Gray Level Co-occurrence Matrix textural analysis for temporal mapping of sea ice in Sentinel-1A SAR imagesAbstract
Sea ice is a critical component of the cryosphere and plays a role in the heat and moisture exchange processes between the ocean and atmosphere, thus regulating the global climate. With climate change, detailed monitoring of changes occurring in sea ice is necessary. Therefore, an analysis was conducted to evaluate the potential of using the Gray Level Co-occurrence Matrix (GLCM) texture analysis combined with the backscattering coefficient (σ°) of HH polarization in Sentinel-1A Synthetic Aperture Radar (SAR) images, interferometric imaging mode, for mapping sea ice in time series. Data processing was performed using cloud computing on the Google Earth Engine platform with routines written in JavaScript. To train the Random Forest (RF) classifier, samples of regions with open water and sea ice were obtained through visual interpretation of false-color SAR images from Sentinel-1B in the extra-wide swath imaging mode. The analysis demonstrated that training samples used in the RF classifier from a specific date can be applied to images from other dates within the freezing period, achieving accuracies ≥ 90% when using 64-bit grayscale quantization in GLCM combined with σ° data. However, when using only σ° data in the RF classifier, accuracies ≥ 93% were observed.
Key words
Synthetic Aperture Radar; Random Forest; Google Earth Engine; Antarctic Peninsula
INTRODUCTION
In the marine region north and west of the Antarctic Peninsula (AP), we found a highly productive ecosystem that supports marine populations whose abundance and geographical distribution are affected by changes in sea ice conditions due to atmospheric warming (Clarke et al. 2006, Montes-Hugo et al. 2009, Moffat & Meredith 2018). One such example is krill (Euphausia sp. and Thysanoessa sp.), which plays an important role in the food chains of whales, seals, penguins, squids, and fish and is one of the most studied species in high-latitude oceans (Nicol & Brierley 2010). Loeb & Santora (2015) found that in the northern AP, during the austral summer between 1992 and 2009, there were seasonal variations in the distribution and occurrence of portions of the populations of four krill species (Euphausia frigida, Euphausia superba, Euphausia triacantha, and Thysanoessa macrura), and that these changes were related to environmental conditions associated with the development of sea ice. In addition, algae associated with sea ice are important food sources for juvenile krill during winter (Kohlbach et al. 2017, Schaafsma et al. 2017). In the Bransfield Strait, the sea ice season is relatively short compared to that in other regions, with pancake ice predominating. Consequently, the algae associated with this type of ice have limited time to become substantial biomass sources (Arrigo 2017). Because this complex ecosystem is linked to the occurrence of sea ice, it is essential to monitor its evolution and understand how climate change may affect its distribution and development throughout the year.
In Antarctica, monitoring the spatial and temporal dynamics of sea ice using remote sensors with optical images is challenging because of low solar luminosity during the austral winter and intense cloud cover during the austral summer. These limitations can be overcome using active and passive microwave remote sensors. When using active microwave remote sensing via Synthetic Aperture Radar (SAR) images for sea ice mapping, the input data for training the classifiers are limited to the different frequencies and polarizations provided by the sensors. Therefore, other techniques have been studied to broaden the range of input data, including Gray Level Co-occurrence Matrix (GLCM) textural analysis (Haralick et al. 1973). The parameter to be configured in the GLCM is the number of gray levels in the generated image specified by quantization (L), which can be 4, 8, 16, 32, or 64. For this parameter, Soh & Tsatsoulis (1999) observed that there would be no noticeable trend in improving the analysis of sea ice texture using different L in the ERS-1 SAR images; however, Park et al. (2020) used L at 64 for the classification of sea ice types in Sentinel-1 SAR images to make the most of the system’s capacity and the computational cost would be lower.
Although there is much related research on sea ice classification and typification, the study regions are predominantly located in the Arctic (Soh & Tsatsoulis 1999, Partington et al. 2010, Zakhvatkina et al. 2017, Murashkin et al. 2018, Park et al. 2020, Mahmud et al. 2022) and use desktop computing for processing, requiring computers with good memory capacity for data processing and storage. To overcome these limitations, Google has made available a platform called the Google Earth Engine (GEE), which comprises a web portal that provides global time-series satellite imagery and vector data, and cloud-based computing with access to software and algorithms to process these data, allowing users to create and run custom algorithms. This allows global-scale analysis to be performed with considerable ease compared with desktop computing (Kumar & Mutanga 2018). In this environment, it is possible to process satellite image classifications using GLCM textural parameters as input data based on available SAR images (Google Earth Engine 2023); however, it does not allow textures to be generated with different L parameters.
This study aimed to adapt the original GEE script related to the GLCM textural analysis, allowing the generation of images with different L levels (4, 8, 16, 32, and 64). From this information combined with σ° data derived from Sentinel-1A SAR images (HH polarization, Interferometric Mode - IW), we will analyze the quality of the temporal mapping of Antarctic Sea ice in an oceanic region located west of AP.
STUDY AREA
The Antarctic Specially Managed Area 7 (ASMA 7) ocean region covers the southwest of Anvers Island, the Palmer Basin, and its adjacent island groups, totaling approximately 3,548 km2. The area includes considerable scientific, tourist, and logistical activities. Scientific research in this area is particularly important for understanding ecosystem interactions and long-term environmental changes in the region and how they relate to Antarctica (Smith & Stammerjohn 2001). ASMA 7 is located within the AP West with spatial delimitation established by the Commission for the Conservation of Antarctic Marine Living Resources, as shown in Fig. 1. This commission established the boundaries of the oceanic regions surrounding the Antarctic continent (approximately between 45° and 60°S) to monitor the evolution of living marine resources such as populations of fish, mollusks, crustaceans, birds, and all other species of living organisms.
This region has a predominantly maritime climate dominated by the adjacent sector of the Bellingshausen Sea (60–100°W) (Martin & Peel 1978). Its distinct characteristics are due to an uninterrupted mountain range with altitudes between 1,400 and 2,000 m in the eastern AP, which forms a climatic barrier (Schwerdtfeger 1984). This influences the distinct behavior of sea ice formation in other regions of the Antarctic continent.
Jacobs et al. (1979) and Meredith et al. (2008) reported that in the western AP, the hydrography is strongly influenced by the air-ocean energy balance, availability of Circumpolar Deep Water along the shelf slope, and melting of sea ice. The surface layer is occupied by Antarctic surface water, which is a relatively cold freshwater body. This layer undergoes significant changes throughout the year, such as heat loss and sea ice formation, during the austral fall and winter.
The geomorphological characterization of the AP is highlighted by an uninterrupted mountain range of steep elevation with an average altitude of 2,800 m, where 62% of the glaciers (282,330.22 km2 ) are discharged, that is, they drain their masses mainly from the ice-covered plateau, ice field, ice cap, ocean, or ice shelf, to the west and east of the AP (Silva et al. 2019). This rugged relief forms a distinct climate barrier; in the western and central regions, there is a maritime climate dominated by the Bellingshausen Sea, whereas on the east coast, there is a continental climate dominated by the Weddell Sea (Martin & Peel 1978).
MATERIALS AND METHODS
Pre-processing of orbital sensor data
The Sentinel-1 mission consists of the 1A and 1B satellites, which, using SAR, can obtain images in up to four imaging modes: IW, with a swath width of 250 km and spatial resolution of 5 × 20 m; Wave mode with coverage of 20 × 20 km and resolution of 5 × 5 m; Strip Map, with a swath width of 80 km and resolution of 5 × 5 m; and Extra-Wide swath (EW), with a swath width of 400 km and resolution of 20 × 40 m. The products also include three levels of processing, including Level-0 (raw data), Level-1 (Single Look Complex and Ground Range Detected [GRD] data), and Level-2 (Level-1 data with oceanographic information in its processing) (ESA 2018). The GEE provides IW and EW modes in its dataset at the GRD processing level. The main limitation was the exclusive availability of HH polarization at Level-1 in our study region for the Sentinel-1A and Sentinel-1B satellites, highlighting the importance of GLCM textural analysis. Although the Sentinel-1B satellite in the EW imaging mode has had a low spatial resolution and data discontinuity since December 18, 2021, it has data in the HH and HV polarizations at Level-1. Because of these factors, the last data source mentioned was not selected, which would allow the provision of other polarimetric characteristics for sea ice mapping through cloud computing.
For the study region, Sentinel-1 images were pre-processed in four stages (Google Earth Engine 2024): a) applying the orbit file, updating the orbit metadata with a restored orbit file; b) border noise removal, removing low-intensity noise and invalid data at the edges of the scene; c) thermal noise removal, removing additive noise in sub-bands to help reduce discontinuities between sub-bands for scenes in multi-band acquisition modes; and d) radiometric calibration, where the backscatter intensity is calculated using sensor calibration parameters in the GRD metadata (Fig. 2).
As the GLCM textures are derived from the backscatter (dB) of the SAR images, it is necessary to consider only regions (pixels of the atmospheric reanalysis model) that have a wind speed of < 10 km/h because the surface roughness produced in places with open water with a wind speed of > 10.8 km/h, can alter the properties of the backscatter, particularly in regions where the angle of incidence of the SAR image is higher (Scharien & Yackel 2005, Kwok et al. 1992). In addition, copolarized data (HH and VV) are more sensitive to wind direction and speed (Leshkevich & Nghiem 2007). To determine the wind speed, the National Centers for Environmental Prediction Climate Forecast System Version 2 atmospheric reanalysis model was used with a spatial pixel resolution of 22,264 m, and forecasts were initialized four times a day (6 h). The wind speed (km/h) was obtained from the u and v wind component data using equation 1:
where 𝑣𝑒𝑙 is the wind speed in km/h, 𝑢 is the wind component parallel to the x-axis (longitude) in m/s, and 𝑣 is the wind component parallel to the y-axis (latitude) in m/s.
Linear and textural parameters applied
Sea ice mapping using automated algorithms on images obtained from single-polarization SAR sensors is inherently limited. This restriction arises from the intricate interactions between the parameters of the RADAR system, such as wavelength, polarization, angle of incidence, and properties of the surface (Bovith & Andersen 2005). Therefore, the exclusive use of the backscattering coefficient (σ°) to discriminate between different sea ice types can be ambiguous. However, complementary methods, such as textural analysis of targets, can improve the classification accuracy (Holmes et al. 1984, Clausi 2002, Zakhvatkina et al. 2013, Liu et al. 2015).
Texture is the pattern of intensity variation in an image that involves information from neighboring pixels. The GLCM method proposed by Haralick et al. (1973) is one of the most widely used methods for calculating second-order texture measures. The GLCM is a four-dimensional matrix P (i,j,d, a) calculated from two shades of gray of the reference pixel i and its neighbor j, with co-occurrence distance d and direction a. Various texture features can be calculated from this GLCM matrix, such as the Angular Second Moment, Contrast, Correlation, Variance, Inverse Difference Moment, Sum Average, Sum Variance, Sum Entropy, Entropy, Difference variance, Difference entropy, Information Measure of Corr. 1, Information Measure of Corr. 2, Max Corr. Coefficient, Dissimilarity, Inertia, Cluster Shade, and Cluster Prominence. Another important factor influencing the features of textural analysis is the quantization of gray levels (L), offering options such as 4, 8, 16, 32, and 64.
In this study, the σ° values of the HH polarization were employed within the programming framework of GEE. Additionally, the GLCM Variance and Sum Average texture attributes, as outlined by Hillebrand et al. (2021), were incorporated. These texture attributes were computed across various levels of L and at four distinct pixel distances, considering the average of all the orientation angles. This comprehensive approach empowered the Random Forest (RF) supervised classifier to effectively differentiate between sea ice and open water. Previous research has reported improvements in classification accuracy by associating GLCM textures with σ° data (Shokr 1991, Hillebrand et al. 2021, Mahmud et al. 2022), but using training and validation samples obtained on the same day as imaging. The mapped sea ice encompassed various developmental stages, including new, young, and first-year ice, as classified by Comiso (2009). For new ice, only the pancake type was identified because of the lateral ice edges and snow accumulation with relatively high backscatter in the SAR images. In contrast, young ice consists of a thin layer of ice reaching 15–30 cm in thickness. The transition to first-year ice depends on factors such as the temperature, wind, and geographical location. In some stages, these two types of ice (young and first-year ice) were difficult to distinguish when the ice surface was deformed and covered in snow.
Classification using the RF algorithm
The RF supervised classifier comprises a machine learning algorithm and is essentially an ensemble of classification trees or classification and regression trees. This algorithm generates many decision trees, in which the features are randomly selected, which makes it a robust classifier with respect to noise. Fernández-Delgado et al. (2014) reported the superiority of RF among classifiers, followed in descending order by support vector machines (SVM), neural networks, and boosting ensembles. If we consider the target of classification as sea ice, Park et al. (2020) highlighted that the SVM is frequently used and works well when the dataset is small; however, the training sample is prepared manually and may be prone to contamination by biased decisions when it is an extensive dataset.
In the training phase of the supervised RF classifier, we used SAR images from Sentinel-1B and employed false-color images (Fig. 3). By synchronizing the acquisition periods of Sentinel-1A and Sentinel-1B images and delineating Regions of Interest (ROIs) with consistent characteristics for each class, we were able to extract the σ° and textural GLCM parameters from the SAR images. To mitigate potential discrepancies resulting from hourly differences in the satellite acquisition times, the samples were deliberately obtained away from the boundaries of the ROIs. This approach minimizes data incompatibility arising from the movement of sea ice fragments caused by ocean currents and wind.
SAR image from Sentinel-1A, IW imaging mode, HH polarization, acquired on 09/04/2020, with respective sea ice and open water sampling locations for training the Random Forest classifier (a). False-color composite (R-HH, G-HV, B-HV/HH) of the Sentinel-1B SAR image, EW imaging mode, acquired on 09/05/2020 (b).
Considering the proposed methodology (Fig. 4), we found two temporal data sets of Sentinel-1A SAR images that allowed us to analyze the feasibility of using GLCM textural analysis with the σ° values of the HH polarization for mapping sea ice. The first analysis was conducted on an SAR image captured on 09/04/2020, with classification by RF using training samples from 09/15/2019 (period i), and then compared with classification by RF using training samples from the same day as the image (period i+1). The second analysis was performed on a SAR image acquired on 09/15/2019, with classification by RF using training samples from 09/20/2018 (period i), and then compared with classification by RF using training samples from the same day as the image (period i+1).
Flowchart of the analyses conducted in GEE to verify the feasibility of using the σ° of the HH polarization with GLCM textural parameters for mapping sea ice in time series with SAR images from Sentinel-1A, imaging mode IW.
For statistical validation, we used the precision (Fawcett 2006) of RF-supervised classification using training samples on a different date than the SAR image as an indicator, applying equation 2:
where 𝑇𝑃 represents true positives, indicating pixels that exhibit consistent classification, regardless of whether the classifier’s training samples originate from another date or coincide with the date of the SAR image. Conversely, 𝐹𝑃 denotes false positives, which are pixels with discrepant classifications, showing that training samples obtained on a different date from the image contribute to error in the supervised classifier.
Programming on the GEE platform
The program was developed on a GEE platform using JavaScript. The image classified for 09/15/2019 is available for consultation at https://code.earthengine.google.com/973f126a58fd796e41f96ba25b58fb90, and the image classified for 09/04/2020 is available at https://code.earthengine.google.com/83f00750e8b0e7ad6d477c0511246b93.
RESULTS
From Table I, we can see that the slightest difference in area was found between the two classifications utilizing training samples from periods i and i+1, Which occurred on 09/04/2020, with 199.14 km². This result was obtained using only σ° data derived from Sentinel-1A SAR images within the RF classifier. This same behavior was observed on 09/15/2019, with a difference of 115.71 km², and was also found using exclusively σ° data. These findings suggest that incorporating the GLCM textural information at different L levels did not improve the results of using training samples for the RF classifier on another date, which was limited to the freezing period.
Presentation of the results obtained using the Random Forest (RF) classifier performed on Sentinel-1A SAR images, IW imaging mode, and HH polarization, at different quantizations of the gray levels (L) in the GLCM textural analysis, combined or not with σ°.
When analyzing the different class precision obtained by the RF classifier (Table II), we observed that the precision was ≥ 93% when using only training samples made up of σ° data for the periods analyzed. For the combination of σ° with GLCM textural analysis, only the configuration with 64 levels of L showed precision ≥ 90%, whereas for the other levels of L (4, 8, 16, and 32), only one of the classifications presented showed accuracy ≥ 90%. This insignificant improvement can be explained by the difficulty in detecting the ranges of values for the GLCM textural parameters for sea-ice and open-water classes related to the dynamic and variable nature of sea ice (Zakhvatkina et al. 2017).
Precisions found for the RF classifier applied to Sentinel-1A SAR images, IW imaging mode, and HH polarization, at different L in the GLCM textural analysis, combined or not with σ°.
Comparing Figs. 5 and 6 shows that in the difference image between the RF classifications, the use of the combination of σ° with GLCM textural analysis causes errors in the delimitation of the sea ice margins, with confusion in the classification of pixels between open water and sea ice (Fig. 5). This problem was no longer verified when we used exclusively σ° samples to train the RF classifier (Fig. 6), where the classification errors were randomly distributed, and this can be seen in the difference image. These errors were concentrated in places where there are leads with open water or thin ice, in most cases presenting low surface roughness and therefore being able to show low σ° values in both SAR bands, HH and HV (Murashkin et al. 2018), being highlighted when applying the GLCM textural analysis.
Results obtained from the Random Forest (RF) classifier, applied to Sentinel-1A SAR images, IW imaging mode, and HH polarization with GLCM at 64 levels of L. a) is the classification of 09/04/2020 from training samples of the 09/15/2019 image. b) is the classification of the 09/04/2020 image from training samples from the same day. c) is the difference image between the classifications obtained in a) and b). d) is the classification of 09/15/2019 from training samples of 09/20/2018. e) is the image classification of 09/15/2019 from training samples of the same day. f) is the difference image between the classifications shown in d) and e).
Results obtained from the RF classifier, applied to Sentinel-1A SAR images, IW imaging mode, and HH polarization. a) is the classification of 09/04/2020 from training samples of the 09/15/2019 image. b) is the classification of the 09/04/2020 image from training samples from the same day. c) is the difference image between the classifications obtained in a) and b). d) is the classification of 09/15/2019 from training samples of 09/20/2018. e) is the image classification of 09/15/2019 from training samples of the same day. f) is the difference image between the classifications shown in d) and e).
Differences were also found in coastal areas, where narrow ice zones near the coast were shown incorrectly in our results, as the action of wind in coastal regions can affect the roughness of thin ice, altering the return intensity in the HH polarization (Partington et al. 2010). A similar problem was encountered by Zakhvatkina et al. (2017) when applying the SVM algorithm to RADARSAT-2 SAR images; they found, for example, the misclassification of thin ice as open water.
DISCUSSION
Studies related to the classification of sea ice using the C-band in SAR sensors have focused on HH and HV polarization. Therefore, one of the possibilities that allow more information to help the training classifiers is the GLCM textural parameters. Most research on this subject has focused on the Arctic using training and validation data from the same day as the SAR image acquisition. The possibility of using SAR images in semi-automated classifications for daily sea ice mapping was addressed by Park et al. (2020), who applied GLCM textural parameters to Sentinel-1A EW imaging mode images with HH and HV polarizations and found 98% accuracy between sea ice (first-year ice and multi-year ice) and open water in the winter period. This good indication of classification quality was also found by Zakhvatkina et al. (2017), with 91% accuracy in discriminating between open water and sea ice, using textural parameters in the SVM algorithm in RADARSAT-2 SAR images but applied to images acquired in the HH and HV polarizations.
The SVM classifier has also been compared in other studies. The research conducted by Liu et al. (2015) compared the results of sea ice concentration mapping between the Maximum Likelihood (ML) and SVM classifiers. For the Radarsat-2 data, in which backscatter data and GLCM textural features were used as input data, the SVM classifier presented the best accuracy, as ML worsened the results with the input of new textural information. Mahmud et al. (2022) also analyzed Arctic Sea ice using the σ° of the HH polarization and its GLCM texture parameters for six classes of ice and open water. In this study, they observed that due to the overlap of σ° signatures in the C band in the ice classes, particularly in the early stage, an accuracy of < 60% was obtained using the SVM algorithm. However, an accuracy of 87% was obtained when discriminating between open water and sea ice. This result was similar to that found by our study, in which, regardless of the L levels of the GLCM textures generated, precision ≥ 85% was obtained when using training samples from a period other than the date of imaging. Although our results are satisfactory in relation to other machine learning techniques reported in other research, Huang et al. (2024) applied deep learning techniques for improved sea ice classification using σ° data and GLCM textural parameters from Sentinel-1A. In their research, they observed a considerable improvement in accuracy and precision indicators compared to machine learning techniques; however, the processing was conducted using desktop computing, and there was no script available for processing through cloud computing.
Our study did not explore sea ice typification, a challenge noted in previous studies involving orbital sensors such as Spaceborne Imaging Radar-C. Wenbo et al. (2015) found that preliminary sea ice classification can be achieved by image classification based on the GLCM; however, it cannot accurately identify sea ice types in areas of speckle noise caused by an inaccurate speckle classification phenomenon. In addition, Shokr (1991) found that the GLCM parameters vary greatly when calculated for first-season ice types (new, young, and first-year smooth ice), which are the predominant typologies in our study area; this can be seen in our study in the difference between the classifications generated from different sampling dates for training the RF classifier, which can be seen in Fig. 5.
In our processing methodology, we disregarded SAR images acquired in conditions of significant wind speeds that could affect the response of σ° as a function of the roughness caused by the undulation of the open water (Scharien & Yackel 2005, Kwok et al. 1992), resulting in this good performance of the RF classifier. However, the slight improvement in accuracy indicators found at different levels of L was also addressed by Soh & Tsatsoulis (1999) in their first analysis of the application of GLCM to ERS-1 SAR images for mapping sea ice, reporting no noticeable trend, indicating that the degree of dissimilarity between samples varies with the number of L, and there are noticeable differences with increasing levels of L, which are more random than systematic.
CONCLUSIONS
To verify the possibility of applying a GLCM textural analysis to the temporal mapping of sea ice in Sentinel-1A SAR images, two studies were conducted using a RF supervised classifier. These studies compared the classifications generated from the training samples from a different date (period i) to the imaging date (period i+1). This study focused on the freezing period and oceanic region west of the AP.
The SAR images were processed, and the results were classified and validated using cloud computing on the GEE platform with a view toward applicability and reproducibility. Our study presented an advancement by implementing a script that allows the configuration of the quantization of gray levels (L) in generating GLCM textural parameters, a technique not available from GEE developers on their websites.
This study used interferometric mode imaging of HH polarization images, analyzing the textural parameters GLCM Variance and Sum Average, as previously discussed by Hillebrand et al. (2021). Based on the results obtained, we found no improvement in the precision of the results when we added GLCM textures to the RF classifier training (precision ranging from 85% to 92%), compared to using only backscattering coefficient data (σ°) in dB, with a precision of ≥ 93%.
ACKNOWLEDGMENTS
This research was supported by Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq; Process 465680/2014-3 – INCT da Criosfera) and was financed by the Fundação de Amparo à Pesquisa do Estado do Rio Grande do Sul (FAPERGS; project numbers 17/2551-0000518-0 and 21/2551-0002042-3).
REFERENCES
- ARRIGO KR. 2017. Sea ice as a habitat for primary producers. In: THOMAS DN (Ed), Sea Ice, New Jersey: Editora John Wiley & Sons Ltda, p. 352-369.
- BOVITH T & ANDERSEN S. 2005. Sea ice concentration from single-polarized SAR data using second-order grey level statistics and learning vector quantization. Danish Meteorological Institute: Scientific Report 05-04.
- CLARKE A, MURPHY EJ, MEREDITH MP, KING JC, PECK LS, BARNES DK & SMITH RC. 2006. Climate change and the marine ecosystem of the western Antarctic Peninsula. Philosophical Transactions of the Royal Society B: Biological Sciences 362: 149-166.
- CLAUSI DA. 2002. An analysis of co-occurrence texture statistics as a function of grey level quantization. Canadian Journal of Remote Sensing 28: 45-62.
- COMISO JC. 2009. Polar Oceans from Space, New York: Springer, 507 p.
- FAWCETT T. 2006. An introduction to ROC analysis. Pattern Recognition Letters 27: 861-874.
- FERNÁNDEZ-DELGADO M, CERNADAS E, BARRO S & AMORIM D. 2014. Do we need hundreds of classifiers to solve real world classification problems? J Mach Learn Res 15: 3133-3181.
-
GOOGLE EARTH ENGINE. 2023. ee.Image.glcmTexture. Available in: https://developers.google.com/earth-engine/apidocs/ee-image-glcmtexture Accessed on: May 13, 2024.
» https://developers.google.com/earth-engine/apidocs/ee-image-glcmtexture -
GOOGLE EARTH ENGINE. 2024. Sentinel-1 Algorithms. Available in: https://developers.google.com/earth-engine/guides/sentinel1 Accessed on: May 13, 2024.
» https://developers.google.com/earth-engine/guides/sentinel1 - HARALICK RM, SHANMUGAM K & DINSTEIN I. 1973. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics SMC-3: 610-621.
- HILLEBRAND FL, DE CARVALHO BARRETO ID, BREMER UF, ARIGONY-NETO J, MENDES-JÚNIOR CW, SIMÕES JC, DA ROSA CN & DE JESUS JB. 2021. Application of textural analysis to map the sea ice concentration with Sentinel-1A in the western region of the Antarctic Peninsula. Polar Science 29: 100719.
- HOLMES QA, NUESCH DR & SHUCHMAN RA. 1984. Textural analysis and real-time classification of sea-ice types using digital SAR data. IEEE Trans Geosci Remote Sens 2: 113-120.
- HUANG Y, REN Y & LI X. 2024. Deep learning techniques for enhanced sea-ice types classification in the Beaufort Sea via SAR imagery. Remote Sens Environ 308: 114204.
- JACOBS SS, GORDON AL & AMOS AF. 1979. Effect of glacial ice melting on the Antarctic Surface Water. Nature 277: 469-471.
- KOHLBACH D, LANGE BA, SCHAAFSMA FL, DAVID C, VORTKAMP M, GRAEVE M, FRANEKER JA, KRUMPEN T & FLORES H. 2017. Ice algae–produced carbon is critical for overwintering of Antarctic krill Euphausia superba. Front Mar Sci 4: 1-16.
- KUMAR L & MUTANGA O. 2018. Google Earth Engine Applications Since Inception: Usage, Trends, and Potential. Remote Sensing 10: 1509.
- KWOK R, CUNNINGHAM G & HOLT B. 1992. An approach to identification of sea ice types from spaceborne SAR data. Microw Rem Sens Sea Ice 68: 355-360.
- LESHKEVICH GA & NGHIEM SV. 2007. Satellite SAR remote sensing of Great Lakes ice cover, part 2. Ice classification and mapping. J Great Lake Res 33: 736-750.
- LIU H, GUO H & ZHANG L. 2015. SVM-based sea ice classification using textural features and concentration from RADARSAT-2 dual-pol ScanSAR data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 8: 1601-1613.
- LOEB VJ & SANTORA JA. 2015. Climate variability and spatiotemporal dynamics of five Southern Ocean krill species. Progr Oceanogr 134: 93-122.
- MAHMUD MS, NANDAN V, SINGHA S, HOWELL SE, GELDSETZER T, YACKEL J & MONTPETIT B. 2022. C-and L-band SAR signatures of Arctic sea ice during freeze-up. Remote Sens Environ 279: 113129.
- MARTIN PJ & PEEL DA. 1978. The spatial distribution of 10 m temperatures in the Antarctic Peninsula. J Glaciol 20: 311-317.
- MEREDITH MP, BRANDON MA, WALLACE MI, CLARKE A, LENG MJ, RENFREW IA & KING JC. 2008. Variability in the freshwater balance of northern Marguerite Bay, Antarctic Peninsula: results from δ18O. Deep Sea Research Part II: Topical Studies in Oceanography 55: 309-322.
- MOFFAT C & MEREDITH M. 2018. Shelf–ocean exchange and hydrography west of the Antarctic Peninsula: a review. Philos Transact A Math Phys Eng Sci 376: 1-17.
- MONTES-HUGO M, DONEY SC, DUCKLOW HW, FRASER W, MARTINSON D, STAMMERJOHN SE & SCHOFIELD O. 2009. Recent changes in phytoplankton communities associated with rapid regional climate change along the western Antarctic Peninsula. Science 323: 1470-1473.
- MURASHKIN D, SPREEN G, HUNTEMANN M & DIERKING W. 2018. Method for detection of leads from Sentinel-1 SAR images. Ann Glaciol 59: 124-136.
- NICOL S & BRIERLEY AS. 2010. Through a glass less darkly - New approaches for studying the distribution, abundance and biology of Euphausiids. Deep Sea Research Part II: Topical Studies in Oceanography 57: 496-507.
- PARK JW, KOROSOV AA, BABIKER M, WON JS, HANSEN MW & KIM HC. 2020. Classification of sea ice types in Sentinel-1 synthetic aperture radar images. The Cryosphere 14: 2629-2645.
- PARTINGTON KC, FLACH JD, BARBER D, ISLEIFSON D, MEADOWS PJ & VERLAAN P. 2010. Dual-polarization C-band radar observations of sea ice in the Amundsen Gulf. IEEE Transactions on Geoscience and Remote Sensing 48: 2685-2691.
- SCHAAFSMA FL, KOHLBACH D, DAVID C, LANGE BA, GRAEVE M, FLORES H & VAN FRANEKER JA. 2017. Spatio-temporal variability in the winter diet of larval and juvenile Antarctic krill, Euphausia superba, in ice–covered waters. Marine Ecol Progr Series 580: 101-115.
- SCHARIEN RK & YACKEL JJ. 2005. Analysis of surface roughness and morphology of first-year sea ice melt ponds: Implications for microwave scattering. IEEE Trans Geosci Remote Sens 43: 2927-2939.
- SCHWERDTFEGER W. 1984. Weather and Climate of the Antarctic, Amsterdam: Elsevier, 262 p.
- SHOKR ME. 1991. Evaluation of second-order texture parameters for sea ice classification from radar images. J Geophys Res: Oceans 96: 10625-10640.
- SILVA N, WAINER I & TONELLI M. 2019. Caracterização de mudanças climáticas na Antártica a partir da segunda metade do Século XX. Revista Brasileira de Geografia Física 12: 2091-2107.
- SMITH RC & STAMMERJOHN SE. 2001. Variations of surface air temperature and sea-ice extent in the western Antarctic Peninsula region. Ann Glaciol 33: 493-500.
- SOH LK & TSATSOULIS C. 1999. Texture analysis of SAR sea ice imagery using gray level co-occurrence matrices. IEEE Transactions on geoscience and remote sensing 37: 780-795.
- WENBO W, YUSONG W, XUE D, XIAOTONG J, YIDA K & XIANGLI W. 2015. Sea ice classification of SAR image based on wavelet transform and gray level co-occurrence matrix. In: 2015 Fifth International Conference on Instrumentation and Measurement, Computer, Communication and Control (IMCCC), Qinhuangdao, China. IEEE Xplore.
- ZAKHVATKINA NY, ALEXANDROV VY, JOHANNESSEN OM, SANDVEN S & FROLOV IY. 2013. Classification of sea ice types in ENVISAT synthetic aperture radar images. IEEE Trans Geosci Remote Sens 51: 2587-2600.
- ZAKHVATKINA N, KOROSOV A, MUCKENHUBER S, SANDVEN S & BABIKER M. 2017. Operational algorithm for ice–water classification on dual-polarized RADARSAT-2 images. The Cryosphere 11: 33-46.
