1.

INTRODUCTION

In recent decades, influenced by global climate change and human activities, shrinking waters has become one of the main environmental problems facing mankind¹. While the remote sensing technology will lead humans’ understanding of the earth’s surface to a new stage, and offer convenience to large-scale, real-time water surveys². Compared with traditional survey methods, remote sensing technology has the characteristics of wide coverage, strong macroscopic, speed, multi-temporal and rich comprehensive information. It can quickly and accurately identify water bodies and monitor the dynamic changes of water bodies in time series, which has become one of the important methods of water environment research³.

The electromagnetic wave wavelength used by multi-spectral remote sensing is generally between 0.4 and 2.5 microns, and the absorption of electromagnetic waves in this band is significantly higher than that of most other ground objects, so the total radiation level of the water body is relatively low. In the infrared band, the energy absorbed by water is higher than that of visible light. Even if it is very shallow, the water absorbs almost all the incident energy in the near-infrared and mid-infrared bands, with very little reflected energy⁴; while vegetation and soil absorb few energy in these two bands and have high reflection, which makes the water body have obvious difference from vegetation and soil in these two wavebands. In the image, the water body presents a dark tone, while the soil and vegetation are brighter⁵.

In recent years, domestic and foreign research methods on water body information extraction mainly include single-band, multi-band and water index methods. Liu used the 4th, 5th and 7th bands of TM remote sensing images, combining with the density segmentation method, obtained the changes in the area of Beijing Miyun Reservoir⁶. Zhong studied the method of extracting wetland water information by using the ratio difference (TM2+TM3)/(TM4+TM5) of different features in TM images⁷. According to the characteristics of high-resolution remote sensing images, Yin used object-oriented methods to extract water from high-resolution remote sensing images⁸.

Previous studies have shown that different methods have different effects in identifying water bodies, but most of these studies focus on water bodies in a single area, there are few effective studies on each method in different areas. So this paper selects two areas as the research objects to compares the three commonly used water extraction models, including the CIWI model, NDWI and near-infrared band model, and multi-band spectrum relationship model and analyzes their effectiveness in the identification of lake water bodies from the Landsat8 data and their applicability in different water areas. The research of this technology can not only monitor the dynamic of boundary rivers in Heilongjiang Province, but also promote the development of corresponding technologies and the management of boundary rivers.

2.

THE RESEARCH AREA AND DATA SOURCES

The study area 1 selects the Northeastern Heilongjiang Province and the border river basins adjacent to Shuangyashan City and Jixi City as the research object. At the same time, in order to compare the extraction effect of the above models in the urban area, the Songhua River basin in the center of Harbin is selected as the study area 2.

According to the requirements of time phase, phenology and cloud cover, the data source is Landsat8 remote sensing image with 30m resolution, one is Landsat8 images of covering the study area in August 2017 and the other is Landsat8 images of covering study area 2 in September 2017. The original image data of the study area is shown in Figure 1.

Figure 1.

Image data of experimental area.

3.

RESEARCH METHOD

By analyzing the characteristics of Landsat8 data, three methods are selected for experiments: CIWI model, NDWI and near-infrared band model, and multi-band spectrum relationship model to verify the applicability of each model in different regions of water bodies.

3.1

CIWI

CIWI is a water body extraction feature model that combines the normalized difference vegetation index (NDVI) and near infrared (NIR). The principle is to add the reflectivity of the near infrared band on the difference of radiation between water bodies, vegetation and land reflected by the NDVI, and to further improve the distinction between water bodies and other features through multiplicative factors. The CIWI calculation formula is as follows:

In the formula: B_nir and represent the reflectivity in the near-infrared band; B_r represents the reflectivity in the red band. For convenient of analysis and comparison, a dimensionless parameter C is introduced in the CIWI index, usually a constant greater than 0, which only amplified and translated the calculation results, to ensure that the calculation results are in the positive interval. And the parameter C is set to 0.8 in this model. If the satellite data has only one near-infrared band, B_nir and take the same near-infrared band value.

3.2

NDWI

NDWI, the normalized difference water index, is calculated based on the reflectivity of the near-infrared band and the green light band⁹. The calculation method is as follows:

B_green and B_nir represent the reflectivity of the green light band and the near-infrared band respectively. In the NDWI and near-infrared band models, the water extraction rules are:

T₁ and T₂ are thresholds. As the wavelength increases, the reflection of the water body gradually weakens from visible light to mid-infrared, and its absorption is strongest in the near-infrared and mid-infrared wavelength range, with almost no reflection⁹. However, the NDWI index of some buildings or soils is also positive, which is easily confused with water bodies, and adding NIR < T₂, conditions can improve the effect of water surface extraction^{10, 11}.

3.3

Multi-spectral relationship method

The water body also has different spectral response characteristics in each band of the image, and the multi-band spectrum relationship method uses this difference characteristics between the spectrum to extract water body information. According to the existing data, only the water body has the characteristics that the green band plus the red band is greater than the near-infrared band plus the short-wave infrared band, so this relationship between the spectrum can be used to extract the water body information. The method of multi-band spectrum relationship based on threshold is usually adopted, that is, by selecting a certain threshold T, the water body meets the following formula, and the other ground features are not¹¹. The extraction model of the water body is as follows:

B_green / B_red / B_nir / B_mir represent the reflectivity of the green, red, near-infrared, and short-wave infrared bands in the image.

4.

RESULTS AND DISCUSSION

The study area is selected for case analysis, and the three models described above are used for experiments. The experimental results are shown in the following figures. Figure 2 is the water body extraction result of the CIWI model in the study area1; Figure 3 is the water body information extraction result of the NDWI and near-infrared waveband model; Figure 4 is the water body information extraction result of the multi-band spectrum relationship model, white represents the water body area, and the black represents non-water body area in each result.

Figure 2.

CIWI model (Area 1).

Figure 3.

CIWI and near-infrared band models (Area 1).

Figure 4.

Multi-band spectrum relationship model (Area 1).

In order to facilitate comparison and analysis, the. extraction results of CIWI, NDWI and near-infrared bands and multi-band spectrum relationship model in the study area 1 was evaluated respectively. 500 sample points were selected from the results of the water body index to construct a confusion matrix and calculate the overall classification accuracy. The accuracy evaluation results show that the accuracy of the three models is 95.0%, 94.4% and 93.4% respectively. The water extraction effect of the CIWI model is slightly better than that of the NDWI and the near-infrared model and the multi-band spectrum relationship model. But there is no significant difference in the extraction effect of these models. All of them can accurately identify most of the water body area, and some water bodies in small areas can also be extracted effectively, which is clearly visible and well-defined.

The difference in accuracy is mainly due to the fact that there are a small number of thick clouds in the northwest corner of the boundary river basin image. In the multi-band spectrum relationship model, the water body and cloud shadows are mixed seriously, so the impact of thick clouds cannot be eliminated well; NDWI and near Infrared bands can improve the discrimination between cloud shadows and water bodies to effectively remove part of influence of cloud shadows, but it is still not ideal in a small part of the area because many large-scale cloud shadows are still mixed in it. However the cloud shadows is smaller than that of multi-band spectrum relationship model and the best one is the CIWI model, in which the outline of the boundary river water body extracted is clear with no confusion of cloud shadows.

Figures 5-7 show the extraction results of CIWI, NDWI and near-infrared bands and multi-band spectrum relationship model in the study area 2. The confusion matrix is also constructed and the overall classification accuracy is calculated. The accuracy analysis results show that the accuracy of the three models is 91.2%, 91.8% and 90.2%, respectively. The water extraction effect of NDWI and near-infrared band models is slightly better than CIWI model and multi-band spectrum relationship model. Among them, although the multi-band spectrum relationship model can completely extract different water body information, it also extracts a large amount of non-water body information, so the extraction effect is the most unsatisfactory compared with other models.

Figure 5.

CIWI model (Area 2).

Figure 6.

CIWI and near-infrared band models (Area 2).

Figure 7.

Multi-band spectrum relationship model (Area 2).

In addition, the three models are able to extract rivers and other small water bodies, but also mistakenly extract many buildings on the south bank of the Songhua River. This is mainly because the information of buildings in residential areas is similar to the inland water body information. But relatively speaking, the NDWI model is better than the other two models in distinguishing water bodies and buildings. Therefore, the further research is needed to effectively eliminate the influence of buildings in water environment research.

5.

CONCLUSION

This paper takes Landsat8 imagery as the data source, using CIWI model, NDWI and near-infrared band model and multi-band spectrum relationship model to extract the water body of the research areas. The overall classification accuracy is selected as the evaluation index to study potential problems in different water body extraction models. The study came to the following conclusions:

In remote sensing images, the spectral characteristics of water bodies are significantly different from other ground objects. Therefore, the three models can effectively extract large areas of water bodies with the overall classification accuracy above 90%. The CIWI model has a great advantage in the discrimination of cloud shadows, while the NDWI and near-infrared band model can effectively eliminate the interference of buildings when used in urban areas.

The CIWI model and the NDWI and near-infrared band model are better choices for water body information extraction. However, the spatial resolution of Landsat8 data is 30m, which is of great value for the monitoring of large-area water bodies and has certain defects in monitoring small-area water bodies, such as the lower extraction accuracy of small area water bodies. In conclusion, analyzing the distribution characteristics of water-body information from Landsat8 data has a better supporting effect on the investigation, macro-monitoring and protection of water resources in boundary rivers.