Speckle filtering of SAR time stacks via temporal decomposition

Abstract

With the launch of the two Sentinel-1 satellites, very dense synthetic aperture radar (SAR) time series have become available. These new time series allow to develop novel speckle supression algorithms. We propose a new speckle filter for hypertemporal SAR image stacks which only operates in the temporal domain. The filtering in the temporal domain fully preserves the spatial resolution. The backscatter time series of a resolution cell is the superposition of subsignals with differing temporal frequency. Under the assumption, that the speckle is changing on every acquisition, the high frequency parts of the signal correspond to speckle. We are using the empirical mode decomposition (EMD), which is a data-driven alternative to Fourier analysis, to decompose the signal into these subsignals with differing frequencies. In contrast to the Fourier analysis, the decomposition does not depend on a priori set base functions. Therefore, the hypertemporal SAR time series is decomposed into components of different frequency which are derived in a fully data-adaptive manner. The high frequency part of the signal is removed to obtain a smoothed SAR time series. The lower frequency sub signals are related to actual biophysical processes and are similar on neighbouring pixels of equal land cover. Therefore, the smoothing over time results also in a spatial smoothing over homogeneous areas. We compare the results of our new approach with the results by the Quegan filter with 3 x 3 and 5 x 5 spatial windows. For the comparison of the spatial smoothing we compute the equivalent number of looks (ENL) over homogeneous forest areas. The ENL is the number of different images which would be needed to obtain the same smoothing by multilooking. With the EMD filter, we obtain a similar ENL compared to the 5x5 Quegan filter and an improved ENL compared to Quegan 3x3. To compare the edge preservation we chose two homogeneous regions with different average backscatter and computed the width of the edge between those regions after applying different filtering approaches. We showed that the EMD filter improves edges in real-world imagery, while the Quegan filter introduces a certain amount of blurring.