Abstract
Abstract. Mapping the exact extent of urban areas is a critical prerequisite in many remote sensing applications, such as hazard evaluation and change detection. The usage of Synthetical Aperture Radar (SAR) data has gained popularity due to the unique characteristics of the backscattered radio signal from human-made targets. The Sentinel-1 (S1) constellation, with a global revisit time of 6–12 days in Interferometric Wide Swath (IW) mode and free and open access to the data, allows the development of new applications to monitor urban sites. However, S1 is rarely considered when fine resolution is required due to the large pixel size and the need for spatial averaging to obtain robust estimators. We propose a method to improve Sentinel-1 urban classification performance by exploiting one Multi-Spectral (MS) image acquired by Sentinel-2 (S2). MS data is used for tracing the precise natural boundaries in a scene through superpixels segmentation. A machine learning approach is then applied to interpret the thematic context of each segment from short temporal stacks of coregistered SAR data. We use a short sensing period (around two months), so rapid changes can be traces. The proposed fusion of S1 and S2 data was tested in the area of Milan (Italy), with a total accuracy of about 90%. The ability to follow high-resolution details in a mixed environment is demonstrated, opening the possibility of efficiently tracing the human footprint.
Highlights
In recent decades, mapping urban areas and change have been vital in facing numerous environmental and cadastral challenges
The approach characterizes objects defined in the optical domain, which are less prone to noise and are available in higher resolution, with features from the Synthetical Aperture Radar (SAR) domain, which are more robust in differentiating urban areas
Efficiency is gained by selecting a set of computationally simple features and the dimension reduction introduced by segmentation
Summary
In recent decades, mapping urban areas and change have been vital in facing numerous environmental and cadastral challenges. SAR data holds great potential in urban mapping because of the sensed signal’s nature, which is very much different between urban and natural areas. While overcast weather limits the possibility of processing continuous time-series using MS data, radio signal penetrates clouds, allowing regular sampling worldwide. The launch of several high-resolution missions made SAR a strong candidate for the task (Esch et al, 2017), especially where cloud coverage prevents regular time sampling by optical means. The work presented here explores the possibility of combining the unique SAR capabilities of detecting stable targets, i.e., urban structures, with the fine resolution of optical surveys. The classification products are given as the percent of urban pixels in a segment, and the value is thresholded to achieve binary classification
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
