Classification of the global Sentinel-1 SAR vignettes for ocean surface process studies

Highlights

• First deep learning model to classify ten geophysical phenomena from S-1 WV SAR data.

• Model performance is evaluated using an independent eye-selected dataset.

• Classified rain cells and sea ice are compared with other satellite measurements.

• The global S-1 SAR data show great potential for sea surface processes studies.

Abstract

Spaceborne synthetic aperture radar (SAR) can provide finely-resolved (meters-scale) images of ocean surface roughness day-or-night in nearly all weather conditions. This makes it a unique asset for many geophysical applications. Initially designed for the measurement of directional ocean wave spectra, Sentinel-1 SAR wave mode (WV) vignettes are small 20 km scenes that have been collected globally since 2014. Recent WV data exploration reveals that many important oceanic and atmospheric phenomena are also well captured, but not yet employed by the scientific community. However, expanding applications of this whole massive dataset beyond ocean waves requires a strategy to automatically identify these geophysical phenomena. In this study, we propose to apply the emerging deep learning approach in ocean SAR scenes classification. The training is performed using a hand-curated dataset that describes ten commonly-occurring atmospheric or oceanic processes. Our model evaluation relies on an independent assessment dataset and shows satisfactory and robust classification results. To further illustrate the model performance, regional patterns of rain and sea ice are qualitatively analyzed and found to be very consistent with independent remote sensing datasets. In addition, these high-resolution WV SAR data can resolve fine, sub-km scale, spatial structure of rain events and sea ice that complement other satellite measurements. Overall, such automated SAR vignettes classification may open paths for broader geophysical application of maritime Sentinel-1 acquisitions.

Introduction

The spaceborne synthetic aperture radar (SAR) is a well-established technique to collect high-resolution sea surface backscatter data during day and night in most weather conditions. Over the ocean, SAR images provide an estimate of the sea surface roughness primarily through backscattering of short waves (Alpers et al., 1981; Hasselmann et al., 1985; Hasselmann and Hasselmann, 1991), where this small-scale (cm) roughness responds to the near-surface ocean winds (Lehner et al., 2000; Winstead et al., 2006; Mouche et al., 2012). In addition, these short waves are also modulated by ocean swell (Heimbach et al., 1998; Lehner et al., 2000; Collard et al., 2009), upper ocean processes (Johannessen et al., 1996; Rascle et al., 2017; Jia et al., 2018), and atmospheric phenomena (Alpers and Brümmer, 1994; Young et al., 2005; Winstead et al., 2006; Li et al., 2007, 2013; Alpers et al., 2016). Beginning with SEASAT in 1978, ocean SAR imagery has been widely used to examine numerous air-sea interaction processes (Meadows et al., 1983; Gerling, 1986; Carsey and Holt, 1987; Fu and Holt, 1982; Katsaros and Brown, 1991). Since then, ever-improving SAR data have been obtained by satellite missions that include ERS-1/2, Envisat/ASAR, RADARSAT-1/2, TerraSAR-X, TanDEM-X and Sentinel-1 constellation.

However, global-scale applications of ocean SAR data remain quite limited. This is largely because the wide swath SAR images are not routinely collected over the open ocean. These acquisitions mainly focus on land, Arctic regions, and near the coasts. Thus, most previous ocean SAR data investigations only involve limited regional or single SAR scene case study (Alpers and Brümmer, 1994; Babin et al., 2003; Sikora et al., 2011; Li et al., 2013; Alpers et al., 2016). One exception is the wave mode (WV) dedicated to retrieving ocean wave proprieties at global scale (Kerbaol et al., 1998; Stopa et al., 2016). The WV has been developed for ERS-1/2 (1991–2003) and Envisat/ASAR (2002–2012), and now introduced to Sentinel-1 (2014-present) and Gaofen-3 (2016-present). It normally collects relative small SAR images (typically 5–10 km square) along the orbit with a distance of about 100 km in between. This is sufficient for ocean wave spectrum retrieval and empirically estimation of the total significant wave height (Heimbach et al., 1998; Collard et al., 2009; Stopa and Mouche, 2017), which can be used in wave forecasting. At present, the routine WV measurements are only available from the Sentinel-1 (S-1) A&B (Torres et al., 2012). It was improved upon Envisat and ERS by having finer spatial resolution (4 m), higher signal-to-noise (which reduces speckle noise), larger scene footprint (20 by 20 km), and increased global sampling.

Wang et al. (2019) demonstrated that the S-1 WV dataset has the potential for new studies on air-sea interactions at scales of 0.5–10 km. The primary advantage of the S-1 WV dataset is its ability of measuring high resolution sea surface roughness globally (~120k images per month). However, without an automated means to identify the geophysical features captured by each image, the potential would remain untapped. For example, previous studies have relied solely on visual inspection to identify SAR images with wind streaks before performing statistical analysis or surface wind direction derivation (Lehner et al., 2000; Levy, 2001; Mouche et al., 2012; Zhao et al., 2016). Such manual classification approach is impractical for the huge volume of S-1 WV data. Similarly, dedicated classic machine learning algorithms have mostly been developed for specific applications such as detection of oil spills and ships. These methods depend on the empirically hand-crafted features, which are usually insufficient to generalize the local variations, shapes and structural patterns of different geophysical phenomena (Topouzelis and Kitsiou, 2015; Zhang et al., 2016).

This study attempts to train a deep convolutional neural network (CNN) to classify the ten prescribed geophysical phenomena seen in WV vignettes. Deep CNN models have been applied with great success in detection, segmentation, and recognition of objects, features, and textures within digital images (LeCun et al., 2015). They have also been applied to hyperspectral and optical remote sensing imagery (Zhao and Du, 2016; Li et al., 2017; Hu et al., 2015; Cheng and Han, 2016; Zhou et al., 2017). However, the primary use of CNN in ocean SAR application has mostly been for target recognition (Zhang et al., 2016; Zhu et al., 2017). In general, CNN is a multilayer architecture that can be trained to automatically extract the optimal image features and to amplify distinctions between images (LeCun et al., 2015; Zhang et al., 2016). A practical and effective way to develop a robust CNN for a specific application is to re-train an existing image recognition model. This so-called transfer-learning or fine-tuning strategy has been proven to be more efficient and practical than creating and training a new CNN architecture from scratch in the case of limited database (Yosinski et al., 2014; Zhu et al., 2017; Cheng et al., 2017; Too et al., 2018; Wang et al., 2018a).

In this paper, we adapt the Inception-v3 CNN (Szegedy et al., 2015) to train a model dedicated to the classification of S-1 WV vignettes, called CMwv. The involved datasets are described in section 2. Section 3 demonstrates the training process of CMwv and illustrates the model performance based on an independent assessment dataset. In section 4, we compare our classification results qualitatively with rain precipitation from Global Precipitation Measurement (GPM) and sea-ice concentration from Special Sensor Microwave Imager (SSM/I). Conclusions follow in section 5.