Technical Framework of Feature Extraction Based on Pixel-Level SAR Image Time Series

Abstract

This study proposes a novel technical framework of feature extraction based on pixel-level synthetic aperture radar (SAR) image time series, to exploit the application potential of SAR image data with low and medium spatial resolution. This framework comprises three key parts: 1) construction of the pixel-level SAR image time series using a new matching technique based on progressive binary partition; 2) pixel-level similarity measurement via dynamic time warping (DTW); and 3) a new spatiotemporal similarity analysis method that improves feature extraction by considering both the similarity of a feature’s pixel-level time series and its spatial correlation. Two locations, covered by 31 low-resolution (150 m) and 26 medium-resolution (30 m) ENVISAT ASAR images, respectively, were selected as test cases to validate the proposed framework. Results show that the framework can identify features with a high level of accuracy, completeness, and correctness, outperforming methods using multitemporal images, as well as the time series-only (nonspatial) method, and other methods of spatiotemporal similarity analysis that use alternative similarity measures.