SAR analysis of wetland ecosystems: Effects of band frequency, polarization mode and acquisition dates

Abstract

The availability of high spatial resolution synthetic aperture radar (SAR) sensors with a wide range of acquisition modes has increased greatly over the past decade and contributed significantly to the study of wetland ecosystems. However, the relative influence of acquisition configurations (i.e. band frequency, polarization mode, number of acquisition dates) in wetland analysis remains poorly explored. This article investigated the relative influence of X-/C-band frequency, dual-/quad-polarization and single-/multiacquisition features on discrimination of vegetation types in 632 ha Ramsar-protected temperate riverine marshes (Mont-Saint-Michel Bay, France). Three SAR datasets (i.e. quad-pol/C-band, dual-pol/C-band and dual-pol/X-band) were generated from five pairs of TerraSAR-X and RADARSAT-2 images. First, a set of 25 SAR features, including backscattering coefficients and polarimetric parameters, was extracted from the SAR datasets. Second, correlation between each pair of images was calculated using the polarimetric parameter Shannon entropy to select the most similar pairs in the time series. Third, the importance of each SAR feature and modeling accuracy were calculated using a conditional random forest model for each of the three datasets. Finally, analysis of variance was performed to assess the impact of band frequency, polarization mode and number of acquisition dates on the classification of vegetation types. The results highlighted that although the time-shift of each pair of TerraSAR-X and RADARSAT-2 images was short (3-11 days), only three pairs were sufficiently similar, highlighting the high variability in wetland ecosystems. The polarimetric parameter Shannon entropy was the most discriminating feature, regardless of the frequency or polarization. Most variance in the model accuracy was explained by the number of acquisition dates (68%), followed by the frequency (23%), while polarization explained little. This article will help select the most suitable SAR sensor acquisition modes for wetland conservation.

On the Role of Polarimetric Decomposition and Speckle Filtering Methods for C-Band SAR Wetland Classification Purposes

“…The NL-SAR, which gave the best results in the segmentation of watercourses images located in marshes [20], also in our study provided high classification accuracy. The most often used filter in wetlands classification is the refined Lee filter [4], [15] with a 5×5 window [2] or a 7×7 window [5]. Several studies applied the boxcar filter with a 5×5 [7], [16], [17] window size, or the improved Lee sigma filter with a 3×3 [3] or 5×5 [1] window.…”

“…I N RECENT years, many studies focused on the application of different synthetic aperture radar (SAR) data in wetlands mapping using various acquisition setups. The studies showed an impact on the classification accuracy of the SAR band, polarization mode, multiple acquisition dates [1]- [5], and incidence angles of the beam [6], [7]. At the same time, the selection of the polarimetric processing parameters of the SAR images was rarely justified, while only several studies demonstrated that these parameters influence the workflow.…”

“…In speckle filtering of full polarimetric images in wetland areas, the most often used methods were the refined Lee filter [4], [15] with a 5×5 [2] or 7×7 window [5], the boxcar filter with a 5×5 window [7], [16], [17], and the improved Lee sigma filter with a 3×3 [3] or 5×5 window [1]. These filters reduce speckle based on the local window, while a recent approach is to use nonlocal filters, which were useful in coastal studies [18], [19] and floodplain boundary preservation [20].…”

“…Many studies have confirmed the applicability of polarimetric SAR data in wetland classification. The Shannon entropy parameter provided the best classification results and was the most discriminating SAR feature regardless of the polarization mode and frequency [5], [15], [24]. Several other studies showed that the Cloude-Pottier decomposition, which also decomposes Shannon entropy, was an important classification feature [6], [25].…”

Coastal Wetland Classification with GF-3 Polarimetric SAR Imagery by Using Object-Oriented Random Forest Algorithm

“…Some studies that used fully polarimetric SAR imagery have noted the influence of different polarimetric scattering features on wetland classification results [ 26 , 27 ]. For example, Millard et al combined SAR and Lidar data to achieve higher accuracy [ 28 ].…”