Repeat-pass SAR interferometry for land cover classification: A methodology using Sentinel-1 Short-Time-Series

• Published in: Remote sensing of environment Vol. 232; p. 111277
• Main Authors: Sica, Francescopaolo, Pulella, Andrea, Nannini, Matteo, Pinheiro, Muriel, Rizzoli, Paola
• Format: Journal Article
• Language: English
• Published: New York Elsevier Inc 01.10.2019; Elsevier BV
• Subjects: Algorithms; Backscatter; Backscattering; C band; Classification; Data acquisition; Forests; Interferometric coherence; Interferometry; Land cover; Land cover classification; Land use; Learning algorithms; Machine learning; Parameter estimation; SAR; Sentinel-1; Synthetic aperture radar interferometry; Target recognition; Temporal decorrelation; Thematic mapping; Time series; Urban areas; Interferometric coherence; Land cover classification; SAR; Sentinel-1; Temporal decorrelation
• ISSN: 0034-4257; 1879-0704
• DOI: 10.1016/j.rse.2019.111277

In this paper we explore the potential of repeat-pass interferometric SAR (InSAR) for land cover classification purposes. We introduce a novel approach for the generation of large-scale thematic maps, based on the use of multi-temporal data, acquired over short observation intervals (short-time-series). A larger interferometric coherence loss is expected with the increasing time difference between two interferometric acquisitions. This phenomenon is normally indicated as temporal decorrelation whose amount differs depending on the type of imaged target on ground. The basic idea is therefore to accurately model the evolution in time of the temporal decorrelation and use the estimated parameters, together with backscatter, as input features for the Random Forest machine learning classification algorithm. The work has been carried out on the case study of land cover mapping over central Europe, considering Sentinel-1 C-band interferometric stacks, acquired over a time span of about one month. Three different land cover classes have been considered: artificial surfaces as e.g. urban areas, forests, and non-forested areas as the ensemble of low vegetation, bare soil, and agricultural areas. The results show a level of agreement above 91%, when compared to the CORINE land cover map product of 2012, which has been used as external reference for both training and testing of the classification algorithm. •We present a novel approach for land classification from multi-temporal InSAR data.•Temporal decorrelation and backscatter are used for a machine learning classifier.•The framework is tested with Sentinel-1 data over a region in Europe.•The results demonstrate the potential of InSAR stacks for land cover mapping.•The approach is suitable for the generation of large-scale land cover maps.