Automated Discrimination of Certain Brightness Fronts in RADARSAT-2 Images of the Ocean Surface

Chris T. Jones Department of Mathematics and Statistics, Dalhousie University, Halifax, Nova Scotia, Canada

Search for other papers by Chris T. Jones in
Current site
Google Scholar
PubMed
Close
,
Todd D. Sikora Department of Earth Sciences, Millersville University, Millersville, Pennsylvania

Search for other papers by Todd D. Sikora in
Current site
Google Scholar
PubMed
Close
,
Paris W. Vachon Defence Research and Development Canada—Ottawa, Ottawa, Ontario, Canada

Search for other papers by Paris W. Vachon in
Current site
Google Scholar
PubMed
Close
,
John Wolfe Defence Research and Development Canada—Ottawa, Ottawa, Ontario, Canada

Search for other papers by John Wolfe in
Current site
Google Scholar
PubMed
Close
, and
Brendan DeTracey Fisheries and Oceans Canada, Bedford Institute of Oceanography, Dartmouth, Nova Scotia, Canada

Search for other papers by Brendan DeTracey in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

Automated classification of the signatures of atmospheric and oceanic processes in synthetic aperture radar (SAR) images of the ocean surface has been a difficult problem, partly because different processes can produce signatures that are very similar in appearance. For example, brightness fronts that are the signatures of horizontal wind shear caused by atmospheric processes that occur independently of properties of the ocean (WIN herein) often appear very similar to brightness fronts that are signatures of sea surface temperature (SST) fronts (SST herein). Using Moderate Resolution Imaging Spectroradiometer (MODIS)-derived SST for validation, 302 SAR SST and 193 SAR WIN signatures were collected from over 250 RADARSAT-2 images of the Gulf Stream region using a Canny edge detector. A vector consisting of textural and contextual features was extracted from each signature and used to train and test logistic regression, maximum likelihood, and binary tree classifiers. Following methods proven effective in the analysis of SAR images of sea ice, textural features included those computed from the gray-level co-occurrence matrix for regions along and astride each signature. Contextual features consisted of summaries of the wind vector field near each signature. Results indicate that signatures labeled SST can be automatically discriminated from signatures labeled WIN using the mean wind direction with respect to a brightness front with an accuracy of between 80% and 90%.

Corresponding author address: Chris T. Jones, 95 Central Avenue, Fairview NS B3N 2H7, Canada. E-mail: jonesc@mathstat.dal.ca

Abstract

Automated classification of the signatures of atmospheric and oceanic processes in synthetic aperture radar (SAR) images of the ocean surface has been a difficult problem, partly because different processes can produce signatures that are very similar in appearance. For example, brightness fronts that are the signatures of horizontal wind shear caused by atmospheric processes that occur independently of properties of the ocean (WIN herein) often appear very similar to brightness fronts that are signatures of sea surface temperature (SST) fronts (SST herein). Using Moderate Resolution Imaging Spectroradiometer (MODIS)-derived SST for validation, 302 SAR SST and 193 SAR WIN signatures were collected from over 250 RADARSAT-2 images of the Gulf Stream region using a Canny edge detector. A vector consisting of textural and contextual features was extracted from each signature and used to train and test logistic regression, maximum likelihood, and binary tree classifiers. Following methods proven effective in the analysis of SAR images of sea ice, textural features included those computed from the gray-level co-occurrence matrix for regions along and astride each signature. Contextual features consisted of summaries of the wind vector field near each signature. Results indicate that signatures labeled SST can be automatically discriminated from signatures labeled WIN using the mean wind direction with respect to a brightness front with an accuracy of between 80% and 90%.

Corresponding author address: Chris T. Jones, 95 Central Avenue, Fairview NS B3N 2H7, Canada. E-mail: jonesc@mathstat.dal.ca

1. Introduction

Images of the ocean surface captured by spaceborne synthetic aperture radar (SAR) have long been known to exhibit spatial variations in surface roughness that correspond to gradients in sea surface temperature (SST) (Vesecky and Stewart 1982). Because SAR penetrates cloud cover, it represents a potential asset to the Canadian Forces Meteorology and Oceanography Centre (MetOc), Halifax, Nova Scotia, Canada, in its effort to provide a continual assessment of the location of tactically important SST fronts that are often obscured by clouds in passive infrared radiometer images. As part of the research and development phase of the Spaceborne Ocean Intelligence Network (SOIN) project, the objective of our research was to develop procedure to identify SST front signatures in RADARSAT-2 SAR images acquired in the vicinity of the Gulf Stream, with particular interest in automated identification of the Gulf Stream north wall (GSNW).

A SAR image reveals spatial variations in the small-scale roughness of the ocean surface (i.e., centimeter-scale waves that induce Bragg scattering) forced by near-surface winds and modulated by longer waves, currents, and surfactants (Holt 2004). There is often unstable atmospheric surface layer stratification on the warm side of an SST front, corresponding to a positive buoyancy flux within the marine atmospheric boundary layer (MABL), manifested by convective processes that transport momentum from the upper MABL toward the surface. Moreover, the atmospheric temperature gradient often present across an SST front can result in a corresponding pressure gradient in the atmosphere (Small et al. 2008). These processes, referred to as the momentum mixing mechanism and the sea level pressure driving mechanism, respectively (Song et al. 2006), enhance near-surface horizontal wind speed on the warm side of the front, with concomitant intensification of surface roughness compared to the cold side. Sweet et al. (1981) collected empirical data suggesting that momentum mixing contributes most to the enhancement of near-surface winds, whereas model results in Wai and Stage (1989) suggest that the sea level pressure gradient contributes most. This subject remains an area of active research, so it is unclear which of these processes, or an altogether different process (e.g., a thermal wind effect; Foster et al. 1999), is usually dominant. Nevertheless, we can say that, individually or in combination, these processes sometimes generate a brightness front in a SAR image that is the signature of an SST front (Sikora et al. 1995; Beal et al. 1997; Holt 2004).

Figure 1 shows two contiguous RADARSAT-2 ScanSAR Narrow A (SCNA) mode image frames acquired at C band with VV polarization on 7 March 2009 at approximately 2230 UTC. The maximum-pixel composite Moderate Resolution Imaging Spectroradiometer (MODIS) SST image in Fig. 2 is from data acquired between 1300 and 1900 UTC on the same day. The complex SST front evident along 39.5°N in Fig. 2 demonstrates that there was energetic horizontal mixing between the cold water mass on the shelf and the warm water mass on the shelf slope. A second SST front is evident along the eastern portion of the image, where an intrusion of cold shelf water penetrated the shelf slope water mass to flow against the eastern side of a meander in the Gulf Stream. The signatures of these fronts are unmistakable in Fig. 1.

Fig. 1.
Fig. 1.

Two contiguous RADARSAT-2 SCNA VV frames acquired on 7 Mar 2009 at approximately 2230 UTC. SST front signatures (indicated by SST) appear as well-defined brightness fronts in the SAR image that match fronts in the MODIS image in Fig. 2. Signatures of pure horizontal wind shear can be seen in the lower portion of the SAR image, identifiable as brightness fronts misaligned with fronts in the MODIS image (indicated by WIN). Other labeled features are described in the text. (RADARSAT-2 data and products ©2009MacDonald, Dettwiler and Associates Ltd.—all rights reserved.)

Citation: Journal of Atmospheric and Oceanic Technology 30, 9; 10.1175/JTECH-D-12-00190.1

Fig. 2.
Fig. 2.

Composite MODIS SST acquired between 1300 and 1900 UTC on the same day as the RADARSAT-2 frames in Fig. 1.

Citation: Journal of Atmospheric and Oceanic Technology 30, 9; 10.1175/JTECH-D-12-00190.1

Examples such as these motivate the pursuit of our objective. The momentum mixing and the sea level pressure driving mechanisms are, however, only two of many processes that render signatures in SAR images of the ocean surface. Signatures of atmospheric internal gravity waves, for example, are evident in Fig. 1 as a series of alternating dark and bright bands in the region of the continental shelf (cf. Thompson et al. 1992). Signatures resembling those produced by the entrainment of surfactants (Ivanov and Ginzburg 2002) and the enhancements of surface roughness in convergence zones associated with strong currents (Lyzenga and Marmorino 1998) are evident as dark and bright filaments, respectively, in the region of the shelf slope. Also evident is a dark signature of areal surfactant accumulation in the lee of the Gulf Stream meander (near 37.5°N, 70°W), where cold water laden with phytoplankton, or possibly sargassum (Beal et al. 1997), had accumulated against the GSNW. Signatures of horizontal wind shear are evident in the southern portion of the image, appearing as brightness fronts not aligned with nearby SST fronts: they provide a good example of how certain signatures of mesoscale variability in the atmosphere can mimic or obscure SST front signatures (cf. Song et al. 2006).

With so many different signatures occurring at the same time, devising an automated classification algorithm for SAR images of the ocean surface is challenging. In a previous effort (Jones et al. 2012), we used a Canny edge detector (Canny 1986) to identify candidate SST front signatures in a set of RADARSAT-2 SCNA VV images, and used logistic regression to assign to each its probability of being an SST front signature. The regression was based on a set of textural and contextual measures associated with each Canny edge and surrounding pixels. Based on its assigned probability, each candidate was classified either as an SST front signature or the signature of some other process. Coincident MODIS and Advanced Very High Resolution Radiometer (AVHRR) SST images were used to validate the classification. A case study, focused on the lower portion of the scene within Fig. 1 herein, demonstrated that the resulting algorithm was unable to discriminate between signatures of SST fronts and those of certain pure atmospheric processes.1

The signatures of some processes can be readily identified by visual inspection of a SAR image. The signatures of atmospheric and oceanic internal gravity waves, for example, can be recognized because they typically occur in trains, and often appear as parallel bands of alternating large and small backscatter in the case of atmospheric internal waves (Vachon et al. 1995), and sometimes as concentric arcs of the same in the case of oceanic internal waves (Apel 2004). Signatures of mesoscale and microscale convection are quite common and can also be easily identified (Young and Winstead 2005). And SST front signatures can sometimes be identified with high confidence without a validating SST image, for example, by the presence of textures that indicate horizontal gradients of static stability in the MABL (Sikora and Ufermann 2004). In the absence of such textures, however, a signature of horizontal wind shear generated by a pure atmospheric process (referred to herein as pure horizontal wind shear or WIN) can easily be confused for an SST front signature (e.g., Fig. 1).

In this study we aimed to identify a set of textural and contextual measures that can be used to automatically discriminate SAR signatures of SST fronts from those of pure horizontal wind shear that are similar in appearance. A noteworthy addition to the methodology of our previous effort is the construction of a local curvilinear coordinate system (LCCS) around each Canny edge, motivated by the extensively studied effect of SST gradients on wind vectors (e.g., Chelton and Xie 2010). Wind vectors projected onto the LCCS provide a means to estimate average wind speed along and across a Canny edge, and the LCCS itself facilitates the extraction of textural measures from either side of the edge. The LCCS also provides a means of quantifying the divergence of winds in the cross-front direction that can be promoted by the momentum mixing and sea level pressure–driving mechanisms, which are a contextual measure we hypothesized would provide a means of discriminating between signatures of SST fronts and signatures of pure horizontal wind shear.

2. Data collection

a. Canny edge validation

A total of 1227 RADARSAT-2 (MacDonald Dettwiler and Associates Ltd., http://www.radarsat2.info/) SCNA VV images were collected by MetOc Halifax from October 2008 to November 2009. We were restricted to use either VV or HH polarization, and because C-band VV images tend to be slightly less noisy than their HH counterparts at incidence angles greater than about 30° (Holt 2004), and edge detection generally works better when there is less noise (Canny 1986), we chose to use VV. Each SAR image covered a 300 km × 300 km area of the ocean surface in the region of the Gulf Stream with the incidence angle ranging from 20° to 46°. In our previous study, we used model-assimilated Hybrid Coordinate Ocean Model (HYCOM, http://hycom.org/) data to identify a GSNW search region (GSNWSR), an envelope in which the GSNW was determined to mostly likely occur (Jones et al. 2012). Only images that overlapped the GSNWSR (indicated by black curves in Fig. 3) were employed for subsequent analysis, of which there were 252.

Fig. 3.
Fig. 3.

Geographic location of the center of each of the 495 Canny edges collected from 252 RADARSAT-2 images is shown as dark-gray circles (SST) and light-gray squares (WIN). Two lines indicate the GSNWSR.

Citation: Journal of Atmospheric and Oceanic Technology 30, 9; 10.1175/JTECH-D-12-00190.1

SAR images were processed using Image Analyst Profressional (IA Pro) (Secker and Vachon 2007), developed by Defense Research and Development Canada (DRDC), to produce images proportional to the amplitude of the radar backscatter (the SAR processor provides amplitude A, which has a smaller dynamic range and is generally less noisy than intensity = A2). Processing included the application of a land mask, a special filter to remove targets of opportunity such as ship signatures, and block averaging to reduce pixel spacing from 25 m (12 000 × 12 000 pixels) to 300 m (1000 × 1000 pixels). Each image was also radiometrically flattened by fitting and subtracting a two-dimensional third-degree polynomial to remove systematic variation in backscatter in both the range and azimuth directions. In a separate processing stream, SAR-derived wind speed (SDW) images were computed using the Institut Français de Recherche pour l'Exploitation de la Mer (Ifremer) version 2 backscatter model relating C-band VV ocean backscatter to wind speed, relative wind direction, and incident angle (Vachon and Dobson 2000). Wind directions were provided by nearest-in-time Quick Scatterometer (QuikSCAT) level 3 data (25 km in resolution) obtained from the National Aeronautics and Space Administration (NASA)'s Physical Oceanography Distributed Active Archive Center (PO.DAAC) website (http://podaac.jpl.nasa.gov/). All subsequent processing was performed using routines written in MATLAB.

A Canny edge detector was applied to each processed SAR image to produce candidate edges. To reduce the candidates to a manageable number, only edges greater than 80 km in length were considered. Edges were then manually compared to MODIS daytime SST averaged over 8 days and 4 km, obtained using the PO.DAAC Ocean Federation of Earth Science Information Partners (ESIP) Tool (http://poet.jpl.nasa.gov/), to determine whether they could be assigned a label, SST or WIN. When comparing a SAR image with its SST counterpart, we often found that some brightness fronts matched the location, orientation, and shape of an SST front. In such cases the brightness fronts were labeled as SST. Brightness fronts that did not match SST fronts were designated to be candidate WIN signatures. To label a signature as WIN with a reasonable level of confidence, we first ruled out other common possibilities. These included bright filaments consistent in appearance to signatures of current shear, brightness fronts that appeared to be part of a train of atmospheric or oceanographic internal waves, and brightness fronts that could be associated with atmospheric convection cells.

To illustrate, the signatures labeled WIN in Fig. 1 are perpendicular to the SST front to the immediate east. A process independent of SST likely generated these signatures. They do not appear to be signatures of any of the enumerated alternatives, and were therefore labeled as WIN signatures with confidence. Admittedly, this kind of manual evaluation is to some degree prone to the uncertainties of subjective judgment. But we believe, having viewed more than 1000 images, including a large number of cases for which WIN signatures could be identified with high confidence (e.g., when supported by surface analysis charts), that this approach is valid. Edges from the WIN/SST pool that could not be labeled with a reasonable level of confidence were not included in the analysis. In total, 495 Canny edges were labeled and considered validated, 302 of which were assigned the label SST and 193 WIN. The geographic locations of the edges are shown in Fig. 3.

It is standard practice to select pixels randomly for validation, for example, in studies using SAR images of vegetative cover where individual pixels are classified (Tso and Mather 2009). The purpose of this is to randomize the effects of unknown and unaccounted for variables in the classifier. In the case of WIN and SST, validation is predicated on the availability of coincident SST data, and is therefore not random. Figure 3 suggests that nonrandom selection did not result in a bias with respect to geographic location; the 495 validated Canny edges appear to be nonsystematic, if not random, in their spatial distribution.

However, there is evidence of a temporal bias in the distribution of the validated edges, as indicated in Fig. 4a. The bar plot, scaled in counts on the right-hand axis, demonstrates a seasonal bias in the distribution of both WIN and SST signatures. This bias is not entirely due to seasonal variations in cloud cover. Estimates of mean percentage clear sky for the region, determined using images of monthly global cloud cover climatology from the International Satellite Cloud Climatology Project (http://isccp.giss.nasa.gov/), are indicated in Fig. 4a by the line plot scaled in percent on the left-hand axis. Although the monthly frequency of SST signatures is highly correlated with percentage clear sky from January to June (ρ = 0.89, p value = 0.017), the correlation during the remainder of the year is not significant. Interestingly, the opposite holds for the monthly frequency of WIN signatures, which is not correlated with percentage clear sky during the first six months of the year, but is from July to December (ρ = 0.85, p value = 0.030).

Fig. 4.
Fig. 4.

(a) Monthly frequencies (scale on the right) of SST (dark gray) signatures, and to some extent WIN (light gray) signatures, are evidently seasonally dependent. This is only partially explained by the monthly-mean percentage clear sky indicated by the line graph scaled on the left. (b) Frequencies suggest that while validated signatures of both types are unlikely to occur when wind speed exceeds about 10 m s−1, validated SST front signatures are relatively likely to occur even at wind speeds below 3 m s−1.

Citation: Journal of Atmospheric and Oceanic Technology 30, 9; 10.1175/JTECH-D-12-00190.1

Cloud cover impacts the ability to validate SST signatures. Other factors may impact how readily signatures can be detected, and thereby contribute to the bias in the monthly frequencies of WIN and SST signatures. These include seasonal variations in wind speed, SST gradient, and phytoplankton biomass. Figure 4b shows the frequency of SST and WIN signatures as a function of mean SDW in the vicinity (within about 10 km) of each signature. Validated SST and WIN signatures occur infrequently at wind speeds greater than about 10 m s−1, and this may partially account for the reduction in the frequency of both signature types during the winter months when the average near-surface wind speed is greatest (the mean wind stress in the region of the Gulf Stream is about 0.15–0.20 N m−2 in January and falls to about 0.05 N m−2 in July; Risien and Chelton 2008). We say partially because, as already mentioned, prevalent cloud cover precludes the ability to validate Canny edges on images acquired during the winter months and therefore also contributed to the paucity of wintertime signatures. SST gradients tend to be greatest during the winter and smallest during the summer in the region of the Gulf Stream (Armstrong et al. 2012). SST front signatures are generally easier to identify when the gradient is more pronounced, but this may be offset by stronger winds in the winter. Surfactant concentrations generated by phytoplankton biomass, because they occur in the region of cooler water to the immediate north of the Gulf Stream but not in the Gulf Stream itself, can enhance the contrast of a signature of the GSNW, thereby making its detection more likely. Surfactant concentrations are generally seasonal: phytoplankton blooms occur during the spring and autumn (Yoder et al. 2002).

We speculate that from April to June, when mean wind speed is below its winter maximum, phytoplankton concentration is high, and SST gradients have not yet reached their summer minimum; this may be a time period when signatures of the GSNW are more likely to be apparent, thus partially explaining the high counts of validated SST front signatures during those months. The decline in validated SST signature frequency during July and August, when mean percent cloud cover is near its minimum, may be explained by a reduction in both phytoplankton biomass and the mean SST gradient. But separating the influences of these contributing factors would be difficult at best, and to test the proposed hypothesis would require a comprehensive multivariate dataset. We leave this to a future effort. The potential impact of the unbalanced temporal distribution of the validated Canny edges will be discussed at the end of the results section.

b. Local curvilinear coordinate system

An LCCS was constructed for each Canny edge, as depicted in Fig. 5. Each edge was first interpolated to have a regular spacing of 300 m along its length. Cross-front transects with the same spacing were then constructed at each point along the edge. Each transect was made to be normal to the edge at the point of intersection (i.e., at grid point gi, the transect was made to be normal to the line segment extending from gi to gi+1) and extend to a distance of 12 km to either side. This distance was chosen based on the assumption that it would be in the region nearest to the front that discriminating measures would be most pronounced, (i.e., within tens of kilometers; Park et al. 2006; Song et al. 2006).

Fig. 5.
Fig. 5.

LCCS constructed for each Canny edge was used for the extraction of contextual measures, such as the mean wind direction with respect to the front (μϕ), as well as various textural measures computed for pixels on the three regions labeled darker, front, and brighter. Note that the scale of the grid shown is coarser than the 300-m resolution that was actually used in the analysis.

Citation: Journal of Atmospheric and Oceanic Technology 30, 9; 10.1175/JTECH-D-12-00190.1

SAR and SDW images were projected onto the LCCS using nearest-neighbor interpolation to produce local SAR (locSAR) and local SDW (locSDW) images. Nearest-in-time QuikSCAT wind data were bilinearly interpolated onto the LCCS to provide wind directions. Alongfront wind (locAW) and cross-front wind (locCW) images were determined by projecting locSDW with the interpolated wind directions onto the cross-front and alongfront components of the LCCS at each grid point, and local divergence (locDIV) and local curl (locCRL) images were computed by projecting the divergence and curl of the SDW onto the LCCS. Each local image was segmented into three regions: a 6.6-km-wide band centered on the front and two 8.7-km-wide bands extending from 3.3 to 12 km on either side of the front. These regions will be referred to hereafter as the darker, front, and brighter regions of the LCCS, as depicted in Fig. 5.

c. Textural measures

A textural measure quantifies some aspect of the spatial distribution of the tonal values of a grayscale image. Such measures can be obtained using methods based on either spatial domain or frequency domain analysis (Tso and Mather 2009). Previous classification studies using a spatial domain approach based on the gray-level co-occurrence matrix (GLCM; Haralick 1973) have shown promising results in the segmentation of SAR images (Barber and LeDrew 1991; Shokr 1991). We therefore decided to try the same methodology.

The approach starts with the computation of the GLCM usually from a window around each image pixel. Assuming a b-bit image, is the joint probability distribution for all pairs of gray levels (i, j) ∈ {1, …, 2b} × {1, …, 2b} that occur with a specified geometric arrangement. A vector consisting of an ordered pair D = (row offset, column offset) specifies the geometry, where offsets are expressed in pixels. Textural measures computed from capture various aspects of the joint probability distribution that can be used to inform a statistical classifier. For example, using a Sea Ice and Terrain Assessment SAR image of a scene of the Canadian Arctic, Barber and LeDrew (1991) computed GLCM measures: contrast, correlation, energy, dissimilarity, and entropy in 25 × 25 pixel windows. Using linear discriminant analysis, they assigned classes—open water, first-year ice, multiyear ice, and land—to each image pixel. The resulting classification skill, expressed by the κ coefficient (described below), was about κ = 0.70.

In contrast to the cited example, our intention was not to compare textures of individual pixels but to compare the darker, front, and brighter regions of the LCCS. To that end, the textural measures—contrast, correlation, energy, dissimilarity, entropy, and homogeneity—were computed from the GLCM obtained from each of the three regions of the LCCS for all of the local images, each expressed as a 6-bit image. The formulas for the textural measures used are listed in Table 1, where Pij denotes the ith row and jth column of , and the mean and standard deviation of the marginal distributions (i.e., the row and column sums) of are denoted by μx and σx, and μy and σy, respectively.

Table 1.

Measures computed from the GLCM.

Table 1.

d. Contextual measures

To test the hypothesis that SST front signatures can be identified by the divergence of wind vectors in the cross-front direction (i.e., the cross-front gradient of the cross-front wind component), the divergence dcw of the wind in the direction of each cross-front transect along a Canny edge was computed using the locSDW (with interpolated QuikSCAT wind directions). The angle ϕ between the SDW vector and the Canny edge was also computed at each LCCS grid point along the Canny edge. This angle was constructed such that ϕ = 0° when the wind vector is parallel to the Canny edge (in the locality of the corresponding grid point), and 90° when the wind vector is perpendicular to the Canny edge.

e. Feature vectors

A feature vector x consisting of 148 components was constructed for each validated Canny edge, as indicated in Table 2. For example, the first feature of x indicated in Table 2 consists of the measure contrast computed from the darker region of the locSAR image (i.e., a feature is defined to be a textural or contextual measure computed on one region of a specified local image). The textural measures listed in Table 1 were determined by computing the average of values obtained using an offset of one pixel in the four cardinal and four intercardinal directions [i.e., D in {(1, 0), (1, 1), (0, 1), (−1, 1), (−1, 0), (−1, −1), (0, −1), (1, −1)}] based on an assumption of isotropy. In addition, means and standard deviations were computed for pixels in all three regions of the local images. The mean and standard deviation of the cross-front divergences, and the same for the angles ϕ between the wind vector and points along the length of the Canny edge, were also computed. The feature vectors were compiled in a 495 × 148 dimensional matrix and, to account for differences in the dynamic ranges of the features, each column of the matrix was divided by its standard deviation prior to analysis.

Table 2.

Schematic of the 148 features tested, consisting of (top portion) 144 textural features and (bottom portion) four contextual features from the locSDW image. For example, the first feature consists of the measure contrast computed from the darker region of the locSAR image. Symbols and abbreviations are explained in the text.

Table 2.

3. Classifiers

The general objective of supervised classification is to partition a k-dimensional feature space into one region for each class. Region boundaries are determined using a set of validated feature vectors, called the training data, in such a way as to minimize classification error. In the current context where there are only two classes, the objective reduces to the production of an optimal separation of the feature space into two half-spaces. The shape of the optimal separating boundary is constrained by the geometry of the particular classifier being used. As different geometries may produce different results, classifiers based on three methods—logistic regression, maximum likelihood analysis, and binary tree analysis—were tested. The following is a brief description of each method.

a. Logistic regression classifier

Logistic regression is a method specific to two classes. It assumes a direction of maximum probability gradient in feature space, specified by a linear combination of the features, such that the probability of one of the two classes (the reference class) is monotonic and follows a sigmoid function (i.e., the logistic function) along the gradient. The probability that a particular location x in feature space corresponds to the reference class is given by
e1
Equation (1) defines contours of equal probability in feature space, and the contour corresponding to prob{ref.class | x, α} = 0.5 is usually taken to be the optimal separating boundary. The polynomial function f(x, α) with its vector of coefficients α determines the location and shape of the boundary.

b. Maximum likelihood classifier

The maximum likelihood classifier is based on the assumption that the feature vectors for each class follow a multivariate normal distribution in k-dimensional feature space, with a different mean vector and covariance matrix for each class. Hence, for the ith class, x ~ N(μi, Σi). For a given location x in feature space, the assigned class is the one that minimizes the equation
e2
On the right-hand side of Eq. (2), dist(x, μi) is the Mahalanobis distance between x and the mean μi of the ith class. The term ln|Σi| is a measure of total variance and can be thought of as a penalty term for classes that are poorly localized in feature space. In the case of two classes, WIN and SST, the contour of equal likelihood on which points x in feature space satisfy ln|ΣSST| + dist(x, μSST) = ln|ΣWIN| + dist(x, μWIN) constitutes the optimal separating boundary.

c. Binary tree classifier

The formulations in Eqs. (1) and (2) both preclude the possibility of one class being concentrated in two or more regions of feature space, thus constraining the probability distribution of each class to have only one mode. A binary tree classifier generally partitions feature space into an array of boxes, and therefore allows the possibility that a class consists of two or more localized modes. With this extra level of complexity, a binary tree classifier may potentially outperform the logistic regression and maximum likelihood classifiers. The optimal separating boundary for the binary tree classifier consists of a set of conjoined truncated hyperplanes aligned with the coordinate axes.

4. Results

a. Univariate analysis

Each feature was considered individually to determine its effectiveness in discriminating between the two classes, WIN and SST. Only one classifier (logistic regression) was used for this purpose, as all three methods were expected to produce similar results in the univariate case. For each feature, the following cross-validation procedure was repeated 300 times. First, 120 feature vectors, 60 from each class, were randomly selected without replacement and set aside as test data. The logistic regression classifier was then trained on the unselected feature vectors. The resulting trained classifier was applied to the test data to assign to each case the probability that it was an SST front signature. All cases for which the assigned probability was more than 0.50 were classified as SST, the remainder as WIN.

For each trial, classification accuracies were calculated from a contingency table commonly referred to as the confusion matrix. An example is shown in Table 3, where the feature tested was the mean wind direction with respect to the front (denoted by μϕ). The columns in Table 3 consist of the randomly selected test cases partitioned across rows according to the assigned class. The proportion of cases of agreement between class and label in Table 3 is pa = (48 + 39)/120 = 0.73. Although this is an intuitive measure of classification accuracy, it can be misleading in cases where the test data are not balanced. For example, an accuracy of 0.73 could be obtained by either classifying all test cases as SST or assigning class based on a coin toss, if it happened that 73% of the test cases were labeled SST. To account for the possibility of imbalances in the test data, investigators commonly compute the so-called κ coefficient to quantify classification skill.

Table 3.

Example of a confusion matrix resulting from a single trial of the cross-validation procedure. In this case the feature tested was the mean angle of the wind with respect to a Canny edge (μϕ). Cases of agreement between classification (rows) and assigned labels (columns) are indicated on the main diagonal, from which a classification accuracy of (48 + 39)/120 = 0.73 was obtained. Classification skill, which takes into account potential imbalances in the assigned labels that determine the probability of agreement between class and label by chance alone, was κ = 0.46 (i.e., 46% better than random chance).

Table 3.

Let pc be the proportion of cases of agreement between class and label expected by chance alone. The κ coefficient is then defined as the ratio κ = (papc)/(1 − pc), which can be interpreted as the proportion of cases of agreement in excess of that expected by chance alone (Cohen 1960). In the current situation, where the test data were purposely balanced, pc = 0.5, the a priori probability that a randomly selected test case was labeled SST (or WIN). In this case, the expression for the κ coefficient simplifies to κ = 2(pa − 0.5): computing from Table 3, κ = 2(0.73 − 0.5) = 0.46. Herein we report both the skill and accuracy of our classifier to make our results readily comparable to other classification studies that used only one or the other.

Table 4 shows the mean κ coefficient (second column) and accuracy (third column) determined from 300 trials for features that resulted in the largest values. The number of trials controls the magnitude of the margin of error for each mean: 300 trials were purposely chosen to reduce the margin of error (ME) for an approximate 99% confidence interval to be no more than 0.010 for the κ coefficient, corresponding to 0.0050 for accuracy. The single most informative feature was μϕ, with mean κ = 0.59, corresponding to a mean accuracy of 0.80. Of the remaining features, only certain textural measures computed from the locCW image proved to contain discriminating information, but all resulted in mean κ coefficients no greater than 0.45. Furthermore, these features were correlated with μϕ, with correlation coefficients |ρ| ≥ 0.70 in most cases.

Table 4.

Cross validation with 300 trials as described in the text was used to produce the mean classification skill (κ coefficient) and mean accuracy for each feature in a univariate logistic regression classifier. The number of trials was sufficient to ensure that the approximate 99% ME was no more than 0.010 for mean skill and 0.0050 for mean accuracy.

Table 4.

b. Bivariate analysis

Having identified x1 = μϕ, the mean wind direction with respect to a detected brightness front, as the most informative feature, the next step in the analysis was to determine whether a second feature would result in a significant increase in the mean κ coefficient. For this purpose, all three methods of classification—logistic regression, maximum likelihood analysis, and binary tree analysis—were tried. Using the same cross-validation procedure as in the univariate analysis, classifiers containing μϕ plus one other feature were tested. It was found that x2, the standard deviation of pixels on the locSAR image in the front region, resulted in the largest improvement in mean κ over that obtained from the univariate analysis.

The distribution of the two classes in two-dimensional feature space is shown in Fig. 6, where features were normalized by their standard deviations to account for differences in their dynamic ranges, and also log transformed to make them more normal in distribution. The thicker curve in Fig. 6 is the mean decision boundary from 300 trials produced by the maximum likelihood classifier. The corresponding mean κ coefficient is 0.63 (mean accuracy = 0.82). This is statistically but only marginally better than mean κ = 0.60 obtained using the logistic regression classifier, the mean decision boundary of which is the thinner curve in Fig. 6. The decision boundary consisting of two orthogonal line segments is the mean that resulted from the binary tree classifier, which matched the mean skill and accuracy of the maximum likelihood classifier. These results are summarized in Table 5.

Fig. 6.
Fig. 6.

Feature vectors labeled SST (dark-gray circles) and WIN (light-gray squares) are fairly well separated in the two-dimensional subspace consisting of features x1 and x2. Each feature was normalized by its standard deviation (sd) to account for differences in dynamic range, and log transformed to make points more normal in distribution. Curved and straight lines are mean decision boundaries, each computed from 300 trials as described in the text.

Citation: Journal of Atmospheric and Oceanic Technology 30, 9; 10.1175/JTECH-D-12-00190.1

Table 5.

All bivariate classifiers containing μϕ together with one other feature were tested using the same cross-validation procedure as in the univariate analysis. Results indicated that the best second feature was x2 (defined in text), which provided a marginal improvement in the mean skill and accuracy. Maximum likelihood and binary tree classifiers resulted in significantly greater classification skill and accuracy compared to the logistic regression classifier. Approximate ME for a 99% confidence interval for the means is given.

Table 5.

c. Multivariate analysis

Although the maximum likelihood and binary tree classifiers resulted in the same mean κ coefficients, the binary tree classifier was chosen because it is comparatively simple and does not depend on assumptions of normality (i.e., it is nonparametric). To determine whether including additional features in the binary tree classifier would result in an improvement in the mean κ coefficient, each feature was combined with x1 and x2 and tested as previously described. None of the three-feature classifiers resulted in a statistically significant improvement in the mean skill over that of the bivariate classifier. It was therefore concluded that none of the features contained discriminating information not already captured by x1 and x2.

d. Error analysis

The binary tree classifier applied to all of the data produces a κ coefficient of 0.63 and accuracy of 0.82 (these are not mean values because they were obtained by applying the classifier to all of the data at once), and we know of no other study that approaches this level of success for the task at hand. Furthermore, it is possible that classification skill may be improved by taking into account certain properties of the dataset not yet included in the analysis, such as the temporal imbalance in the data across months (see Fig. 4a), the lag between the acquisition times of each pair of RADARSAT-2 and QuikSCAT images, and wind speed in the vicinity of each Canny edge.

A possible consequence of the unbalanced distribution of the validated Canny edges across months is a classifier that exhibits greater skill when applied to signatures that occur during the months containing more edges. To test this, the skill of the binary tree classifier in Fig. 6 was computed for each month separately (note that it is essential to use classification skill here so as to take into account the variability in the ratio of WIN to SST labels across months—see explanation of κ coefficient). The correlation between the monthly κ coefficient and the number of Canny edges collected was then computed and found to be statistically insignificant (ρ = 0.18, p value = 0.57). This result indicates that the unbalanced temporal distribution of the validated edges had no significant impact on monthly classification skill.

The mean absolute temporal separation tlag between the acquisition times for the RADARSAT-2 images and their nearest-in-time QuikSCAT images was 38 min, with a standard deviation of 19 min. That the orientation of near-surface winds with respect to a Canny edge is the key discriminant in the classifier suggests the possibility that the classification skill may be proportional to tlag, which serves as a proxy for inaccuracies in estimates of wind direction. To test this, the mean classification skill of the binary tree classifier in Fig. 6 was computed on bins of tlag centered at 10-min intervals up to 90 min. The correlation between the resulting vector of κ coefficients and the bin centers was found to be statistically insignificant (ρ = 0.03, p value = 0.93). It may therefore be concluded that the temporal separation between acquisition times had no impact on the performance of the classifier, and that temporal lags up to at least 88 min (the maximum in the dataset) are acceptable in the application of the classifier to new data.

The mean SAR-derived wind speed on the LCCS was introduced into the binary tree classifier to determine its impact on classification skill. This brought to light the existence of two sets of anomalous edges. The first consisted of three SST front signatures (but no WIN signatures) corresponding to a wind speed of less than 1.4 m s−1 for which μϕ < 23°, which indicates that the general pattern that signatures labeled WIN are more likely when μϕ < 23° may not be true at very low wind speed (but WIN signatures are uncommon at a wind speed below about 4 m s−1; see Fig. 4b). The second anomalous set consisted of four SST and nine WIN signatures (i.e., 70% WIN) with μϕ between 23° and 32° and a wind speed in excess of 12 m s−1. This indicates that the general pattern that edges labeled SST are more likely when μϕ is at least 23° may break down at very high wind speed. It would seem that accounting for extremes in wind speed might improve the skill of the binary tree classifier. For example, removing the anomalous edges from the analysis and applying the binary tree classifier of Fig. 6 to the remaining data produces a κ coefficient of 0.66 and an accuracy of 0.85 compared to κ = 0.63 and accuracy = 0.82 when the same model is applied to all of the data. However, given the small number of data points concerned (a total of only 16 out of 495 cases), this conclusion is only tentative. More data corresponding to wind speeds greater than 12 m s−1 and less than about 2 m s−1 would be required before any conclusion about the impact of wind speed on classification accuracy could be made with confidence.

5. Discussion

Unlike readily identifiable signatures of processes such as current shear, surfactant accumulation, and atmospheric internal waves in SAR images, the signatures of pure horizontal wind shear are often similar in appearance to those of SST fronts. This observation motivated our objective of finding an automated means to discriminate between signatures labeled SST and WIN. Our initial hypothesis that the divergence of winds in the cross-front direction can provide a means to accomplish this was rejected by the results of the preceding analysis. Instead, it was the mean wind direction with respect to each brightness front that proved to have the greatest discriminating power.

The estimated probability distributions of μϕ for Canny edges labeled SST and WIN are shown in Fig. 7. The single mode of the distribution for Canny edges labeled WIN (gray curve) centered near 15° suggests that pure horizontal wind shear tends to be closely aligned with its brightness front signature (i.e., its Canny edge). Horstmann and Koch (2005) showed that the SAR signatures of wind streaks that occur on a scale of about 200 m tend to be well aligned with the mean surface wind direction. Our results indicate that this might be true for some mesoscale wind signatures as well, meaning that a Canny edge identified as a signature of pure horizontal wind shear can potentially be used to infer wind direction (up to 180° ambiguity) within a certain locality.

Fig. 7.
Fig. 7.

Estimated probability distributions of μϕ for signatures labeled WIN (gray curve) and SST (black curve) exhibit some separation. Vertical line at 23° indicates the optimal decision bound: edges to the left are more likely signatures of pure horizontal wind shear, while edges to the right are more likely SST front signatures. Edges near the boundary where the probability densities are approximately equal are easily misclassified, suggesting that, to reduce classification error, it may be preferable to leave Canny edges that fall near the boundary (within the gray band) unclassified.

Citation: Journal of Atmospheric and Oceanic Technology 30, 9; 10.1175/JTECH-D-12-00190.1

The distribution of μϕ for signatures labeled SST (Fig. 7, black curve) appears to be bimodal, and it is possible that this reflects differences in atmospheric dynamics when the wind blows from the warm side of an SST front toward the cold side as compared to when the wind blows from cold to warm. To test this, edges labeled SST were separated into two groups, cold-to-warm and warm-to-cold advection. A comparison of the group means (43° for cold to warm and 47° for warm to cold) was made using a t test under the assumption of unequal variances after a test indicated that the two groups had different standard deviations (p value < 0.01). This resulted in an approximate 95% confidence interval for the mean difference of (0.53°, 8.3°), indicating that the difference is significant (p value = 0.03). Hence, the bimodality of the distribution for Canny edges labeled SST evident in Fig. 7 is at least partly due to a difference between cold-to-warm and warm-to-cold advection.

Using μϕ alone as a discriminating variable, the optimal decision bound between the two classes occurs at μϕ = 23°, indicated by the vertical line in Fig. 7. The probability densities for the two classes overlap heavily near this bound, and the classification error is therefore expected to be high for Canny edges with μϕ close to 23°. To mitigate the impact of this relationship, μϕ can be partitioned into three regions by expanding the decision bound to cover an interval on which edges would be flagged to indicate that they could not be classified with a reasonable level of certainty. The automated decision rule would take on the following form:
e3
The radii R1 and R2 can be adjusted to enhance the classification skill without producing an excessive number of flagged edges. For example, when the decision rule in Eq. (3) is applied to all of the data with R1 = R2 = 8.75° (indicated by the gray band in Fig. 7), the κ coefficient increases to 0.70, corresponding to an accuracy of 0.89, at the expense of 30% of the edges flagged. The bounds need not be equal. When the left-hand boundary is shifted to the right (R1 = 1.5° and R2 = 8.75°), the skill increases to 0.74, accuracy is 0.88, and only 17% of the edges are unclassified. In an operational setting, edges flagged by the automated procedure might be classified manually by an analyst using all available contextual information, including visual cues in the SAR image commonly associated with SST fronts, previous determinations of the location of the GSNW, MODIS and/or AVHRR SST, surface weather analyses, as well as the classes assigned to other edges on the same image.

The intention of our analysis was to identify textural measures that could discriminate between SST and WIN signatures following the approach used in the analysis of sea ice. Our results suggest that no such measures exist. We see the observation that the mean wind angle with respect to a Canny edge can discriminate between the two signature types as a promising starting point toward an automated algorithm that can accurately classify a broader range of signature types. We have no doubt that there is room for future refinements to our algorithm. The projection of QuikSCAT winds with a resolution of 25 km onto an LCCS only 12.5 km across may have been a source of error in our analysis, for example. It might be possible to reduce such errors by increasing the span of the LCCS. The use of high-resolution model wind data may also engender improvements in classification accuracy, and indeed, with the demise of QuikSCAT, we have turned to the Global Environmental Multiscale model (Côté et al. 1998) to provide near-surface wind vectors. Other means of improvement will undoubtedly arise. Nevertheless, we believe that our algorithm will prove to be useful in the identification of SST front signatures in SAR images of the Gulf Stream.

6. Conclusions

Our objective was to develop an automated procedure to classify certain brightness fronts in SAR images of the ocean surface in the region of the Gulf Stream identified by a Canny edge detector as being either SST front signatures or signatures of horizontal wind shear generated by pure atmospheric processes. Our analysis was based on a set of 144 textural features similar to those used in the segmentation of SAR images of sea ice, and four contextual features extracted from wind vectors and motivated by studies of the effects of SST fronts on near-surface winds. The single most significant discriminating feature was the mean wind direction with respect to a given brightness front. It alone resulted in a mean classification accuracy of 80%, and therefore it provides a basis for semiautomated classification of brightness fronts in RADARSAT-2 images in the region of the Gulf Stream. Furthermore, a decision rule [Eq. (3)] that leaves ambiguous cases unclassified can increase automated classification accuracy to nearly 90%.

Acknowledgments

The authors thank the Canadian Space Agency (CSA) for supporting this study via its Government Related Initiatives Program (GRIP).

REFERENCES

  • Apel, J. R., 2004: Oceanic internal waves and solitons. Synthetic aperture radar marine user's manual, C. R. Jackson and J. R. Apel, Eds., U.S. Department of Commerce, NOAA/NESDIS, 189–206.

  • Armstrong, E. M., Wagner G. , Vazquez-Cuervo J. , and Chin T. M. , 2012: Comparisons of regional satellite sea surface temperature gradients derived from MODIS and AVHRR sensors. Int. J. Remote Sens., 33, 66396651.

    • Search Google Scholar
    • Export Citation
  • Barber, D. G., and LeDrew E. F. , 1991: SAR sea ice discrimination using texture statistics: A multivariate approach. Photogramm. Eng. Remote Sensing, 57, 385395.

    • Search Google Scholar
    • Export Citation
  • Beal, R. C., Kudryavtsev V. N. , Thompson D. R. , Grodsky S. A. , Tilley D. G. , Dulov V. A. , and Graber H. C. , 1997: The influence of the marine atmospheric boundary layer on ERS 1 synthetic aperture radar imagery of the Gulf Stream. J. Geophys. Res., 102 (C3), 57995814.

    • Search Google Scholar
    • Export Citation
  • Canny, J., 1986: A computational approach to edge detection. IEEE Trans. Pattern Anal. Mach. Intell., 8, 679698.

  • Chelton, D. B., and Xie S. P. , 2010: Coupled ocean–atmosphere interaction at oceanic mesoscales. Oceanography, 23, 5269.

  • Cohen, J., 1960: A coefficient of agreement for nominal scales. Educ. Psychol. Meas., 20, 3746.

  • Côté, J., Gravel S. , Méthot A. , Patoine A. , Roch M. , and Staniforth A. , 1998: The operational CMC–MRB Global Environmental Multiscale (GEM) model. Part I: Design considerations and formulation. Mon. Wea. Rev., 126, 13731395.

    • Search Google Scholar
    • Export Citation
  • Foster, R. C., Brown R. A. , and Enloe A. , 1999: Baroclinic modification of midlatitude marine surface wind vectors observed by the NASA scatterometer. J. Geophys. Res., 104 (D24), 31 25531 237.

    • Search Google Scholar
    • Export Citation
  • Haralick, R. M., Shanmugam K. , and Dinstein I. , 1973: Textural features for image classification. IEEE Trans. Syst. Man Cybern., 9, 610621.

    • Search Google Scholar
    • Export Citation
  • Holt, B., 2004: SAR imaging of the ocean surface. Synthetic aperture radar marine user's manual, C. R. Jackson and J. R. Apel, Eds., U.S. Department of Commerce, NOAA/NESDIS, 25–79.

  • Horstmann, J., and Koch W. , 2005: Measurement of ocean surface winds using synthetic aperture radars. IEEE J. Oceanic Eng., 30, 508515.

    • Search Google Scholar
    • Export Citation
  • Ivanov, A. Yu., and Ginzburg A. I. , 2002: Oceanic eddies in synthetic aperture radar images. Proc. Pan Ocean Remote Sensing Conf., Goa, India, Indian Academy of Sciences, 281295.

  • Jones, C. T., Sikora T. D. , Vachon P. W. , and Wolfe J. , 2012: Toward automated identification of sea surface temperature front signatures in RADARSAT-2 images. J. Atmos. Oceanic Technol., 29, 89102.

    • Search Google Scholar
    • Export Citation
  • Lyzenga, D. R., and Marmorino G. O. , 1998: Measurements of surface currents using sequential synthetic aperture radar images of slick patterns near the edge of the Gulf Stream. J. Geophys. Res., 103 (C9), 18 76918 777.

    • Search Google Scholar
    • Export Citation
  • Park, K., Cornillon P. , and Codiga D. , 2006: Modification of surface winds near ocean fronts: Effects of Gulf Stream rings on scatterometer (QuikSCAT, NSCAT) wind observations. J. Geophys. Res.,111, C03021, doi:10.1029/2005JC003016.

  • Risien, C. M., and Chelton D. C. , 2008: A global climatology of surface wind and wind stress fields from eight years of QuikSCAT scatterometer data. J. Phys. Oceanogr., 38, 23792413.

    • Search Google Scholar
    • Export Citation
  • Secker, J., and Vachon P. W. , 2007: Exploitation of multi-temporal SAR and EO satellite imagery for geospatial intelligence. Proc. 10th Int. Conf. on Information Fusion, Quebec City, QC, Canada, IEEE, 17451752.

  • Shokr, M. E., 1991: Evaluation of second-order texture parameters for sea ice classification from radar images. J. Geophys. Res., 96 (C6), 10 62510 640.

    • Search Google Scholar
    • Export Citation
  • Sikora, T. D., and Ufermann S. , 2004: Marine atmosphere boundary layer cellular convection and longitudinal roll vortices. Synthetic aperture radar marine users manual, C. R. Jackson and J. R. Apel, Eds., U.S. Department of Commerce, NOAA/NESDIS, 321–330.

  • Sikora, T. D., Young G. S. , Beal R. C. , and Edson J. B. , 1995: Use of spaceborne synthetic aperture radar imagery of the sea surface in detecting the presence and structure of the convective marine atmospheric boundary layer. Mon. Wea. Rev., 123, 36233632.

    • Search Google Scholar
    • Export Citation
  • Small, R. J., and Coauthors, 2008: Air–sea interaction over ocean fronts and eddies. Dyn. Atmos. Oceans, 45, 274319.

  • Song, Q., Cornillon P. , and Hara T. , 2006: Surface wind response to oceanic fronts. J. Geophys. Res.,111, C12006, doi:10.1029/2006JC003680.

  • Sweet, W., Fett R. , Kerling J. , and Laviolette P. , 1981: Air–sea interaction effects in the lower troposphere across the north wall of the Gulf Stream. Mon. Wea. Rev., 109, 10421052.

    • Search Google Scholar
    • Export Citation
  • Thompson, R. E., Vachon P. W. , and Borstad G. A. , 1992: Airborne synthetic aperture radar imagery of atmospheric gravity waves. J. Geophys. Res., 97 (C9), 14 24914 257.

    • Search Google Scholar
    • Export Citation
  • Tso, B., and Mather M. , 2009: Classification Methods for Remotely Sensed Data. CRC Press, 356 pp.

  • Vachon, P. W., and Dobson F. W. , 2000: Wind retrieval from RADARSAT images: Selection of a suitable C-band HH polarization wind retrieval model. Can. J. Remote Sens., 28, 306313.

    • Search Google Scholar
    • Export Citation
  • Vachon, P. W., Johannessen J. A. , and Browne D. P. , 1995: ERS-1 SAR images of atmospheric gravity waves. IEEE Trans. Geosci. Remote, 32, 10141025.

    • Search Google Scholar
    • Export Citation
  • Vesecky, J. F., and Stewart R. H. , 1982: The observation of ocean surface phenomena using imagery from the SEASAT synthetic aperture radar: An assessment. J. Geophys. Res.,87 (C5), 33973340.

    • Search Google Scholar
    • Export Citation
  • Wai, M. M.-K., and Stage S. A. , 1989: Dynamical analyses of marine atmospheric boundary layer structure near the Gulf Stream oceanic front. Quart. J. Roy. Meteor. Soc., 115, 2944.

    • Search Google Scholar
    • Export Citation
  • Yoder, J. A., Schollaert S. E. , and O'Reilly J. E. , 2002: Climatological phytoplankton chlorophyll and sea surface temperature patterns in continental shelf and slope waters off the Northeast U.S. coast. Limnol. Oceanogr., 47, 672682.

    • Search Google Scholar
    • Export Citation
  • Young, G., and Winstead N. , 2005: Meteorological phenomena in high-resolution SAR wind images. High Resolution Wind Monitoring with Wide Swath SAR: A User's Guide, R. C. Beal et al., Eds., U.S. Department of Commerce, NOAA/NESDIS, 13–31.

1

Following Jones et al. (2012), we define pure atmospheric processes to consist of horizontal and vertical modes of oscillation in the MABL that impact surface roughness and that in principle occur independently of spatial patterns in SST. When such processes occur, the resulting SAR signatures need not be correlated with coincident SST fronts, although they may be similar in appearance to SST front signatures. For example, we consider the signatures labeled WIN in Fig. 1 to be due to pure atmospheric processes, as they do not coincide with any of the SST fronts evident in Fig. 2.

Save
  • Apel, J. R., 2004: Oceanic internal waves and solitons. Synthetic aperture radar marine user's manual, C. R. Jackson and J. R. Apel, Eds., U.S. Department of Commerce, NOAA/NESDIS, 189–206.

  • Armstrong, E. M., Wagner G. , Vazquez-Cuervo J. , and Chin T. M. , 2012: Comparisons of regional satellite sea surface temperature gradients derived from MODIS and AVHRR sensors. Int. J. Remote Sens., 33, 66396651.

    • Search Google Scholar
    • Export Citation
  • Barber, D. G., and LeDrew E. F. , 1991: SAR sea ice discrimination using texture statistics: A multivariate approach. Photogramm. Eng. Remote Sensing, 57, 385395.

    • Search Google Scholar
    • Export Citation
  • Beal, R. C., Kudryavtsev V. N. , Thompson D. R. , Grodsky S. A. , Tilley D. G. , Dulov V. A. , and Graber H. C. , 1997: The influence of the marine atmospheric boundary layer on ERS 1 synthetic aperture radar imagery of the Gulf Stream. J. Geophys. Res., 102 (C3), 57995814.

    • Search Google Scholar
    • Export Citation
  • Canny, J., 1986: A computational approach to edge detection. IEEE Trans. Pattern Anal. Mach. Intell., 8, 679698.

  • Chelton, D. B., and Xie S. P. , 2010: Coupled ocean–atmosphere interaction at oceanic mesoscales. Oceanography, 23, 5269.

  • Cohen, J., 1960: A coefficient of agreement for nominal scales. Educ. Psychol. Meas., 20, 3746.

  • Côté, J., Gravel S. , Méthot A. , Patoine A. , Roch M. , and Staniforth A. , 1998: The operational CMC–MRB Global Environmental Multiscale (GEM) model. Part I: Design considerations and formulation. Mon. Wea. Rev., 126, 13731395.

    • Search Google Scholar
    • Export Citation
  • Foster, R. C., Brown R. A. , and Enloe A. , 1999: Baroclinic modification of midlatitude marine surface wind vectors observed by the NASA scatterometer. J. Geophys. Res., 104 (D24), 31 25531 237.

    • Search Google Scholar
    • Export Citation
  • Haralick, R. M., Shanmugam K. , and Dinstein I. , 1973: Textural features for image classification. IEEE Trans. Syst. Man Cybern., 9, 610621.

    • Search Google Scholar
    • Export Citation
  • Holt, B., 2004: SAR imaging of the ocean surface. Synthetic aperture radar marine user's manual, C. R. Jackson and J. R. Apel, Eds., U.S. Department of Commerce, NOAA/NESDIS, 25–79.

  • Horstmann, J., and Koch W. , 2005: Measurement of ocean surface winds using synthetic aperture radars. IEEE J. Oceanic Eng., 30, 508515.

    • Search Google Scholar
    • Export Citation
  • Ivanov, A. Yu., and Ginzburg A. I. , 2002: Oceanic eddies in synthetic aperture radar images. Proc. Pan Ocean Remote Sensing Conf., Goa, India, Indian Academy of Sciences, 281295.

  • Jones, C. T., Sikora T. D. , Vachon P. W. , and Wolfe J. , 2012: Toward automated identification of sea surface temperature front signatures in RADARSAT-2 images. J. Atmos. Oceanic Technol., 29, 89102.

    • Search Google Scholar
    • Export Citation
  • Lyzenga, D. R., and Marmorino G. O. , 1998: Measurements of surface currents using sequential synthetic aperture radar images of slick patterns near the edge of the Gulf Stream. J. Geophys. Res., 103 (C9), 18 76918 777.

    • Search Google Scholar
    • Export Citation
  • Park, K., Cornillon P. , and Codiga D. , 2006: Modification of surface winds near ocean fronts: Effects of Gulf Stream rings on scatterometer (QuikSCAT, NSCAT) wind observations. J. Geophys. Res.,111, C03021, doi:10.1029/2005JC003016.

  • Risien, C. M., and Chelton D. C. , 2008: A global climatology of surface wind and wind stress fields from eight years of QuikSCAT scatterometer data. J. Phys. Oceanogr., 38, 23792413.

    • Search Google Scholar
    • Export Citation
  • Secker, J., and Vachon P. W. , 2007: Exploitation of multi-temporal SAR and EO satellite imagery for geospatial intelligence. Proc. 10th Int. Conf. on Information Fusion, Quebec City, QC, Canada, IEEE, 17451752.

  • Shokr, M. E., 1991: Evaluation of second-order texture parameters for sea ice classification from radar images. J. Geophys. Res., 96 (C6), 10 62510 640.

    • Search Google Scholar
    • Export Citation
  • Sikora, T. D., and Ufermann S. , 2004: Marine atmosphere boundary layer cellular convection and longitudinal roll vortices. Synthetic aperture radar marine users manual, C. R. Jackson and J. R. Apel, Eds., U.S. Department of Commerce, NOAA/NESDIS, 321–330.

  • Sikora, T. D., Young G. S. , Beal R. C. , and Edson J. B. , 1995: Use of spaceborne synthetic aperture radar imagery of the sea surface in detecting the presence and structure of the convective marine atmospheric boundary layer. Mon. Wea. Rev., 123, 36233632.

    • Search Google Scholar
    • Export Citation
  • Small, R. J., and Coauthors, 2008: Air–sea interaction over ocean fronts and eddies. Dyn. Atmos. Oceans, 45, 274319.

  • Song, Q., Cornillon P. , and Hara T. , 2006: Surface wind response to oceanic fronts. J. Geophys. Res.,111, C12006, doi:10.1029/2006JC003680.

  • Sweet, W., Fett R. , Kerling J. , and Laviolette P. , 1981: Air–sea interaction effects in the lower troposphere across the north wall of the Gulf Stream. Mon. Wea. Rev., 109, 10421052.

    • Search Google Scholar
    • Export Citation
  • Thompson, R. E., Vachon P. W. , and Borstad G. A. , 1992: Airborne synthetic aperture radar imagery of atmospheric gravity waves. J. Geophys. Res., 97 (C9), 14 24914 257.

    • Search Google Scholar
    • Export Citation
  • Tso, B., and Mather M. , 2009: Classification Methods for Remotely Sensed Data. CRC Press, 356 pp.

  • Vachon, P. W., and Dobson F. W. , 2000: Wind retrieval from RADARSAT images: Selection of a suitable C-band HH polarization wind retrieval model. Can. J. Remote Sens., 28, 306313.

    • Search Google Scholar
    • Export Citation
  • Vachon, P. W., Johannessen J. A. , and Browne D. P. , 1995: ERS-1 SAR images of atmospheric gravity waves. IEEE Trans. Geosci. Remote, 32, 10141025.

    • Search Google Scholar
    • Export Citation
  • Vesecky, J. F., and Stewart R. H. , 1982: The observation of ocean surface phenomena using imagery from the SEASAT synthetic aperture radar: An assessment. J. Geophys. Res.,87 (C5), 33973340.

    • Search Google Scholar
    • Export Citation
  • Wai, M. M.-K., and Stage S. A. , 1989: Dynamical analyses of marine atmospheric boundary layer structure near the Gulf Stream oceanic front. Quart. J. Roy. Meteor. Soc., 115, 2944.

    • Search Google Scholar
    • Export Citation
  • Yoder, J. A., Schollaert S. E. , and O'Reilly J. E. , 2002: Climatological phytoplankton chlorophyll and sea surface temperature patterns in continental shelf and slope waters off the Northeast U.S. coast. Limnol. Oceanogr., 47, 672682.

    • Search Google Scholar
    • Export Citation
  • Young, G., and Winstead N. , 2005: Meteorological phenomena in high-resolution SAR wind images. High Resolution Wind Monitoring with Wide Swath SAR: A User's Guide, R. C. Beal et al., Eds., U.S. Department of Commerce, NOAA/NESDIS, 13–31.

  • Fig. 1.

    Two contiguous RADARSAT-2 SCNA VV frames acquired on 7 Mar 2009 at approximately 2230 UTC. SST front signatures (indicated by SST) appear as well-defined brightness fronts in the SAR image that match fronts in the MODIS image in Fig. 2. Signatures of pure horizontal wind shear can be seen in the lower portion of the SAR image, identifiable as brightness fronts misaligned with fronts in the MODIS image (indicated by WIN). Other labeled features are described in the text. (RADARSAT-2 data and products ©2009MacDonald, Dettwiler and Associates Ltd.—all rights reserved.)

  • Fig. 2.

    Composite MODIS SST acquired between 1300 and 1900 UTC on the same day as the RADARSAT-2 frames in Fig. 1.

  • Fig. 3.

    Geographic location of the center of each of the 495 Canny edges collected from 252 RADARSAT-2 images is shown as dark-gray circles (SST) and light-gray squares (WIN). Two lines indicate the GSNWSR.

  • Fig. 4.

    (a) Monthly frequencies (scale on the right) of SST (dark gray) signatures, and to some extent WIN (light gray) signatures, are evidently seasonally dependent. This is only partially explained by the monthly-mean percentage clear sky indicated by the line graph scaled on the left. (b) Frequencies suggest that while validated signatures of both types are unlikely to occur when wind speed exceeds about 10 m s−1, validated SST front signatures are relatively likely to occur even at wind speeds below 3 m s−1.

  • Fig. 5.

    LCCS constructed for each Canny edge was used for the extraction of contextual measures, such as the mean wind direction with respect to the front (μϕ), as well as various textural measures computed for pixels on the three regions labeled darker, front, and brighter. Note that the scale of the grid shown is coarser than the 300-m resolution that was actually used in the analysis.

  • Fig. 6.

    Feature vectors labeled SST (dark-gray circles) and WIN (light-gray squares) are fairly well separated in the two-dimensional subspace consisting of features x1 and x2. Each feature was normalized by its standard deviation (sd) to account for differences in dynamic range, and log transformed to make points more normal in distribution. Curved and straight lines are mean decision boundaries, each computed from 300 trials as described in the text.

  • Fig. 7.

    Estimated probability distributions of μϕ for signatures labeled WIN (gray curve) and SST (black curve) exhibit some separation. Vertical line at 23° indicates the optimal decision bound: edges to the left are more likely signatures of pure horizontal wind shear, while edges to the right are more likely SST front signatures. Edges near the boundary where the probability densities are approximately equal are easily misclassified, suggesting that, to reduce classification error, it may be preferable to leave Canny edges that fall near the boundary (within the gray band) unclassified.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 185 44 3
PDF Downloads 127 37 8