**1. Introduction**

Sea ice refers to saltwater ice that is directly frozen from seawater, and also includes continental glaciers (icebergs and Iceland), river ice, and lake ice that enter the ocean. Sea ice greatly inhibits the heat and steam exchange between the ocean and atmosphere and alters the radiation budget and energy balance of the ocean's surface. These changes have an important impact on oceanic hydrological conditions, atmospheric circulation, and ocean climate [1]. In addition, the production and disappearance of sea ice greatly affects human marine activities. For example, sea ice significantly affects the development of marine resources and marine transportation [2]. Moreover, sea ice presents a potential freshwater resource [3]. Therefore, accurate real-time monitoring of sea ice bears an important application value and theoretical significance.

The bottom of the Yellow River Delta contains a small flat slope, shallow water low salinity, as a result of which seawater is easily frozen [4]. The development of conditions conducive to ice formation and persistence in this sea area is unstable; that is, the ice disappears as the temperature rises and reappears as the temperature drops [5]. Sea ice

**Citation:** Qiu, H.; Gong, Z.; Mou, K.; Hu, J.; Ke, Y.; Zhou, D. Automatic and Accurate Extraction of Sea Ice in the Turbid Waters of the Yellow River Estuary Based on Image Spectral and Spatial Information. *Remote Sens.* **2022**, *14*, 927. https://doi.org/ 10.3390/rs14040927

Academic Editor: Simona Niculescu

Received: 29 December 2021 Accepted: 7 February 2022 Published: 14 February 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

is formed at a rapid rate and responds more closely to the local climate. Sea ice disasters frequently occur in the waters of the Yellow River Delta, and sea ice often has a major impact on fishing ports, wharves, shallow beach aquaculture, and offshore infrastructure [6]. For example, the operation area of the Shengli Oilfield, China's second-largest oilfield, is mainly concentrated in the Yellow River Delta, the Bohai Sea and its adjacent waters. Sea ice severely affects and threatens offshore oilfield production operations and various engineering facilities in the winter [7]. According to statistics, severe and relatively serious sea ice disasters occur roughly once every five years in China, and in some sea areas, sea ice disasters occur almost every year [8]. In the winter of 2009–2010, the Bohai Sea and the northern part of the Yellow Sea experienced the worst ice conditions in nearly 30 years. The severe ice conditions had a significant impact on the society and economy of the provinces (cities) along the Yellow Sea and the Bohai Sea. According to statistics, sea ice disasters caused economic losses totaling nearly USD 900 million [9]. Therefore, in the context of global warming, it is entirely possible that continuous low temperatures and severe sea ice disasters will occur in some areas. In particular, with the rapid economic development of the Bohai Rim region, the losses suffered by sea ice disasters are increasing year by year [10]. In the new era, the country should promote the understanding of the ocean, rationally develop and utilize marine resources, protect the rights and interests of the ocean, and insist on the harmonious coexistence of man and the ocean. Faced with the new requirements as China enters a new era and accelerates the building of its maritime power, the existing sea ice disaster prevention and mitigation capabilities can no longer fully meet the actual needs of the economic and social development in icy sea areas, and sea ice monitoring capabilities remain relatively weak [11]. It is the trend of future development to improve the three-dimensional marine disaster observation network that combines coastal observation, offshore platform and satellite remote sensing, so as to improve the ability of marine disaster observation [12]. Therefore, the precise extraction of sea ice provides the basis for strengthening the analysis of and research on sea ice conditions, and is necessary for improving early sea ice warning technology, as well as ice prevention and disaster reduction capabilities.

Traditional sea ice monitoring methods, such as shore-based observations and icebreaker observations, cannot obtain large-scale sea ice information in a timely and accurate manner. Remote sensing monitoring technology has a high timeliness, can obtain repeated observations on a large scale, and is relatively inexpensive, providing long-term data support for dynamic and efficient sea ice monitoring [13–15].

Large-scale monitoring of sea ice in high latitudes can be carried out using remote sensing technologies such as microwave and optical remote sensing. Passive microwave and synthetic aperture radar (SAR) imagery enables all-weather observations and the ability to penetrate through clouds [16,17]. The automatic segmentation algorithm based on statistical distance realizes the classification of C-band fully polarized sea ice data [18]. A sea ice classification method is proposed for X-band, C-band and L-band fully polarized synthetic aperture radar images. By extracting the polarized features of sea ice classification, the feature vector provides input into the neural network classifier to realize the extraction of sea ice [19]. However, it is challenging to obtain due to the high cost and the long revisit period for most of them [20]. Optical remote sensing data, although limited by weather conditions, usually delivers better spatial resolution, lower cost, and shorter revisit times. For example, satellites such as Moderate Resolution Imaging Spectroradiometer (MODIS) [21], Advanced Very High-Resolution Radiometer (AVHRR), Geostationary Ocean Color Imager (GOCI) [22], and Feng Yun 3 (FY-3) [23] has a high time resolution and can be used for continuous monitoring of sea ice in a large-scale area and a long time sequence. The disadvantage is that the spatial resolution is low, and it is difficult to perform refined regional sea ice monitoring. High spatial resolution remote sensing data represented by GF1 [24], Landsat [25], and Sentinel-2 [26] can be used to achieve refined sea ice monitoring through data mining, and the effective combination of multi-source medium and highresolution satellite data compensates for the low time resolution of the data and further improves its sea ice monitoring capabilities.

Three main types of feature parameters are used for sea ice extraction: spectral features, spatial features, and temperature features. The spectrum processing methods commonly used for sea ice extraction include the band ratio, band difference, and various normalized indexes, which highlight the sea ice information [27]. In addition, the sample point selection and machine learning methods have also been used to extract spectral information about sea ice. The second concerns the spatial features. The spatial feature commonly used for sea ice extraction is the textural feature based on the gray-level co-occurrence matrix. Generally, the surface of sea ice is rough and has conspicuous irregular and unstable textural characteristics; while the surface of sea water is smooth and has constant textural characteristics [28]. The third refers to the temperature feature. This parameter is used to retrieve the surface temperature through the thermal infrared band of the remote sensing image. The temperature feature is simple and easy to use, but has fewer data sources and a low spatial resolution; the accuracy of the temperature retrieval algorithm is also not very high, limiting its application [29].

Current sea ice extraction methods can be divided into threshold segmentation methods, machine learning methods, and digital image processing methods. The threshold segmentation method mainly involves setting the threshold value of the sea ice's spectral, textural, temperature, and other parameters and determining the threshold values using an artificial or bimodal histogram, scatter plots, and other methods, such as using the red band and the ratio threshold of the near-infrared band to achieve rapid extraction of sea ice [27]. Su et al. used the red and near-infrared bands of Sentinel-3 images to establish a sea ice information index that highlights the spectral information of sea ice, and employed the Jenks method to determine the segmentation threshold of ice water [30]. Hayashi et al. used reflectance scatter plots of MODIS bands 1 and 2 to derive a formula suitable for extracting the area of thin ice [31]. Ice conditions in the Gulf of Riga in the Baltic Sea were counted by a bimodal histogram method, the statistical results were limited by the spatial resolution of MODIS [32]. The threshold segmentation method is simple and fast, but it is difficult to determine the threshold value using this method, and the threshold value of different images will be slightly different. Machine learning methods such as support vector machines and Classification and Regression Trees (CART) decision trees select certain sample points and then classify the sea ice and seawater areas [33]. The machine learning method is simple and easy to use, but the classification process is unclear, and the sample points need to be manually selected, making it difficult to achieve automation. In addition, the final accuracy depends entirely on the selected sample points. The digital image processing method is to highlight the sea ice information by processing and transforming the image. For example, Li et al. proposed a linear spectral decomposition method based on MODIS images with multiple constrained end members [34]. The pixels are decomposed to extract the range of the sea ice. Liu et al. used a wavelet transform to extract the textural information from a SAR image, converted a China–Brazil Earth Resources Satellite (CBERS-02B) optical image from Red-Green-Blue space to Hue-Saturation-Intensity space, and finally employed Principal Component Analysis (PCA) to fuse the HSI image and the texture images [35]. Digital image processing methods can highlight sea ice information more intuitively, but the process is more complicated and difficult to automate.

The seawater turbidity of the Yellow River Delta is high, and the changes are quite drastic, which intensifies the distinction between sea ice and seawater. Aiming at the problems of the low accuracy and efficiency of the current sea ice extraction methods used in the Yellow River Delta and based on multi-source remote sensing images, in this study, a set of sea ice automatic extraction method systems suitable for the Yellow River Delta was developed and the spatial and textural information about the sea ice was fully excavated. The automatic extraction method will be further extended to other areas such as Liaodong Bay in order to provide important technical support for the rational development of sea ice resources and to improve the early warning capabilities of sea ice disaster prevention and

mitigation systems. This method explores the spectral information and texture information of different ice types. The spectral information, texture information and edge information of sea ice are used to extract sea ice, and the completeness and accuracy of sea ice extraction results are improved through multi-scale segmentation of images. This method can provide a reliable method for extracting sea ice extent from high-resolution optical remote sensing data such as GF1.

#### **2. Materials and Methods**

#### *2.1. Study Area and Data*

The Yellow River Delta is located in the southeastern part of Bohai Bay and the northwestern part of Laizhou Bay (Figure 1). The sea area is located in a typical monsoon climate zone. The winter is controlled by the Eurasian continental high pressure system. The north and northwest winds are dominant, and the weather is dry and cold. When strong cold air invades, it is often accompanied by processes such as strong winds, snowfall, and sharp temperature drops. The coastline is more than 100 km long, and the sea water near the coast mostly contains suspended sediment from the mouth of the Yellow River. Therefore, the sea area has a wide intertidal bandwidth, a small and flat bottom slope, and shallow water depths. In winter, this area is greatly affected by the meteorological conditions and the continent. In addition, affected by the runoff from the Yellow River and other rivers entering the sea, the salt content of the seawater in this sea area is relatively low. The aforementioned special geographical environment and climatic conditions provide sufficient and necessary conditions for the freezing of seawater in this sea area. Sea ice often has a significant impact on fishing ports, wharves, shallow beach aquaculture, and marine infrastructure in this area. With the gradual expansion of the economic scale of the marine aquaculture industry in this area, the impact of sea ice on production activities has become more significant.

**Figure 1.** The study area.

According to the development stage of sea ice, the sea ice in the Yellow River Delta can be divided into new ice (NI), ice rind, nilas (NL), grey ice (GR), grey-white ice (GW), and white ice. New ice is formed by direct freezing of seawater or snow falling when the temperature is low, and the sea surface is not melted. It is mostly needle-like, flake-like, grease-like, or sponge-like. Ice rind is formed by the freezing of new ice or direct freezing of the calm sea surface. The surface of the ice crust is smooth, moist, and shiny. Its thickness is about 5 cm. It can fluctuate with the wind and is easily broken by wind and waves. Nilas ice is a thin, elastic ice crust with a thickness of less than 10 cm. It easily bends and breaks under external forces and can produce a finger-like overlapping phenomenon. Grey ice is

an ice cap layer with a thickness of 10–15 cm. It is developed from nilas. The surface is flat and moist. It is mostly grey. It is less elastic than nilas ice. It is easily broken by swells and overlaps when squeezed. Grey-white ice is an ice layer with a thickness of 15–30 cm, and is developed from grey ice. It has a rough surface, is greyish-white, and mostly forms when ice ridges are squeezed. White ice describes an ice layer with a thickness of greater than 30 cm. It develops from grey-white ice, has a rough surface, and is mostly white [36].

In this study, Sentinel-2 and Landsat8 images from 2017–2019 were selected to verify the applicability of the method. The time phase and image quality information obtained from the data is presented in Table 1.



The GF1 data used in this article were obtained from the China Resources Satellite Application Center (http://www.cresda.com/CN/, accessed on 21 January 2017). The PIE-Basic software was used for geometric correction, atmospheric correction, orthorectification, image clipping, and other pre-processing work. The Sentinel-2 data were obtained from the European Space Agency's (ESA) data sharing website (https://scihub.Coppe-rnicus.eu/, accessed on 12 January 2018). The Sentinel-2 data released is a product of the Top-of-Atmosphere (TOA) reflectance that has been geometrically corrected and radiometrically corrected, so it was only necessary to perform atmospheric correction of this dataset. The SNAP software officially provided by the ESA was used to perform the atmospheric correction on the downloaded data. The Landsat data were obtained from https://landlook. usgs.gov/ (accessed on 23 January 2019) and the ENVI software was used to perform the FLAASH atmospheric correction. In order to facilitate the subsequent statistical analysis, calculations, and other operations and to reduce the data storage space, all of the reflectance data were expanded by 10,000 times and rounded.

#### *2.2. Sea–Land Separation*

In order to avoid interference from land information, the sea and land need to be separated before the sea ice extraction. The Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Water Index (NDWI) are the most commonly used indexes for water and land separation. They both use the normalized ratio of the reflectance between visible and near-infrared light. The difference is that the NDVI uses the green and near-infrared bands; while the NDWI uses the red and near-infrared bands. In the Yellow River Delta, the concentration of suspended sediment in some of the seawater is extremely high, which improves the reflectivity of the seawater in the near-infrared band. Figure 2 shows the spectral curve of the seawater and sea ice in the GF1 image. The reflectivity of the clean seawater is significantly higher in the visible light range than in the near-infrared band, while the reflectivity of the highly turbid seawater is higher in the near-infrared band than in the blue-green band. However, it is still lower than the reflectivity in the red band since the reflectivity in the blue and green bands will gradually become saturated as the concentration of suspended sediment increases, while the reflectivity in the red and near-infrared bands will continue to increase as the concentration of suspended sediment increases [37]. However, the near-infrared reflectivity is always lower than the red-band reflectivity. Sea ice 1 and sea ice 2 are defined as sea ice in clean water and turbid water, respectively, and the reflectivity in the near-infrared band is also lower than that in the

red band. Therefore, the effect of using the NDVI for sea–land separation is better. Pixels with NDVI values of less than 0 are classified as seawater, and pixels with NDVI values of greater than 0 are classified as land. In addition, after separation, it is necessary to filter out discrete areas such as rivers and lakes on the land and ultimately to only retain the ocean area. The results of the sea–land separation are shown in Figure 3.

**Figure 2.** The reflectance of the ice and water in the GF1 image.

**Figure 3.** (**a**) True color GF1 image acquired on 12 January 2018; and (**b**) the results of the sea–land separation.

#### *2.3. Sea Ice Spectral Information Extraction*

Spectral information is the most commonly used feature for extraction. In this section, a sea ice spectral information index suitable for different suspended particulate matter concentrations based on the spectral characteristics of sea ice is developed. Generally, the reflectance of seawater in clear seas is higher than that of seawater. However, the suspended particulate matter (SPM) in the Yellow River Estuary increases the reflectance of the seawater. The reflectance of seawater with a high SPM content is even higher than that of sea ice with a lower SPM content (Figure 4). As can be seen from the box plot in Figure 4, as the wavelength increases, the reflectivity of the seawater fluctuates more widely, and the blue band is relatively less affected by the SPM content. This is due to the fact that as

the SPM content increases, the blue band reaches saturation first, followed by the green and red bands. From the perspective of the type of sea ice, in addition to grey ice and grey-white ice, a considerable part of the new ice and ice rind has the same reflectance as seawater. This is because grey ice and grey-white ice are thicker and have a much higher reflectivity than seawater, whereas new ice and ice rind are thin ice with a thickness of less than 10 cm and have a lower reflectivity. Therefore, a single waveband cannot distinguish all types of sea ice from seawater. The difficulty of sea ice extraction in the Yellow River Estuary is mainly the extraction of the new ice and ice rind from seawater with different SPM contents.

**Figure 4.** Sea ice and water reflectance box plot.

The commonly used spectral index such as the NDVI and NDSI, cannot eliminate the influence of SPM on sea ice extraction. Therefore, in this study, the spectral information concerning the sea ice was extracted by searching for an optimal band combination method. Since the visible light and near-infrared bands are the most important bands for extracting sea ice, these four bands were selected as the best band combination from the data source. The scatter plot can intuitively reflect the separation of the different samples. As shown in Figures 3 and 4, each point in the scatter plot formed by any two bands represents the position of the sea ice or seawater in this two-dimensional space. If the points of sea ice and seawater are scattered together (Figure 5a), then the separation of the sea ice and seawater in the two-dimensional space formed by these two wavebands is poor; and if the sea ice and seawater each gather in one location, it indicates that the sea ice and seawater can be separated well using this waveband combination (Figure 5b). In this way, a straight line separating the sea ice and seawater can be drawn in this two-dimensional space.

**Figure 5.** (**a**) Band 1 and Band 2 exhibit good separation in this scatter plot; (**b**) Band 1 and Band 2 display poor separation in this scatter plot.

In order to test all of the band combinations as much as possible, in this study, combinations of the four red, green, blue, and near-infrared bands were tested first. The results of these combinations are presented in Table 2. There are 28 results in total. Two of these 28 bands were chosen to construct a two-dimensional scatter plot, with a total of 378 combinations. For example, if the NDVI can better distinguish between sea ice and seawater, then the scatter plots constructed with the NIR − R and NIR + R bands will exhibit better separation. Finally, the reflectivity of grey ice and grey-white ice is much higher than that of primary ice and ice skins, and the extraction is relatively simple. In order to make the scatter plot show the separation of new ice, ice rind, and seawater better, first only the sample points of new ice, ice rind, and seawater were selected to construct the scatter plot. The optimal band combination was selected by averaging the Euclidean distance and inter-class variance combined with visual interpretation.


**Table 2.** First band combination list.

The average Euclidean distance is

$$\mathbf{x} = \frac{1}{n} \sum\_{i=1}^{n} \mathbf{x}\_{i\prime} \tag{1}$$

where *x*(*y*) is the average Euclidean distance of the sea ice or seawater on the *x*-axis or *y*-axis, *n* is the number of sea ice or seawater sample points, and *x*<sup>i</sup> is the reflectivity of sea ice or seawater.

$$\mathbf{U} = \sqrt{(\mathbf{x}\_{\text{ice}} - y\_{\text{ice}})^2 + (\mathbf{x}\_{\text{sea}} - y\_{\text{sea}})^2} \,\text{}\tag{2}$$

where U is the average Euclidean distance between the sea ice and seawater in twodimensional space, *x* and *y* are the average Euclidean distances of the sea ice sample points on the *x*-axis and *y*-axis, respectively, and *x* and *y* are the sea ice sample points on the *x*-axis and *y*-axis, respectively.

The variance between classes is

$$
\sigma = \sum\_{i=1}^{n} (\mathbf{x}\_i - \overline{\mathbf{x}}\,)^2,\tag{3}
$$

where σ is the between-class variance of sea ice or seawater, and xi is the number of sample points of sea ice or seawater. The larger the average Euclidean distance and the smaller the variance between classes, the higher the degree of separation is.

According to the principle of a larger average Euclidean distance and a smaller variance between classes, the top 12 band combinations were selected and their scatter plots were examined. As shown in Figure 6, the abscissas and ordinates of the scatter plots are the reflectance of the band after a band combination, for example, B is the reflectance of the green band, and R/B is the ratio of the reflectance of the red band to the green band. It was found that in the scatter plots of B and R/B, sea ice and seawater can basically be separated by a straight line. The mixing of seawater is more serious. Therefore, the band combination with the best reflectivity ratio between the green band and the red band was finally selected.

**Figure 6.** Scatter plots of band combinations for seawater and sea ice.

R/B and B were used as the *x*-axis and *y*-axis, respectively, to draw the scatter plots of the seawater and different ice types (Figure 7). The seawater reflectance samples exhibit good linearity in the scatter plots. The different types of sea ice are located above the straight line. Therefore, the linear equation of seawater was obtained through linear fitting as the dividing line.

**Figure 7.** Scatter plot of the reflectivity of the different types of sea ice and seawater.

The dividing line equation is

$$\mathbf{y} = 269\mathbf{x} + 611,\tag{4}$$

where x is the reflectance ratio of the red band to the green band, and y is the reflectance of the green band.

In order to make the value of the sea ice a positive number to facilitate subsequent statistical analysis and the determination of the threshold, only the slope of the seawater linear fitting line was used, and finally the sea ice spectral information index was constructed:

$$\mathbf{y} = \mathbf{B} - 2\epsilon\mathbf{9} \times (\mathbf{R}/\mathbf{B}),\tag{5}$$

where y is the calculated reflectance value of each pixel in the image (sea ice has a larger value and seawater has a relatively small value); B is the reflectance value of the green band; and R/B is the ratio of the reflectance in the red band to that in the green band.

#### *2.4. Sea Ice Spatial Information Extraction*

As shown in Figure 8, in optical images, such as GF1, Landsat, and Sentinel-2 images, some thin ice such as new ice and ice rind exhibits spectra very similar to that of seawater containing suspended particles. Sea ice in seawater with a higher concentration of suspended particulate matter cannot be extracted using spectral information alone. Therefore, it is necessary to distinguish this part of the sea ice from the seawater based on the spatial information concerning the sea ice. Compared with the smooth spatial characteristics of the seawater surface, the surface of sea ice is generally rough, with conspicuous irregular and unstable textural characteristics. In this section, three different spatial information extraction schemes are designed to explore the applicability of the textural features and edge features of the gray-level co-occurrence matrix to the extraction of various types of ice; and through comparison of these schemes, the best method for extracting sea ice spatial information is determined.

**Figure 8.** The similarity of the sea ice and seawater spectral curves for the different sensors.

The textural features represent the surface conditions of the object, such as smooth or rough, which helps to distinguish homogeneous and heterogeneous regions. Since sea ice has irregular and unstable textural characteristics, the textural characteristics of the image were added when the sea ice was extracted in order to solve the problem of sea ice and seawater bearing similar spectra. At present, the commonly used method of extracting sea ice textural features is the gray-level co-occurrence matrix method. The gray-level cooccurrence matrix is a matrix that counts the gray-level relationship between pixels within a certain interval in a local area of an image. The factors affecting the gray level co-occurrence matrix are the image quantization level, the size of the moving window, the movement direction, and the movement step length. The gray-level co-occurrence matrix provides information about the image's gray direction, interval, and change range. Based on the gray-level co-occurrence matrix, the statistical attributes that quantitatively describe the textural features are extracted. Haralick et al. (1973) defined 14 textural features [38]. The feature statistics commonly used to extract textural information from remote sensing images mainly include the mean, variance, homogeneity, contrast, dissimilarity, entropy, angular second moment, and correlation. Recently, many related studies have been conducted on the extraction of sea ice information based on the textural features of the gray level co-occurrence matrix, but there remains a lack of research on the applicability of textural features to various types of ice. Therefore, the applicability and advantages of the different textural features in extracting sea ice types were explored.

Sea ice has conspicuous edge characteristics under different SPM contents. The Sobel edge detection operator has the advantage of easy calculations and a strong anti-noise ability. The edge detection image value of the Sobel operator represents the gradient value of the pixels in the region, and the edge of the sea ice has a higher gradient value. Therefore, the edge points of the sea ice were extracted using the Sobel operator, and the edge point density map is generated with the number of sea ice edge points in a certain range. The edge point density map represents the density of the sea ice edges in a local area. The higher the sea ice density, the greater the number of sea ice edge points. The edge point density map was used to explore the edge characteristics of the sea ice.

In order to further explore the best scheme for extracting the spatial information about the sea ice and to delve deeper into the sea ice spatial information, Scheme 3 combines the edge point density map and the statistics of each textural feature through multiplication, and further explores the ability to combine the edge and textural features to extract the sea ice.

#### *2.5. Object-Oriented Extraction of Sea Ice Extent*

Object-oriented classification refers to the segmentation of images to form objects with adjacent homogeneous pixels, which overcomes the limitations of traditional remote sensing image classification methods that use pixels as the basic classification and processing unit to contain more semantic information [39]. The object is the processing unit, which can achieve a higher level of remote sensing image classification. The object-oriented remote sensing image classification method is not only based on the spectral features but also uses the textural features of the image to segment and classify the image. The classification results avoid speckle noise and have good integrity. Image segmentation is the most basic and critical step in the object-oriented classification method, which directly determines the accuracy of the classification results and the workload of the classification process. In order to improve the accuracy and efficiency of the segmentation, in this paper study, the edge detection segmentation algorithm and the full lambda schedule merge algorithm in ENVI were used.

The blue band is less affected by a high SPM content, and the texture and edge information about the sea ice is clearer. Therefore, the blue band of the image was selected as the reference image to segment the sea ice spectral information and the spatial information in the image. The mean attribute of the segmented object was used to extract the sea ice. Figure 9 shows that the object-oriented method can not only reduce the speckle noise in the classification results, but also limit the influence of the spatial feature window factor when using the spatial features to extract the sea ice and can improve the accuracy of the classification results.

**Figure 9.** Object-oriented segmentation results.
