Nothing Special   »   [go: up one dir, main page]
Next Article in Journal
Adaptive Event-Triggered Consensus Control of Nonlinear Multi-Agent Systems via Output Feedback Methodology: An Application to Energy Efficient Consensus of AUVs
Next Article in Special Issue
Detection of Marine Oil Spill from PlanetScope Images Using CNN and Transformer Models
Previous Article in Journal
A Comprehensive Review of an Underwater Towing Cable Array: A Discussion on the Dynamic Characteristics of the Towing Cable Array During the Outspread Process
Previous Article in Special Issue
A Deep Learning Strategy for the Retrieval of Sea Wave Spectra from Marine Radar Data
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
JMSE
Volume 12
Issue 10
10.3390/jmse12101881
jmse-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

A Novel Method for the Estimation of Sea Surface Wind Speed from SAR Imagery

by
Zahra Jafari
1,2,*,
Pradeep Bobby
2,
Ebrahim Karami
1 and
Rocky Taylor
3
1
Department of Electrical and Computer Engineering, Memorial University of Newfoundland, St. John’s, NL A1B 3X5, Canada
2
C-CORE, St. John’s, NL A1B 3X5, Canada
3
Department of Mechanical Engineering, Memorial University of Newfoundland, St. John’s, NL A1B 3X5, Canada
*
Author to whom correspondence should be addressed.
J. Mar. Sci. Eng. 2024, 12(10), 1881; https://doi.org/10.3390/jmse12101881
Submission received: 16 September 2024 / Revised: 11 October 2024 / Accepted: 16 October 2024 / Published: 20 October 2024
(This article belongs to the Special Issue Remote Sensing Applications in Marine Environmental Monitoring)
Browse Figures
Figure 1
<p>Distribution of wind direction and wind speed.</p> "> Figure 2
<p>NRCS vs. incidence angle for different wind speeds and directions using CMOD5N and CMOD7 functions.</p> "> Figure 3
<p>Scatter plots of real versus calculated wind speed using (<b>a</b>) CMOD5, (<b>b</b>) CMOD-IFR, and (<b>c</b>) CMOD7 models with HH polarization.</p> "> Figure 4
<p>Scatter plots of real versus calculated wind speed using (<b>a</b>) CMOD5, (<b>b</b>) CMOD-IFR, and (<b>c</b>) CMOD7 models after compensation for polarization.</p> "> Figure 5
<p>Distribution of intensities for HH and HV polarizations at high and low wind speeds.</p> "> Figure 6
<p>Block diagram of proposed system.</p> "> Figure 7
<p>Effect of despeckling filter on RCM image.</p> "> Figure 8
<p>Histogram of the introduced feature extracted from calibrated data, with orange representing low wind, green representing mid wind, and purple representing high wind.</p> "> Figure 9
<p>Histogram of the introduced feature extracted from uncalibrated data, with orange representing low wind, green representing mid wind, and purple representing high wind.</p> "> Figure 10
<p>Comparisons of retrieved SSWS using concatenated models with different features from the calibrated RCM dataset.</p> "> Figure 11
<p>Comparisons of retrieved SSWS using concatenated models with different features from the uncalibrated RCM dataset.</p> "> Figure 12
<p>The closest region, where both RCM data and buoy station data are available.</p> "> Figure 13
<p>ERA5 vs. buoy wind speeds for the south of Greenland across all seasons in 2023.</p> "> Figure 14
<p>Testing the proposed model in the south of Greenland using buoy wind speed data.</p> ">
Versions Notes

Abstract

:
Wind is one of the important environmental factors influencing marine target detection as it is the source of sea clutter and also affects target motion and drift. The accurate estimation of wind speed is crucial for developing an efficient machine learning (ML) model for target detection. For example, high wind speeds make it more likely to mistakenly detect clutter as a marine target. This paper presents a novel approach for the estimation of sea surface wind speed (SSWS) and direction utilizing satellite imagery through innovative ML algorithms. Unlike existing methods, our proposed technique does not require wind direction information and normalized radar cross-section (NRCS) values and therefore can be used for a wide range of satellite images when the initial calibrated data are not available. In the proposed method, we extract features from co-polarized (HH) and cross-polarized (HV) satellite images and then fuse advanced regression techniques with SSWS estimation. The comparison between the proposed model and three well-known C-band models (CMODs)—CMOD-IFR2, CMOD5N, and CMOD7—further indicates the superior performance of the proposed model. The proposed model achieved the lowest Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE), with values of 0.97 m/s and 0.62 m/s for calibrated images, and 1.37 and 0.97 for uncalibrated images, respectively, on the RCM dataset.

1. Introduction

Surface winds over the ocean are crucial for various climate processes and directly influence ocean circulation, water mass formation, energy transfer, and the motion of icebergs, posing a threat to offshore structures [1]. Studying wind speed is vital for improving tropical cyclone forecasting and protecting coastal communities and offshore infrastructure [2]. Microwave satellite sensors are essential for wind speed estimation over the sea surface. Among these, the spaceborne synthetic aperture radar (SAR) stands out as a high-precision, all-weather imaging radar unaffected by clouds, rain, or solar altitude [3]. The SAR’s capability to operate in various environments, including oceans, makes it particularly valuable for this purpose. By emitting microwaves towards the sea surface and detecting backscattered signals from capillary waves, the SAR can be used to estimate offshore wind speeds. This method provides more accurate sea wind speed measurements than the other scatterometry techniques, making the SAR a superior tool for marine meteorological research [4].
Among the various algorithms for extracting SSWS from backscattering intensity, geophysical model function (GMF) approaches have been validated using large datasets [5]. Each GMF model has distinct advantages and disadvantages [6]. CMOD4 [7] is the most widely used model for general purposes due to its reliable performance, but it tends to underestimate high wind speeds. CMOD_IFR2 [8] is optimized explicitly for the ERS-1 and ERS-2 satellites, offering improved accuracy for these platforms but is less versatile for other missions. CMOD5 [9] addresses the high wind speed underestimation of CMOD4, providing better accuracy in high wind conditions at the expense of increased complexity. CMOD5N [10] further refines CMOD5, enhancing performance across a broader range of conditions but introducing additional complexity and potential overfitting issues. Inaccurate or unavailable wind direction and NRCS values directly impact GMF accuracy [11].
The azimuth cut-off method is independent of wind direction and image calibration but only provides a one-dimensional wind component that is limited to a low incidence angle [12]. The optimal inversion method uses both SAR data and background model data, but it requires a priori information and a set of assumptions to be practical [11]. Multiple NRCS methods developed by He et al. [13] are independent of a priori information; however, they tend to underestimate wind speed in the near range and face a 180-degree ambiguity in wind direction. A method for retrieving the wind vector using dual-polarization ASAR images, proposed by Song et al. [14], is independent of a priori information and eliminates the 180-degree ambiguity in wind direction, but it requires SAR data with both co-polarization and cross-polarization.
In the absence of both co-pol and cross-pol or uncalibrated data, a neural network (NN) facilitates the direct extraction of SSWS from the SAR image intensity without needing any information about the NRCS. This approach bypasses the necessity for explicit models of the SAR imaging process, making it versatile and applicable to various system configurations, including different polarizations and incidence angles [15]. Recent advances in NN methods for wind speed retrieval include a method based on the NN technique and Bayesian regularization algorithm to retrieve SSWS for C-band SAR sensors [16]. YunXiang Liu developed a multi-hidden layer NN for estimating SSWS using Global Navigation Satellite System (GNSS) reflection measurements [17]. Additionally, a back propagation NN for SSWS retrieval was developed, achieving an improved RMSE [18]. Another approach proposed a deep learning-based, end-to-end modified convolutional neural network (CNN) model, which further improved RMSE [19]. These innovative methods demonstrate significant progress in SSWS estimation accuracy and efficiency. Xiaoxu Liu et al. proposed a new hybrid neural network model—a recurrent deep neural network using a feature attention mechanism—for GNSS-R global SSWS retrieval [20]. Recently, Jinwei Bu et al. introduced the GloWS-Net deep learning model for retrieving sea surface wind speed using spaceborne GNSS-R data from CYGNSS mission observations [21]. However, these approaches require VV polarization. In this paper, we propose a novel technique that extracts the top seven statistical features from uncalibrated (normalized) SAR data and predicts the wind speed using various ML regression algorithms. The proposed approach outperforms existing method without requiring wind direction, incidence angle, and works with either co-pol or cross-pol data.
The rest of the paper is structured as follows: Section 2 reviews the materials, including the dataset and ground truth acquired from ERA5. Section 3 outlines the wind speed estimation based on a GMF approach, while Section 4 presents the proposed method. Section 5 discusses the results obtained from calibrated and uncalibrated images. Finally, the paper is concluded and future work is presented in Section 6.

2. Materials and Methods

2.1. Datasets

2.1.1. RCM

The RADARSAT Constellation Mission (RCM) is a Canadian SAR initiative designed to provide enhanced Earth observation capabilities. It comprises three identical satellites equipped with C-band SAR instruments, offering improved spatial coverage and temporal resolution. This allows for frequent monitoring of the Earth’s surface regardless of weather conditions. The multi-polarization modes of RCM, including HH, VV, HV, VH, and compact polarization, enable comprehensive data collection on surface properties, orientation, and changes. RCM data supports a wide range of spatial resolutions, from 5 m to 100 m [22].
In this research, we used a subset of RCM data (https://www.eodms-sgdot.nrcan-rncan.gc.ca/index-en.html, accessed on 17 January 2024), consisting of 150 images collected from January 2022 to December 2023. The medium resolution 50 m mode, acquired in dual polarization (HH + HV), is available. These images were taken on various days from open water areas, ensuring there were no targets or sea ice present. Table 1 summarizes the dataset and Figure 1 displays the distribution of wind direction and wind speed for the date and geographical locations where our dataset was collected. As seen in Figure 1, the most frequent wind speed is around 8 m/s, and the wind direction peaks around 300 degrees in the dataset we used.

2.1.2. ERA5

The Copernicus Climate Change Service (C3S) provides ECMWF reanalysis data products, including ERA5, the fifth-generation atmospheric reanalysis of the global climate. ERA5 combines model and observational data to create a comprehensive global dataset, replacing its predecessor, ERA-Interim. In our paper, we used ERA5 hourly data accessed on January 2022 to December 2023 as ground truth (https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?tab=form, accessed on 17 January 2024). This dataset includes 2 m temperature, 10 m u-component, v-component of wind with spatial resolution of 0.25° × 0.25°, and a temporal resolution of one hour.

3. CMODs and Assessments

Three GMF models were used in this study: CMOD-IFR2, CMOD5N, and CMOD7 [23].
According to the CMOD algorithm for the C-band (wavelength = 5.8 GHz) SAR, the NRCS ( σ 0 ) from the SAR satellite image is modeled using the following GMF:
σ 0 = B 0   1 + B 1   cos ⁡ ( ϕ ) + B 2   cos ⁡ ( 2 ϕ ) γ
where B 0   , B 1   , and B 2   are functions of the incidence angle, wind speed, and wind direction, ϕ represents the relative angle between the wind direction and the radar look direction, and the exponent γ is 1 for CMOD-IFR2 and 1.6 for CMOD5N. These parameters are typically calibrated for the VV polarization, which is more sensitive to wind speed than HH and HV polarizations, to ensure accurate wind speed retrieval from SAR imagery [4].
The CMOD7 algorithm, which is the latest version, improves upon its predecessors by refining the functional parameters to better accommodate varying environmental conditions. The equation for CMOD7 remains structurally similar but incorporates more advanced parameterization [23].
It is worth noting that in this study, we employed GMF approaches with HH polarization images, and the wind direction data were extracted from the ERA5 dataset. The main weaknesses of CMOD algorithms are that they require the σ V V 0 and wind direction [24]. However, this information is often unavailable or uncertain, leading to potential inaccuracies in wind speed estimation from SAR imagery.
Figure 2 presents four subplots comparing the NRCS as a function of incidence angle (θ) for various wind speeds and two wind directions (45° and 90°) using the CMOD5N and CMOD7 models. The NRCS decreases with increasing incidence angle across all wind speeds, with the CMOD7 model generally predicting higher NRCS values than CMOD5N for similar conditions. A red dashed line indicates the Noise Equivalent Sigma Zero (NESZ) threshold of −22 dB, below which NRCS measurements are considered unreliable. This threshold is critical, as NRCS values below it, particularly at lower wind speeds, may not accurately reflect the actual conditions due to measurement noise.

3.1. Result of CMODs Using HH Polarization

In scatterometer data analysis and wind speed retrieval, VV polarization is typically preferred for its sensitivity to wind speed and direction over the ocean. However, there are situations where HH polarization data are used instead [25]. One primary reason for opting for HH polarization is the unavailability of VV polarization data. In such cases, HH polarization serves as an alternative, ensuring that wind speed estimations can still be obtained. While HH polarization may not be as sensitive as VV polarization to wind-related surface roughness, it provides a valuable fallback option [26]. However, using HH instead of VV in CMODs can underestimate the predicted wind speed. This can be compensated by using a polarization ratio to transform the NRCS of HH to VV, using a factor that depends on the data, frequency band, and incidence angle [27].
The results obtained from different GMF-based algorithms using HH polarization are summarized in Table 2 and Figure 3. The first three columns in Table 2 show the average, maximum, and minimum of the wind speed predicted using various CMOD algorithms and the last two columns present the error metrics MAE and MSE for these algorithms. Figure 3 illustrates scatterplots for CMOD_IFR2, CMOD5N, and CMOD7, highlighting that estimation errors increase with wind speed across all three algorithms. This phenomenon is primarily attributed to the difference between σ H H 0 and σ V V 0 , where a typical 2 dB difference is observed in calm ocean waters and up to 5 dB difference in high wind conditions. Utilizing σ H H 0 instead of σ V V 0 introduces additional errors, which become more significant at higher wind speeds, leading to a consistent underestimation of wind speed by all three algorithms.

3.2. Result of CMODs After Polarization Compensation

Due to using HH polarization instead of VV polarization, the slope of the scatterplot between the calculated and actual wind speeds was much smaller (around 0.6) than unity, which is what we would expect if VV data were used. To address this discrepancy, we proposed a heuristic approach that compensates for the scatterplot slope. For each CMOD, after splitting the data into validation (20%) and test sets (80%), we used the validation data to calculate the slope and intercept of the scatterplots of the estimated wind speed versus the actual wind speed. These calculated parameters were subsequently applied to adjust the wind speed predictions in the test dataset. This method allowed us to account for any systematic biases and improve the overall accuracy of our wind speed estimations by ensuring consistency between the calculated and ground truth values across different data subsets. As shown in Table 3 and Figure 4, the performance of the CMOD algorithms is significantly improved after compensation for polarization.
Comparing the results of these algorithms in Table 3, CMOD7 generally provided the most accurate wind speed estimates due to its refined parameterization. CMOD5N performed well but tended to overestimate wind speeds under certain conditions, whereas CMOD-IFR2, while simpler, showed limitations in accuracy compared to the more recent models.
Figure 4 depicts the scatterplots for the CMOD algorithms. It is evident that the slope of the regression line has been significantly improved and is close to unity for all three algorithms. This indicates that the underestimation problem due to using σ H H 0 instead of σ V V 0 has been addressed.

4. Proposed Method

In this paper, we focus on the extraction of statistical features from RCM HH and HV polarization images. As shown in Figure 5, low and high wind speeds have clearly distinct intensity distributions. These extracted features are combined with the incidence angle and wind direction (upon its availability), forming a feature vector for each image patch. This feature vector is used to estimate SSWS. The system architecture is depicted in Figure 6. As shown, RCM images undergo initial preprocessing, followed by feature extraction. The resultant features are then merged, and regression is performed using three ML techniques: NN, RF [28], and CatBoost [29]. The outputs from these methods are aggregated through calculating their median value. Further details regarding each component are elaborated below.

4.1. Data Preprocessing

For this research, the 16-bit digital number (DN) values, which range from 0 to 65,535 for HH and HV polarizations, were provided for different days from the east coast of Canada. Note that in order to reduce complexity of processing, we reduced the data resolution to 8 bits, i.e., ranging from 0 to 255. We extracted 85,000 pairs of image patches, each being 100 × 100 pixels in size, from the original HH and HV RCM images. The positions of these patches were carefully selected from areas with only open water (away from land, sea ice, ships, and icebergs) and were spaced far apart to ensure uncorrelated data samples.
Filtering images affects the signal’s statistical parameters, which are crucial for estimating wind speed. Therefore, we skipped filtering to preserve these parameters. As shown in Figure 7, despeckling filters significantly alter the signal’s statistical properties. To extract the maximum amount of statistical information from the images, we avoided using any smoothing filters.
The DN images were calibrated by C-CORE to recover the initial NRCS values. We used these calibrated images for CMOD algorithms. Additionally, we used these calibrated images with the proposed algorithm for comparison purposes.

4.2. Feature Extraction

4.2.1. Statistical Features

As can be seen in Figure 5, the distribution of intensities for any polarization is influenced by wind speed. Therefore, we exploit this relationship by extracting statistical features. To extract statistical features, each HH and HV patch was processed individually. After investigating various statistical features, the following ones were selected and extracted from the intensity values of each patch:
Minimum and Maximum Intensities: The minimum and maximum intensities are the smallest and the largest pixel values in the image.
Minimum and maximum of windowed min/max ratio (M1, M2): To extract more subtle spatial information from the intensity variation, we computed the minimum/maximum intensity ratio over 4 × 4 non-overlapped windows, and then found the maximum and minimum of this ratio for all 625 windows.
Mean Intensity (μ): The average intensity is the mean pixel value of all the pixels in the image.
Standard Deviation of Intensities: The standard deviation is a measure of the dispersion of pixel values around the mean.
σ = ∑ x   –   μ 2 N
where Σ is the summation symbol over the target area, x is each pixel value, and N is the number of pixels.
Entropy: The entropy is the randomness or uncertainty of a random variable.
E n t r o p y = − ∑ P x l o g 2 P x
where P(x) is the probability of occurrence of value x.

4.2.2. Analysis of Statistical Features

During the regression analysis to estimate wind speed from HH and HV channels, we evaluated statistical features and assessed the impact of individual features by analyzing the mutual information [30] between each feature and wind speed, and the probability density function of each feature versus wind speed. Table 4 and Table 5 present the mutual information between various statistical features and the target output for both calibrated and normalized images, respectively. For calibrated images (Table 4), the HV channel shows higher mutual information across most features compared to the HH channel, particularly for the Mean (0.78) and Max (0.66) features, indicating a stronger relationship with the target output. In contrast, for normalized images (Table 5), the HV channel demonstrates a significantly higher mutual information for Entropy (0.52) and Std (0.52), suggesting that normalization enhances the relevance of these features.
Furthermore, by comparing the histograms of each attribute across three wind speed ranges—low (0.17 to 6.97 m/s), mid (6.97 to 9.79 m/s), and high (9.79 to 25.63 m/s)—as shown in Figure 8 and Figure 9 for calibrated and uncalibrated images, respectively, we can discern the influence of each feature on regression performance. Note that we selected thresholds of 6.97 and 9.79 to divide the data into three ranges, maintaining an equal number of samples in each range. Otherwise, if one range had significantly fewer data points than another, the histogram would not effectively represent each feature’s contribution. Through these histograms, we aimed to illustrate how each individual feature helps predict wind speeds and how effectively they can separate different wind speed ranges.
If the histograms for the three wind speed categories show significant overlap for a particular feature, it suggests that this feature provides limited useful information for wind speed estimation, e.g., the Std for both HH and HV channel in Figure 8. Conversely, if the histograms for the three categories are distinct for a specific feature, it indicates that this feature imparts valuable information related to the target output and significantly contributes to the overall performance. For example, in Figure 8, the min and max intensities for both HH and HV provide meaningful information; however, as shown in Figure 9, they do not provide useful information for uncalibrated data.

4.3. Regression Models

After selecting the best features, we developed a model to predict their ability to estimate wind speed. For this purpose, we tested multiple machine learning regressors, including Random Forest (RF), Support Vector Machines (SVMs), neural networks (NNs), XGBoost, LightGBM, and CatBoost. After evaluating these models on the SAR dataset, we achieved the best accuracy with RF, CatBoost, and the NN, respectively. Therefore, we selected these three models and combined their outputs. The details of each regressor are as follows:
RF: The Random Forest regressor, known for its robustness and simplicity, was configured with 2000 trees (n_estimators = 2000) and a fixed random seed (random_state = 42). RF works by constructing multiple decision trees during training and outputting the average prediction of the individual trees, which helps to improve predictive accuracy and control overfitting.
CatBoost: CatBoost is a powerful gradient boosting algorithm that uses a combination of ordered boosting and oblivious trees, along with an innovative “Bayesian bootstrap” technique to enhance generalization. For our model, we empirically optimized the following parameters, iterations = 500, depth = 6, learning_rate = 0.2, and used log loss as the loss function. The remaining parameters were kept at their default settings, reducing the need for extensive hyperparameter tuning.
NN: For the neural network, we used two fully connected (FC) layers. The hidden layer comprised 96 neurons, followed by a single-neuron output layer for regression. To mitigate overfitting, we applied L2 norm kernel regularization with a factor of 0.01 and included a dropout layer with a rate of 0.2. For the output layer, we used a linear activation function to predict continuous values.

Model Concatenation

To enhance the prediction accuracy, we employed a model concatenation approach, integrating the outputs of the three mentioned models. Each of these models brings unique strengths to the table, contributing to a more robust and reliable prediction system. RF provides strong performance due to its ensemble learning technique, which mitigates overfitting and improves generalization. CatBoost offers advanced gradient boosting capabilities with optimized handling of categorical features and efficient training. The NN adds flexibility and the ability to capture complex patterns through its layered architecture. By concatenating the outputs of these models, we aim to leverage their individual advantages and create a composite model that outperforms each individual component, leading to more accurate and reliable predictions.

5. Validation and Comparison Results

In this section, we assess the performance of the proposed method using different regression algorithms: RF, CatBoost, and NN, for both calibrated and uncalibrated images, as presented in Table 6 and Table 7. Table 6 shows that when the proposed method uses only features extracted from the HH channel, the performance of all algorithms is suboptimal. This is because the effect of wind speed on the HH channel is highly dependent on the incidence angle and wind direction, which were excluded from the features. However, since cross-polarization backscatter from the ocean surface is independent of the wind direction and incidence angle [31], the performance of the proposed method with HV-extracted features is significantly better than with HH features and surpasses all CMOD-based algorithms, especially at a high wind speed. This is a significant outcome, as wind direction data are not always available.
When features extracted from both polarizations are combined, the performance of the proposed method improves. This improvement occurs because the absence of wind direction data is somewhat compensated by the cross-pol value, which is independent of wind direction. Therefore, adding HH features provides additional information to HV. Including the wind direction in the feature set further enhances the algorithm’s performance.
Figure 10 also illustrates the performance of wind speed prediction using concatenated RF, CatBoost, and NN models with various combinations of input features based on NRCS values. Each scatter plot compares the real versus calculated wind speeds, with performance metrics including the RMSE, MAE, R2, and Bias, which are prominently displayed. The model using only HH features (top left) shows the highest RMSE (2.95 m/s) and MAE (2.20 m/s), indicating lower accuracy. This accuracy deteriorates further at high wind speeds. However, HV features (top right) significantly enhance the model’s accuracy, particularly at high wind speeds, as demonstrated by a lower RMSE (1.79 m/s) and a higher R2 (0.90). The incorporation of the incidence angle and HH and HV features (bottom left) further enhances performance, reducing the RMSE to 1.36 m/s and increasing the R2 to 0.94. The most accurate predictions are achieved by including wind direction information (bottom right), resulting in the lowest RMSE and MAE, especially at high wind speeds. At a wind speed of 25 m/s, the estimation error is as low as 0.1 m/s. These results indicate the importance of utilizing a comprehensive statistical and physical feature set to achieve an optimal wind speed prediction accuracy.
Calibrated SAR images are not always available, often due to insufficient memory storage. Therefore, we evaluated the performance of the proposed algorithm on uncalibrated HH and HV images with DN values, and the results are summarized in Table 7. As expected, the accuracy of the wind speed estimation for uncalibrated data is lower than for calibrated data. However, it still outperforms the results obtained from CMOD algorithms, which require VV backscattering data. Similar to Table 6, the best performance is achieved when the features extracted from both polarizations are combined. This demonstrates the robustness of the proposed method, even when using uncalibrated images.
Figure 11 illustrates the performance of concatenated models in predicting SSWS using various feature combinations from the uncalibrated RCM dataset. The top-left and top-right plots show that using HV polarization improves accuracy compared to HH polarization alone. Adding both polarizations in our methodology, along with the incidence angle, further lowers the RMSE to 1.81 m/s and increases the R2 to 0.90. The inclusion of wind direction further enhances model performance, achieving an RMSE of 1.37 m/s, a MAE of 0.97 m/s, and an R2 of 0.94. This demonstrates that our method works effectively with uncalibrated data, even without wind direction and VV polarization, outperforming GMF models. However, as expected, the best estimation accuracy is achieved when the proposed model uses the wind direction as an additional feature, with this improvement being more pronounced at high wind speeds.

Testing the Model with Buoy Data

As our study is limited to the specific region (East Coast of Canada) and there are no alternative sources of wind speed data for this area aside from ERA5, we searched for nearby buoy stations and identified the closest one to be south of Greenland. We tested our model in this region, where both RCM data and buoy station data are available, as shown in Figure 12.
Figure 13 compares the wind speed data from ERA5 with buoy measurements for the region across all seasons of 2023. Figure 14 presents the results of testing our model using buoy data as the ground truth, along with the calculated RMSE and MAE of wind speeds derived from our model. These results indicate a slight increase in the model’s error.
We expected an increase in the model error when the ground truth for the testing data changed from ERA5 to buoy data, while the training data were based on ERA5. Nevertheless, we believe that using buoy data as the ground truth would significantly reduce the error.
Furthermore, as our research involves using ML to measure SSWS from SAR images (RCM dataset), our methodology can capture underlying patterns and relationships in the data, even though ERA5 is less accurate and more challenging to extract reliable distributions from than buoy data. Despite this, we compared our model, which was trained on ERA5 data, with wind speeds measured by buoys. The results showed only a slight decline in performance, as illustrated in Figure 14. This difference is likely due to the fact that the training and testing datasets were not the same; buoy data were used for testing, while ERA5 was used for training.

6. Conclusions

In this research, we propose a novel methodology for estimating wind speed over open water, which is versatile enough to be generalized across various polarization types and data formats, including both normalized and calibrated datasets. Our approach involves extracting features from co-polarization (HH) and cross-polarization (HV) backscattering data and then estimating wind speed using various machine learning algorithms, including Random Forest, CatBoost, and Neural Networks, as well as combinations of these models. We compared the proposed method with CMOD algorithms CMOD5, CMOD-IFR2, and CMOD7, compensating for the lack of VV data. Our results show that the proposed method outperforms these algorithms, even without requiring wind direction and calibrated values. When wind direction information is available, the proposed method achieved an average RMSE and MAE as low as 0.97 m/s and 0.62 m/s, respectively, for calibrated data, and 1.37 m/s and 0.98 m/s for uncalibrated data. At the highest wind speed, the MAE was reduced to less than 0.1 m/s.
The main limitation of our research is the lack of scatterometer and buoy data for the dates and region investigated. In future work, we plan to extend our research to regions where such data are available. This will not only allow us to compare our results but also enable us to leverage the combination of RCM and scatterometer data to improve the accuracy of wind speed estimation.

Author Contributions

Conceptualization, Z.J. and E.K.; Methodology, Z.J.; Software, Z.J.; Validation, Z.J. and E.K.; Investigation, P.B.; Resources, P.B. and R.T.; Data curation, P.B.; Writing—original draft, Z.J.; Writing—review and editing, P.B., E.K. and R.T.; Visualization, E.K.; Supervision, P.B.; Project administration, P.B. and R.T.; Funding acquisition, R.T. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by Equinor ASA.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The data used to produce the figures and analyses in this article include the following: ERA5, a comprehensive global atmospheric reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), which can be accessed at https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?tab=form, accessed on 17 January 2024 and the RCM dataset, available upon request from the corresponding regional climate data provider (https://www.eodms-sgdot.nrcan-rncan.gc.ca/index-en.html, accessed on 17 January 2024). For the analysis of satellite images, ArcGIS Pro software was employed, while SNAP software was used to calibrate SAR imagery. The proposed method was implemented by TensorFlow ver. 2.0 in Python environment.

Acknowledgments

This research was funded by Equinor, whose support is gratefully acknowledged. We would like to thank Maria Yulmetova, GIS Specialist at C-CORE, for providing us with the calibrated data from RCM, which was crucial for our analysis. We also appreciate the technical support provided by Mark Howell and Ian Turnbull, whose expertise significantly contributed to the success of this project.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Carter, L.; Bostock-Lyman, H.; Bowen, M. Water masses, circulation and change in the modern Southern Ocean. In Antarctic Climate Evolution; Elsevier: Amsterdam, The Netherlands, 2022; pp. 165–197. [Google Scholar]
  2. Zhang, B.; Perrie, W. Recent progress on high wind-speed retrieval from multi-polarization SAR imagery: A review. Int. J. Remote Sens. 2014, 35, 4031–4045. [Google Scholar] [CrossRef]
  3. Asiyabi, R.M.; Ghorbanian, A.; Tameh, S.N.; Amani, M.; Jin, S.; Mohammadzadeh, A. Synthetic aperture radar (SAR) for ocean: A review. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sensing. 2023, 16, 9106–9138. [Google Scholar] [CrossRef]
  4. Kim, H.G.; Hwang, H.J.; Lee, S.H.; Lee, H.W. Evaluation of SAR wind retrieval algorithms in offshore areas of the Korean Peninsula. Renew. Energy 2014, 65, 161–168. [Google Scholar] [CrossRef]
  5. Gao, Y.; Sun, J.; Zhang, J.; Guan, C. Extreme wind speeds retrieval using Sentinel-1 IW mode SAR data. Remote Sens. 2021, 13, 1867. [Google Scholar] [CrossRef]
  6. Lin, H.; Xu, Q.; Zheng, Q. An overview on SAR measurements of sea surface wind. Prog. Nat. Sci. USA 2008, 18, 913–919. [Google Scholar] [CrossRef]
  7. Stoffelen, A.; Anderson, D. Scatterometer data interpretation: Estimation and validation of the transfer function CMOD4. J. Geophys. Res. Oceans 1997, 102, 5767–5780. [Google Scholar] [CrossRef]
  8. Vachon, P.W.; Dobson, F.W. Wind retrieval from RADARSAT SAR images: Selection of a suitable C-band HH polarization wind retrieval model. Can. J. Remote Sens. 2000, 26, 306–313. [Google Scholar] [CrossRef]
  9. Hersbach, H. CMOD5: An Improved Geophysical Model Function for ERS C-Band Scatterometry; ECMWF Technical Memorandum; ECMWF: Reading, UK, 2003; 50p. [Google Scholar]
  10. Hersbach, H. CMOD5.NAC-Band Geophysical Model Function for Equivalent Neutral Wind, Report of the European Centre Medium-range Weather Forecasts; ECMWF: Reading, UK, 2008. [Google Scholar]
  11. Lehner, S.; Schulz-Stellenfleth, J.; Schattler, B.; Breit, H.; Horstmann, J. Wind and wave measurements using complex ERS-2 SAR wave mode data. IEEE Trans. Geosci. Remote Sens. 2000, 38, 2246–2257. [Google Scholar] [CrossRef]
  12. Portabella, M.; Stoffelen, A.; Johannessen, J.A. Toward an optimal inversion method for synthetic aperture radar wind retrieval. J. Geophys. Res. 2002, 107, 1. [Google Scholar] [CrossRef]
  13. He, Y.; Perrie, W.; Zou, Q.; Vachon, P.W. A new wind vector algorithm for C-band SAR. IEEE Trans. Geosci. Remote Sens. 2005, 43, 1453–1458. [Google Scholar]
  14. Song, G.; Hou, Y.; Qi, P. Wind vector retrieval using dual polarization imagery of ASAR. Prog. Nat. Sci. USA 2006, 16, 1183–1187. [Google Scholar]
  15. Li, X.; Yang, D.; Yang, J.; Zheng, G.; Han, G.; Nan, Y.; Li, W. Analysis of coastal wind speed retrieval from CYGNSS mission using artificial neural network. Remote Sens. Environ. 2021, 260, 112454. [Google Scholar] [CrossRef]
  16. Yu, P.; Xu, W.; Zhong, X.; Johannessen, J.A.; Yan, X.H.; Geng, X.; He, Y.; Lu, W. A neural network method for retrieving sea surface wind speed for C-Band SAR. Remote Sens. 2022, 14, 2269. [Google Scholar] [CrossRef]
  17. Guo, W.; Du, H.; Guo, C.; Southwell, B.J.; Cheong, J.W.; Dempster, A.G. Information fusion for GNSS-R wind speed retrieval using statistically modified convolutional neural network. Remote Sens. Environ. 2022, 272, 112934. [Google Scholar] [CrossRef]
  18. Li, X.M.; Qin, T.; Wu, K. Retrieval of sea surface wind speed from spaceborne SAR over the Arctic marginal ice zone with a neural network. Remote Sens. 2020, 12, 3291. [Google Scholar] [CrossRef]
  19. Liu, X.; Bai, W.; Tan, G.; Huang, F.; Xia, J.; Yin, C.; Sun, Y.; Du, Q.; Meng, X.; Liu, C.; et al. GNSS-R Global Sea Surface Wind Speed Retrieval Based on Deep Learning. IEEE Trans. Geosci. Remote Sens. 2023, 61, 4207215. [Google Scholar] [CrossRef]
  20. Liu, X.; Bai, W.; Xia, J.; Huang, F.; Yin, C.; Sun, Y.; Du, Q.; Meng, X.; Liu, C.; Hu, P.; et al. FA-RDN: A hybrid neural network on GNSS-R sea surface wind speed retrieval. Remote Sens. 2021, 13, 4820. [Google Scholar] [CrossRef]
  21. Bu, J.; Yu, K.; Zuo, X.; Ni, J.; Li, Y.; Huang, W. Glows-net: A deep learning framework for retrieving global sea surface wind speed using spaceborne gnss-r data. Remote Sens. 2023, 15, 590. [Google Scholar] [CrossRef]
  22. Dabboor, M.; Olthof, I.; Mahdianpari, M.; Mohammadimanesh, F.; Shokr, M.; Brisco, B.; Homayouni, S. The RADARSAT constellation mission core applications: First results. Remote Sens. 2022, 14, 301. [Google Scholar] [CrossRef]
  23. Stoffelen, A.; Verspeek, J.A.; Vogelzang, J.; Verhoef, A. The CMOD7 geophysical model function for ASCAT and ERS wind retrievals. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2017, 10, 2123–2134. [Google Scholar] [CrossRef]
  24. Wei, S.; Yang, S.; Xu, D. On accuracy of SAR wind speed retrieval in coastal area. Appl. Ocean. Res. 2020, 95, 102012. [Google Scholar] [CrossRef]
  25. Geldsetzer, T.; Khurshid, S.K.; Warner, K.; Botelho, F.; Flett, D. Wind speed retrieval from simulated RADARSAT constellation mission compact polarimetry SAR data for marine application. Remote Sens. 2019, 11, 1682. [Google Scholar] [CrossRef]
  26. Rivas, M.B.; Stoffelen, A.; Zadelhoff, G.J.V. The Benefit of HH and VV Polarizations in Retrieving Extreme 491 Wind Speeds for an ASCAT-Type Scatterometer. IEEE Trans. Geosci. Remote Sens. 2014, 461, 52. [Google Scholar]
  27. Shao, W.; Li, X.M.; Lehner, S.; Guan, C. Development of polarization ratio model for sea surface wind field retrieval from TerraSAR-X HH polarization data. Int. J. Remote Sens. 2014, 35, 4046–4063. [Google Scholar] [CrossRef]
  28. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  29. Prokhorenkova, L.; Gusev, G.; Vorobev, A.; Dorogush, A.V.; Gulin, A. CatBoost: Unbiased boosting with categorical features. Adv. Neural Inf. Process. Syst. 2018, 31, 6639–6649. [Google Scholar]
  30. Jafari, Z.; Karami, E. Breast cancer detection in mammography images: A CNN-based approach with feature selection. Information 2023, 14, 410. [Google Scholar] [CrossRef]
  31. Vachon, P.W.; Wolfe, J. C-band cross-polarization wind speed retrieval. IEEE Geosci. Remote Sens. Lett. 2010, 8, 456–459. [Google Scholar] [CrossRef]
Jmse 12 01881 g001
Figure 1. Distribution of wind direction and wind speed.
Figure 1. Distribution of wind direction and wind speed.
Jmse 12 01881 g001
Jmse 12 01881 g002
Figure 2. NRCS vs. incidence angle for different wind speeds and directions using CMOD5N and CMOD7 functions.
Figure 2. NRCS vs. incidence angle for different wind speeds and directions using CMOD5N and CMOD7 functions.
Jmse 12 01881 g002
Jmse 12 01881 g003
Figure 3. Scatter plots of real versus calculated wind speed using (a) CMOD5, (b) CMOD-IFR, and (c) CMOD7 models with HH polarization.
Figure 3. Scatter plots of real versus calculated wind speed using (a) CMOD5, (b) CMOD-IFR, and (c) CMOD7 models with HH polarization.
Jmse 12 01881 g003
Jmse 12 01881 g004
Figure 4. Scatter plots of real versus calculated wind speed using (a) CMOD5, (b) CMOD-IFR, and (c) CMOD7 models after compensation for polarization.
Figure 4. Scatter plots of real versus calculated wind speed using (a) CMOD5, (b) CMOD-IFR, and (c) CMOD7 models after compensation for polarization.
Jmse 12 01881 g004
Jmse 12 01881 g005
Figure 5. Distribution of intensities for HH and HV polarizations at high and low wind speeds.
Figure 5. Distribution of intensities for HH and HV polarizations at high and low wind speeds.
Jmse 12 01881 g005
Jmse 12 01881 g006
Figure 6. Block diagram of proposed system.
Figure 6. Block diagram of proposed system.
Jmse 12 01881 g006
Jmse 12 01881 g007
Figure 7. Effect of despeckling filter on RCM image.
Figure 7. Effect of despeckling filter on RCM image.
Jmse 12 01881 g007
Jmse 12 01881 g008
Figure 8. Histogram of the introduced feature extracted from calibrated data, with orange representing low wind, green representing mid wind, and purple representing high wind.
Figure 8. Histogram of the introduced feature extracted from calibrated data, with orange representing low wind, green representing mid wind, and purple representing high wind.
Jmse 12 01881 g008
Jmse 12 01881 g009
Figure 9. Histogram of the introduced feature extracted from uncalibrated data, with orange representing low wind, green representing mid wind, and purple representing high wind.
Figure 9. Histogram of the introduced feature extracted from uncalibrated data, with orange representing low wind, green representing mid wind, and purple representing high wind.
Jmse 12 01881 g009
Jmse 12 01881 g010
Figure 10. Comparisons of retrieved SSWS using concatenated models with different features from the calibrated RCM dataset.
Figure 10. Comparisons of retrieved SSWS using concatenated models with different features from the calibrated RCM dataset.
Jmse 12 01881 g010
Jmse 12 01881 g011
Figure 11. Comparisons of retrieved SSWS using concatenated models with different features from the uncalibrated RCM dataset.
Figure 11. Comparisons of retrieved SSWS using concatenated models with different features from the uncalibrated RCM dataset.
Jmse 12 01881 g011
Jmse 12 01881 g012
Figure 12. The closest region, where both RCM data and buoy station data are available.
Figure 12. The closest region, where both RCM data and buoy station data are available.
Jmse 12 01881 g012
Jmse 12 01881 g013
Figure 13. ERA5 vs. buoy wind speeds for the south of Greenland across all seasons in 2023.
Figure 13. ERA5 vs. buoy wind speeds for the south of Greenland across all seasons in 2023.
Jmse 12 01881 g013
Jmse 12 01881 g014
Figure 14. Testing the proposed model in the south of Greenland using buoy wind speed data.
Figure 14. Testing the proposed model in the south of Greenland using buoy wind speed data.
Jmse 12 01881 g014
Table 1. Specification of RCM mode used for this study.
Table 1. Specification of RCM mode used for this study.
ModeRes mLooks rng × azSwath Width kmNominal NESZ (dB)Polarization Mode
Medium Resolution50 m504 × 1350 *−22HH, HV
* For medium resolution, a larger swath width up to 600 km is available but with degraded performance.
Table 2. Comparison of SAR wind retrieval algorithms using HH polarization.
Table 2. Comparison of SAR wind retrieval algorithms using HH polarization.
Wind SpeedMean (m/s)Max (m/s)Min (m/s)MAE (m/s)RMSE (m/s)
ERA5 wind8.7725.620.17--
CMOD_IFR26.4822.220.662.793.43
CMOD_5N6.6322.300.582.643.22
CMOD76.2325.080.223.203.78
Table 3. Comparison of SAR wind retrieval algorithms after compensation for polarization.
Table 3. Comparison of SAR wind retrieval algorithms after compensation for polarization.
Wind SpeedMean (m/s)Max (m/s)Min (m/s)MAE (m/s)RMSE (m/s)
ERA5 wind8.7725.620.17--
CMOD_IFR28.7631.99−0.102.352.95
CMOD_5N8.7529.52−0.152.262.80
CMOD78.7728.522.331.732.21
Table 4. Mutual information between each statistical feature and target output from calibrated images.
Table 4. Mutual information between each statistical feature and target output from calibrated images.
ChannelM1M2EntropyMinMaxMeanStd
HH0.200.280.040.340.340.400.15
HV0.140.320.080.580.660.780.10
Table 5. Mutual information between each statistical features and target output from normalized images.
Table 5. Mutual information between each statistical features and target output from normalized images.
ChannelM1M2EntropyMinMaxMeanStd
HH0.190.010.290.000.200.150.26
HV0.290.030.520.000.020.290.52
Table 6. Performance (RMSE/MAE m/s) of individual and concatenated models on statistical features from calibrated images.
Table 6. Performance (RMSE/MAE m/s) of individual and concatenated models on statistical features from calibrated images.
ModelsRFCatBoostNNConcat. Models
Statistical features_HH3.00/2.242.94/2.192.96/2.222.95/2.20
Statistical features_HV1.83/1.411.79/1.401.82/1.421.79/1.39
Statistical features_HH + HV + Inc1.38/1.061.40/1.091.44/1.121.36/1.04
Statistical features_HH + HV + Inc + Dir0.99/0.701.06/0.791.19/0.910.99/0.69
Table 7. Performance (RMSE/MAE m/s) of individual and concatenated models on statistical features from uncalibrated images.
Table 7. Performance (RMSE/MAE m/s) of individual and concatenated models on statistical features from uncalibrated images.
ModelsRFCatBoostNNConcat. Models
Statistical features_HH3.29/2.503.17/2.433.30/2.493.14/2.40
Statistical features_HV2.38/1.902.37/1.812.49/1.952.35/1.79
Statistical features_HH + HV+ Inc1.81/1.401.78/1.391.75/1.351.70/1.30
Statistical features_HH + HV + Inc + Dir1.46/1.101.40/1.011.65/1.281.37/0.97
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Jafari, Z.; Bobby, P.; Karami, E.; Taylor, R. A Novel Method for the Estimation of Sea Surface Wind Speed from SAR Imagery. J. Mar. Sci. Eng. 2024, 12, 1881. https://doi.org/10.3390/jmse12101881

AMA Style

Jafari Z, Bobby P, Karami E, Taylor R. A Novel Method for the Estimation of Sea Surface Wind Speed from SAR Imagery. Journal of Marine Science and Engineering. 2024; 12(10):1881. https://doi.org/10.3390/jmse12101881

Chicago/Turabian Style

Jafari, Zahra, Pradeep Bobby, Ebrahim Karami, and Rocky Taylor. 2024. "A Novel Method for the Estimation of Sea Surface Wind Speed from SAR Imagery" Journal of Marine Science and Engineering 12, no. 10: 1881. https://doi.org/10.3390/jmse12101881

APA Style

Jafari, Z., Bobby, P., Karami, E., & Taylor, R. (2024). A Novel Method for the Estimation of Sea Surface Wind Speed from SAR Imagery. Journal of Marine Science and Engineering, 12(10), 1881. https://doi.org/10.3390/jmse12101881

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop