**3. Results**

#### *3.1. Analysis of Sea Ice Spectral Information Index*

Based on a GF1 image acquired on 12 January 2018, 800 sea ice and seawater sample points were selected, and the sea ice spectral information index was used to plot the distribution ranges of the different types of sea ice and seawater reflectance values (Figure 10). The results revealed that after the sea ice spectral information index was constructed, the reflectance values of the different types of sea ice were larger, while the reflectance values of the seawater were concentrated within a small interval, indicating that the sea ice spectral information index can effectively extract the sea ice spectral information.

**Figure 10.** The range of sea ice spectral information index of sea ice and sea water.

The sea ice can be initially extracted by selecting a suitable threshold. Figure 11 shows the sea ice extraction results. It can be seen that the sea ice spectral information index can effectively extract the sea ice in seawater with different suspended sediment concentrations. The new ice and ice rind in the high suspended sediment area can also be extracted more accurately. However, there are still some problems in the classification results. First, there is the salt and pepper phenomenon, which is a common problem in pixel-based classification methods. This will be solved by object segmentation and extraction. Second, there is still a small amount of confusion between seawater and sea ice in area c, which is mainly concentrated in the areas where the concentration of suspended particles changes drastically. This is because the seawater in these areas display spectral curves that are extremely similar to those of some of the types of sea ice such as new ice and ice rind, and this phenomenon is present in the GF1, Landsat, and Sentinel-2 images (Figure 8). Therefore, it is not possible to completely distinguish between sea ice and seawater using only the spectral characteristics of the image, thus necessitating the addition of the spatial characteristics of the image.

**Figure 11.** Sea ice extraction results using spectral information. (**a**) GF1 image (R/G/B); (**b**–**d**) The three sub-areas of the study area; (**e**–**h**) The extraction results using spectral information, respectively.

#### *3.2. Optimization of Spatial Feature Extraction Scheme*

Scheme 1: The textural feature parameters based on the gray-level co-occurrence matrix mainly include the quantization level, the size of the moving window, and the movement direction and step length. Since the directional characteristics of sea ice are not evident, default values (0,1) were used for the movement directions of the x-axis and y-axis. Moreover, the movement step length was set to a default value of 1. The following section only discusses the quantization level of the gray-level co-occurrence matrix and the moving window size in detail.

Without compressing the gray level of the original image, the size of the gray level co-occurrence matrix is the square of the gray level of the original image, which will greatly increase the calculation load of the gray level co-occurrence matrix. Therefore, in practical applications, in order to improve the efficiency of the calculation of the textural features, the gray level of the original image is usually compressed, and quantization levels of 64, 32, and 16 are generally used.

Figure 12 shows the characteristics of the sea ice and seawater in the GF1 images under different quantitative levels. It can be observed that the images with 64 quantization levels maintain the textural characteristics of the original images better; the images with 32 quantization levels display a reduced ability to maintain details; and the images with 16 quantization levels have lost a significant amount of textural information. Therefore, the higher the quantization level, the better the textural details of the original image are preserved. However, images with high quantization levels are not suitable for the extraction of sea ice textural information. Figure 13 shows the four textural feature indexes of the homogeneity, dissimilarity, entropy, and second moment under different quantization levels. Due to the drastic changes in the concentration of the suspended sediment in the Yellow River Delta, the images with 64 quantization levels exhibit a large amount of speckle noise in the seawater area. In the 32 quantization level images, this speckle noise is greatly suppressed. In addition, since the calculation load increases with increasing quantization level, the calculation efficiency is lower. Therefore, the quantization level of the gray-level co-occurrence matrix was finally set to 32.

**Figure 12.** Image features at different quantization levels. (**a**) Sea ice areas in GF1 images; (**b**–**d**) sea ice images at 64, 32, and 16 quantization levels, respectively; (e) Sea water areas in GF1 images; (**f**–**h**) sea water images at 64, 32, and 16 quantization levels, respectively.

**Figure 13.** Images with different textural feature parameters under different quantization levels. (**a**–**d**) Texture image of homogeneity, dissimilarity, entropy, second moment at 64 quantization levels; (**e**–**h**) Texture image of homogeneity, dissimilarity, entropy, second moment at 32 quantization levels.

The moving window is an important factor that affects the textural feature extraction of the gray-level co-occurrence matrix. Figure 14 shows the distribution range of the textural feature values of various types of sea ice and seawater for different window sizes. It can be seen that the size of the window has little effect on the textural characteristics of the sea ice and seawater, but as the texture window increases, the calculation load increases greatly, thus the window size selected in this study was 3. Based on the statistical results of the textural feature index values of the various types of sea ice and seawater, grey ice and grey-white ice have a higher degree of discrimination from seawater in terms of each textural feature value. The types of thin ice such as new ice, ice rind, and nilas cannot be completely distinguished from the textural characteristics of seawater. This is because the surfaces of the ice rind and nilas are relatively smooth, which is similar to the textural characteristic value of seawater. The surfaces of grey ice and grey-white ice are rough, and the textural characteristic value of seawater is quite different.

**Figure 14.** *Cont*.

**Figure 14.** Plots of the ice water textural characteristic indicators for different window sizes. (**a**) Ice and water texture value distribution in 3 window sizes; (**b**) Ice and water texture value distribution in 5 window sizes; (**c**) Ice and water texture value distribution in 7 window sizes; (**d**) Ice and water texture value distribution in 11 window sizes.

Scheme 2 uses the Sobel operator to generate an edge point density map to highlight the edge features of the sea ice. The distribution ranges of the edge density values of the different types of sea ice and the seawater for different window sizes were calculated (Figure 15), and the optimal calculation window size for the sea ice edge points was compared and analyzed.

**Figure 15.** The effect of the window size on sea ice extraction using an edge point density map.

When the window was small, the edge density value of the seawater basically approached 0, and the edge density values of the grey ice, grey-white ice, and seawater were significantly different. The edge density values of the new ice, ice rind, nilas, and seawater partially overlapped. The overlapping area mainly contained the inner smooth sea ice. As the window size increased, the number of edge points that were detected inside the thin ice region such as new ice, ice rind, and nilas increased, and the edge point density value gradually increased. When the window size reached 45, the edge point density values of the various types of sea ice were significantly different from those of the seawater. As the window continued to grow, it greatly increased the amount of calculation load, thus 45 was selected as the best window size.

In Scheme 3, the texture feature window size was set to 3 × 3, and the quantization level was set to 32. After the edge point density map was combined with the various textural feature indicators, the distribution ranges of the various types of sea ice and the seawater were determined (Figure 16). It can be seen from Figure 16 that the combination of textural feature indicators such as the variance, homogeneity, and contrast with the edge point density map failed to produce a better extraction effect. After the mean textural feature was combined with the edge point density map, the range of the seawater decreased further and became more concentrated, and the distinction between the various types of sea ice and the seawater increased further. Therefore, the edge point density map combined with the mean textural feature index was selected as the final solution for extracting the sea ice spatial information.

**Figure 16.** Box plots for the combinations of the edge density map and textural feature.

Figure 17 shows the comparison between the edge texture information extraction results and the spectral information extraction results. The edge texture image can extract the extent of the sea ice as a whole and can extract the types of ice such as new ice, nilas, grey ice, and grey-white ice. The most important factor is that the texture images at the edges can compensate for the similarity between the spectra of the sea ice and the seawater. As shown in Figure 17j,o, the extraction accuracy of the spectral information is lower in areas where the concentration of the suspended particulate matter changes drastically. The edge texture images solve this problem. Although the seawater in the crevices between portions of ice can also be identified as sea ice, it can be combined with the spectral information to achieve a more accurate sea ice extraction.

#### *3.3. Accuracy Verification*

Figure 18 shows the sea ice extraction results obtained using the different methods for a GF1 image acquired on 12 January 2018. Four scenes including new ice, ice rind, nilas, grey ice, and grey-white ice were selected to illustrate the results of the sea ice extraction. In addition, the results were compared with the sea ice extraction results obtained using the K-Means and SVM methods. Taking into account the complexity of the changes between the various types of sea ice in the seawater with different suspended particulate matter concentrations in the Yellow River Delta, in order to improve the accuracy of the K-Means and SVM methods as much as possible, the K-Means method categories were set to 2–10 categories, and then, the classification post-processing was performed. The image was finally divided into two categories, namely, sea ice and seawater. When the SVM method was employed to select the sample points, the sample points were selected according to the types of ice in the Yellow River delta, turbid seawater, and clear seawater in order to improve the accuracy of the sea ice extraction.

**Figure 17.** Comparison of edge texture results and spectral results. (**a**–**e**) GF1 true color images; (**f**–**j**) results of sea ice extraction from spectral information; and (**k**–**o**) results of the sea ice extraction from edge texture information.

**Figure 18.** Comparison of the sea ice extraction results obtained using different methods. (**a**,**e**,**i**,**m**) True color images of the GF1 image acquired on 12 January 2018; (**b**,**f**,**j**,**n**) Classification results for the method proposed in this paper; (**c**,**g**,**k**,**o**) K-Means classification results; (**d**,**h**,**l**,**p**) SVM classification results.

It can be observed from Figure 18c,h that the K-Means method cannot completely extract the sea ice when extracting thin ice such as new ice and ice rind in seawater with a high suspended particulate matter concentration. As Figure 18g,o shows, most of the seawater was classified as sea ice in the areas with high suspended particulate matter concentrations near the shore and in the clear water areas. This demonstrates that the K-Means method is greatly affected by suspended sediment. The results of the SVM method of extracting sea ice were generally better than those of the K-Means method, but most of the seawater remained classified as sea ice in the areas with high suspended particulate matter concentrations. In addition, there is a significant salt and pepper phenomenon present in the extraction results. The method proposed in this paper can accurately extract the various types of sea ice in both turbid seas and clean seas. It also greatly reduces the salt and pepper phenomenon and improves the integrity of the sea ice extraction.

In order to quantitatively evaluate the accuracy of the sea ice extraction, the overall accuracies and kappa coefficients of the classification results for the GF1, Landsat 8, and Sentinel-2 images were compared and analyzed and additionally compared with those of the K-Means and SVM methods. In addition, the method was applied to the Yellow River Delta and Liaodong Bay. The results are presented in Table 3. The overall accuracy of the method proposed in this paper is basically >95%, the kappa coefficient is > 80%, and the accuracy is 5% higher than those of the SVM and K-Means methods. On 21 January 2017, there were mixed pixels of clouds and water in some areas, which affected the accuracy of the final sea ice extraction. In Liaodong Bay, the accuracy of the SVM was close to that of the method proposed in this paper. This is because the sea ice in Liaodong Bay is predominantly thick ice such as grey ice and grey-white ice, and is less affected by suspended sediment. Therefore, both the proposed method and the SVM method achieved better accuracies.


**Table 3.** Accuracy evaluation table.

The final results show that the accuracy of the K-Means method was the lowest among the three methods. This is due to the similarity between the spectra of the highly turbid seawater and thin ice sheets in the Yellow River Delta and the complexity of the various types of sea ice in the different turbid seawater regions. This led to the relatively low classification accuracy of the K-Means method. The SVM method exhibited a better classification accuracy than the K-Means method overall, but it only used the spectral information, thus the classification accuracy of the ice types, such as in the high suspended sediment areas and for ice rind, was lower. In addition, the SVM method is reliant on prior knowledge. It is a time-consuming process to manually select sample points, and the quality of the sample points directly affects the accuracy of the final classification. The method proposed in this paper attained good accuracy in both the turbid water and clear water areas, and achieved automation of the sea ice extraction. All processing methods were carried out in ENVI. The ENVI functions are called using IDL and can be easily automated.

#### **4. Discussion**

In recent years, extreme weather such as high temperatures, droughts and floods have occurred frequently, and climate anomalies have become the norm, which has led to people's cognitive thinking on global climate change and human living environment [40,41]. As an indicator of global climate change, sea ice change is related to global warming, rises in sea levels and other issues [42,43]. The development of ice conditions in the Yellow River Delta waters in China is unstable, and the formation of sea ice is rapid, which responds more closely to local regional climates. Accurate monitoring of sea ice extent is therefore crucial. Suspended sediment in the mouth of the Yellow River significantly affects the accuracy of sea ice extent extraction. This paper proposes an automatic extraction method of sea ice that combines texture, edge and spectral information, which improves the accuracy of sea ice extraction under highly dynamic suspended sediment changes. Compared with SVM and K-Means, the accuracy is improved by more than 5%. This method provides a basis for accurate sea ice identification using GF1 images, and also offers a method for other optical remote sensing data. High-resolution satellite data based on multiple sources can compensate for the lack of data time resolution and further improve its sea ice monitoring capabilities. Therefore, sea ice monitoring based on multi-source remote sensing data will be the key direction of future development. Moreover, this method provides an approach for other optical remote sensing data, which is of great significance for making full use of multi-source remote sensing data to study the law of sea ice change. Accurate identification of sea ice extent is of great significance to sea ice monitoring, sea ice prediction, disaster prevention and mitigation, and climate research in the Yellow River Delta region. Although this paper discusses the characteristics of various sea ices in detail and enables higherprecision sea ice extraction, it does not distinguish between various sea ice types. Accurate identification of sea ice types is of great significance to the study of sea ice production, ablation and migration. Most of the sea ice in the Yellow River Delta is less than 30 cm thick, and it remains difficult to classify them with greater precision. In addition, the spectrum, texture, and edge information of coastal ice and floes such as grey and white ice are relatively close, and it is difficult to distinguish between coastal ice and floating ice. Therefore, in the future, we will study the distinction of various sea ice types and realize the identification methods of different types of sea ice.

#### **5. Conclusions**

The automatic and accurate extraction of sea ice is essential for studying the laws of sea ice generation and migration, improving sea ice disaster prevention and mitigation, and monitoring climate change. Accurate real-time observations of sea ice bear an important application value and is of theoretical significance.

In order to solve the problem of the low sea ice extraction accuracy caused by the influence of the suspended sediment in the Yellow River Delta, in this study, an automatic sea ice extraction method combining sea ice spectral, texture and edge information is proposed, where the sea ice extraction accuracy can reach over 93%, which is more than 5% higher than SVM and K-means. Compared with previous studies, the sea ice spectral information index suitable for different suspended sediment concentrations is constructed by a two-dimensional scatter diagram of characteristic bands, which improves the applicability of sea ice spectral information index. In changing from discussing the texture characteristics of sea ice as a whole in the past, this study discusses the texture characteristics and edge characteristics of various sea ice types in the Yellow River Delta in detail, laying a foundation for the classification of sea ice types. In addition, the automatic determination of threshold based on OTSU can realize the automatic extraction of sea ice. The method in this paper uses only four bands of visible light and near-infrared to extract sea ice, thus providing a method to be extended to other high-resolution optical remote

sensing data and is of great significance to maximally utilize multi-source remote sensing data for real-time monitoring of sea ice.

In future research, we may expand the research area to the Bohai Sea in China, and realize real-time observation of sea ice through Landsat, Sentinel-2, GF1 and other optical remote sensing data. In terms of data sources, in order to improve the frequency of sea ice monitoring, SAR data may also be applied. We hope to conduct high-precision and high-frequency sea ice monitoring, so as to make a certain contribution to preventing disasters and studying climate change around the Bohai Sea.

**Author Contributions:** Conceptualization, H.Q. and Z.G.; methodology, H.Q.; software, K.M.; validation, H.Q., J.H. and Y.K.; formal analysis, D.Z.; investigation, Z.G. and Y.K.; resources, H.Q.; data curation, H.Q.; writing—original draft preparation, H.Q.; writing—review and editing, H.Q.; visualization, H.Q.; supervision, H.Q.; project administration, H.Q.; funding acquisition, Z.G. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by National Natural Science Foundation of China grant numbers 41971381 and 42071396. This research was funded by National Key R&D Program of China grant number 2017YFC0505903.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Not applicable.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**

