the target years. Second, due to the different collection processes of training and evaluation data, the proposed rigorous refining method for the preparation of training data has shown good performance. The effect of using two VV and VH polarization and also the slope derived from the SRTM  Digital Elevation Model in the input feature set on the final accuracy of classification was also investigated. According to the results, the simultaneous use of these three features produces more accurate results in both target classes. However, the results show that the use of VV polarization increases the final classification accuracy compared to VH polarization. The presence of slope feature along with both polarizations has also increased the classification accuracy of each class, especially the background class. However, the exclusion of both VV and VH features from the input feature set has resulted in a tangible decreasein overall classification accuracy.