Abstract
Urbanization has attracted wide and active interests due to the impact on regional sustainable development. As an important indicator of urbanization, impervious surface area (ISA) should be accurately monitored. In scenario of identifying ISA by supervised classification from satellite images, the training samples are usually labeled manually, which is highly labor-intensive and time-consuming. High-resolution nighttime light image provides a unique footprint of human activities and settlements which are strongly correlated with ISA. In view of this, a novel ISA training sample selection method is proposed by integrating the JL1-3B high-resolution nighttime light imagery and Sentinel-2 time series imagery, and the random forest is applied to classify ISA from Sentinel-2 imagery. The quality of the automatically selected samples was quantitatively validated. There were over three study areas, and the overall classification accuracies were above 97%, showing reliable and robust performance. Compared with conventional methods, the proposed approach achieves satisfactory results in separating bare land from ISA. This study provides a data fusion way which can automatically generate sufficient and high-quality training samples for ISA mapping, and suggests that high-resolution nighttime imagery could demonstrate a promising potential for urban remote sensing.
Highlights
U RBANIZATION is regarded as an essential trend of the Earth’s terrestrial surface changes
The FROM-GLC10 impervious surface area (ISA) product was extracted from FROM-GLC10 as it has a category of impervious surface
With the proposed sample selection scheme, by the use of iterative processing and the time series spectral information and nighttime light imagery (NTL) brightness information, it is efficient to automatically select reliable and diverse training samples, which ensure the accuracy of classification results
Summary
U RBANIZATION is regarded as an essential trend of the Earth’s terrestrial surface changes. The generation of training samples is usually through screen digitalization in high-resolution imagery manually, in which the operator should have a comprehensive knowledge of land cover types for the study area [13]. Though the above datasets are global products, they could not satisfy the current requirements of urban monitoring due to the low spatial resolution and limited accuracy in local region These datasets have strong limitations in dynamic monitoring especially when cities are experiencing rapid urban sprawl, as it can only provide the distribution of ISA for certain years. To achieve this aim, an automatic sample selection method for reliable training-samples generation is developed first, and random forest (RF) model is established for ISA extraction based on features derived from Sentinel-2 optical data.
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
