Toward accurate mapping of 30-m time-series global impervious surface area (GISA 2.0)

As an important indicator of urbanization, accurate and long-term global artificial impervious surface area (ISA) monitoring is vital to biodiversity, water quality assessment, urban heat island, etc. However, the existing several 30-m global ISA datasets exhibit large inconsistencies, due to their differences in training samples and mapping methods. In this context, we proposed a global ISA mapping method by considering the inconsistency of the existing products, based on which we further generated a new 30-m global ISA dataset (GISA 2.0). Specifically, we divided the mapping area into A-Grids and M-Grids in terms of their consistency degrees. An automatic mapping method was proposed for classifying the A-Grids, by extracting training samples from the consistent regions of existing datasets. In the case of M-Grids, where the existing ISA datasets showed large inconsistency, we proposed to add manually interpreted samples, to strengthen the classification in these areas. We randomly selected over 120,000 test samples from 207 global grids. The results showed that GISA 2.0 achieved a F1-score of 0.935, better than GISA 1.0 (0.893), GAIA (0.721) and GAUD (0.809). A further assessment based on 118,822 ZY-3 test samples indicated that the overall accuracy and F1-score of GISA 2.0 outperformed the existing ones.