A method for extracting water body time series of disc-shaped sub-lake based on multi-source remote sensing data

Abstract

The application discloses a kind of based on multi-source remote sensing data's disc type sub-lake water body time sequence extraction method, comprising the following steps: obtaining Landsat 8 remote sensing image, MODIS data;All Landsat and MODIS image are mutually registered, and after registration, image is uniformly cut to image by vector data according to the scope of study area;For the month of no cloud of Landsat image, the water surface area is extracted using ISODATA method based on MNDWI, and the accuracy after extraction is verified;For the month of cloud of Landsat image, model is established, and the scale is reduced based on MNDWI;Step five, the water surface area is extracted using ISODATA method based on MNDWI, to obtain the water body reaching Landsat resolution in cloud rain period, fill the water body sequence of missing period of Landsat.The application data is complete, spatial resolution is high, and precision is high.

Description

TECHNICAL FIELD

[0001] The present application relates to a remote sensing extraction method, in particular to a disc type sub-lake water body time series extraction method based on multi-source remote sensing data. BACKGROUND

[0002] Water surface area is the most basic hydrological variable to describe the change of lake water regime, which can reflect the degree of influence of climate factors and human factors on the lake. At present, there are various ways to obtain lake water surface area by remote sensing. According to different satellite sensors, it can be mainly divided into visible light (near infrared) method, microwave remote sensing method and multi-sensor joint method.

[0003] Nowadays, most studies usually combine single-band method and multi-band method to extract water surface area. The band ratio method is based on the multi-band characteristics of multi-spectral remote sensing image, selects the strongest and weakest reflection bands of water body, calculates the ratio between the two, and then sets the optimal threshold to extract the water surface area. Commonly used water body index includes McFeeters' normalized difference water index (NDWI) based on TM image using green band and near infrared band, and Xu's improved normalized difference water index (MNDWI). Numerous studies have shown that using band ratio method to extract water surface area has become the most widely used method in the current research of remote sensing to obtain water area.

[0004] In view of the limitations of single remote sensing sensor in time and space continuity and water surface extraction accuracy, in recent years, the joint processing of multi-source and multi-generation remote sensing data has become one of the frontiers of international earth observation technology development. The research results show that based on the joint means of different sensors, the advantages of different remote sensing data can be fully complementary, so as to expand the application range of various data, truly realize the purpose of obtaining lake water surface area all day and all weather, and it is the inevitable trend of land surface water body remote sensing development.

[0005] Due to the difference of imaging principle and the limitation of technical conditions, the detection performance and application range of any single sensor are limited. Although the medium and high resolution optical satellite image has high precision in extracting water surface, its time resolution is low, which is easily affected by cloudy and rainy days, resulting in the loss of monthly scale data. The time resolution of low resolution optical satellite image is high, which can obtain higher frequency water boundary, and can be used as a supplement to the missing months of medium and high resolution satellite data, but it is limited by low spatial resolution, resulting in low precision.

[0006] And, the lake surface area of the complex terrain condition of the lake, because the disc type sub-lake has the characteristics of small area, rapid change in different periods of the year, if the continuous change information in a long time sequence is explored, it is difficult to meet the observation demand of high spatial resolution and time resolution at the same time by using a single sensor. SUMMARY

[0007] The purpose of the application is to overcome the deficiencies in the prior art, and the purpose of the application is to provide a disc type sub-lake water body time sequence extraction method based on multi-source remote sensing data, which uses scale conversion to reduce the scale of low-resolution remote sensing data to decompose water surface information and obtain medium-high resolution water surface information.

[0008] Technical scheme: The disc type sub-lake water body time sequence extraction method based on multi-source remote sensing data, comprising the following steps:

[0009] Step one, acquire Landsat 8 remote sensing images and MODIS data in the research area, and take Landsat and MODIS satellite images as the main data source and take GF-1 data as the verification data;

[0010] Step two, mutually register all Landsat and MODIS images, and uniformly cut the images according to the research area range by using vector data after registration;

[0011] Step three, for the cloud-free months of Landsat images, use the ISODATA method based on MNDWI to extract the water surface area, and compare the accuracy after extraction with the NDWI threshold method, the NDWI-ISODATA method and the MNDWI threshold method;

[0012] Step four, for the cloud months of Landsat images, establish a model and perform scale reduction based on MNDWI;

[0013] Step five, use the ISODATA method based on MNDWI to extract the water surface area to obtain the water body reaching the resolution of Landsat in the cloud and rain period, and fill in the water body sequence of the missing period of Landsat.

[0014] Further, in step one, the accuracy of Landsat 8 remote sensing images is 30m, and the accuracy of MODIS data is 500m, the preprocessing includes image registration and image cutting, the accuracy of registration is that the root mean square error is less than 0.5 pixels, and the images after registration are uniformly cut by using vector data according to the research area range.

[0015] Further, in step three, the sample points for verifying the accuracy are generated in the rectangular range of the three disc type lake boundaries by using ArcGIS with n x n grid.2 A point, wherein the points falling within the lake area are taken as verification sample points, the extraction results of each water body extraction method are compared with the reference value, the number of water bodies and non-water bodies falling in the three disc-shaped lakes of the sample points are respectively counted, and a confusion matrix between the reference value is established, and the precision evaluation index is calculated by using the confusion matrix.

[0016] Further, the precision evaluation index includes misclassification rate, omission rate, overall precision and Kappa coefficient, and the highest precision in which water body extraction method is determined. The misclassification rate refers to the ratio of sample points that are actually non-water bodies but are incorrectly classified as water bodies to the total sample points, and the value range is 0-1. The omission rate refers to the ratio of sample points that are actually water bodies but are not classified as water bodies to the total sample points, and the value range is 0-1. The overall precision is the percentage of the number of all correctly classified samples to the total number of samples, and the value range is 0-1.

[0017] Further, the calculation formula of Kappa coefficient is:

[0018]

[0019] Wherein, ∑=(TP+FP)×(TP+FN)+(FN+TN)×(FP+TN), T represents the number of all sample points, TP is the number of samples that are actually water bodies but extracted as water bodies, FP is the number of samples that are actually non-water bodies but extracted as water bodies, FN is the number of samples that are actually water bodies but extracted as non-water bodies, and TN is the number of samples that are actually non-water bodies but extracted as non-water bodies.

[0020] Further, in step four, according to the MODIS data and Landsat data, MNDWI is calculated, MODIS_MNDWI and Landsat_MNDWI are obtained, and MODIS_MNDWI is resampled to be consistent with the resolution of Landsat_MNDWI. The resampling method is the nearest neighbor method, and based on each pixel, a linear regression model between the resampled MNDWI value and the corresponding high-resolution MNDWI value is established:

[0021]

[0022] In the formula, And Respectively represent the high-resolution image MNDWI value and the low-resolution image MNDWI value at time t, pixel position (i, j); a i,j , b i,j Indicate the correlation regression coefficient.

[0023] The MODIS low-resolution MNDWI value is taken as the independent variable, the regression model is used for downscaling, and the Downscaled_MNDWI of the disc-shaped sub-lake with 30m resolution is obtained.

[0024] Working principle: for the image cloud-free months, directly extract the water area; for the image cloud months, use different remote sensing data to establish a linear regression model, use the downscaling method to obtain the missing data, and then extract the water area, so as to realize the water body time series extraction of multi-source remote sensing data.

[0025] Beneficial effects: compared with the prior art, the present application has the following remarkable features: complete data on a monthly scale, high spatial resolution and high precision; solves the problem of water surface information acquisition during the cloud period of the image; high water extraction precision. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is the flowchart of the present application;

[0027] Figure 2 is the location map of the research area of the present application;

[0028] Figure 3 is the schematic diagram of the water area extraction method of the present application in the research area;

[0029] Figure 4 is the water extraction result comparison chart of the present application;

[0030] Figure 5 is the verification sample point distribution diagram of the present application;

[0031] Figure 6 is the comparison chart of the water surface extraction result of the saucer-shaped lake in the dry season of the present application;

[0032] Figure 7 is the comparison chart of the water surface extraction result of the saucer-shaped lake in the wet season of the present application;

[0033] Figure 8 is the comparison chart of the water surface extraction result of the saucer-shaped lake in the dry season of the present application;

[0034] Figure 9 is the scatter diagram between Landsat and MODIS of the present application;

[0035] Figure 10 is the scatter diagram between Landsat and Downscaled_MODIS of the present application. DETAILED DESCRIPTION

[0036] As Figure 1 , the saucer-shaped sub-lake water body time series extraction method based on multi-source remote sensing data includes the following steps:

[0037] Step 1, data preparation. Taking the saucer-shaped sub-lake of Poyang Lake as the research object, the research area is as Figure 2The Landsat (30m), MODIS (500m) satellite images and other data are used as the main data sources, and the GF-1 data is used as the verification data. The research data is shown in Table 1.

[0038] Table 1 Research data

[0039]

[0040] Step two, data preprocessing. Landsat 8 remote sensing images and MODIS data in the research area are obtained and preprocessed. The preprocessing mainly includes image registration and image cropping. Specifically, in order to ensure that different data are matched in space, all Landsat and MODIS images are mutually registered, and the registration accuracy is less than 0.5 pixels. In order to reduce the amount of calculation, the registered images are cropped according to the research area range, and the images are uniformly cropped by using vector data.

[0041] Step three, water area extraction and accuracy verification. For the months without cloud in the Landsat image, the ISODATA method based on MNDWI is used to extract the water area. In order to verify the accuracy and compare with other methods, three typical disc-shaped lakes in the Poyang Lake area, namely Banghu, Linchong Lake and Taiyang Lake, are taken as the research objects, which are located in different spatial positions of the Poyang Lake (east, south and west of the Poyang Lake). Among them, Banghu is located in the west of the Poyang Lake, with a geographical position of 115°54'~116°01'E, 29°10'~29°17'N, and an area of about 71.22km 2 ; Linchong Lake is located in the south of the Poyang Lake, with a geographical position of 116°14'~116°19'E, 28°49'~29°53'N, and an area of about 28.72km 2 ; Taiyang Lake is located in the east of the Poyang Lake, with a geographical position of 116°35'~116°37'E, 29°13'~29°15'N, and an area of about 7.58km 2 . The Landsat 8 image used in the experiment is dated November 20, 2015, at which time the Poyang Lake is in the dry season, and various types of ground objects such as water area, shoal, vegetation coexist, which is more typical and suitable for the research of water area extraction method. The positions of the three disc-shaped lakes in the Poyang Lake are shown in Figure 3 .

[0042] Since ISODATA algorithm is a pixel-based region segmentation algorithm, there may be "salt and pepper noise" in the classification results, which may be caused by the "same object different spectrum" phenomenon of image tone caused by image imaging or noise in the target scene. Therefore, when using the water body index-ISODATA method to extract the water surface area, the different categories in the initial classification results are continuously merged in combination with the results of artificial visual discrimination until a better classification effect is achieved, so as to eliminate the influence of noise.

[0043] Based on the Landsat image of the study area, the extraction results of the water body index threshold method, the ISODATA method based on NDWI and the ISODATA method based on MNDWI are compared, and the extraction results are as follows Figure 4 The four methods can effectively extract most of the water bodies in the study area. In order to further intuitively compare the advantages and disadvantages of several methods, the accuracy of the extraction results needs to be evaluated. The accuracy evaluation of remote sensing image feature classification is to compare and analyze the existing reference classification results with the classification results to be tested, so as to quantitatively judge the accuracy of the water body information extraction method. The reference value is obtained by using the ISODATA method to extract the water surface area of the GF-1 No. 2m fusion image with a date close to the image date. For the selection of verification sample points, n 2 points are generated in the rectangular range of the three disc-shaped lake boundaries by using ArcGIS, and the points falling in the lake area are used as verification sample points, which are distributed as shown in Figure 5 .

[0044] The extraction results of each water body extraction method are compared with the reference value, and the number of water bodies and non-water bodies falling in the three disc-shaped lakes is counted, and the confusion matrix between the reference value and the reference value is established. The confusion matrix (Confusion Matrix) method compares the categories of the selected sample points and the corresponding categories in the image classification results and counts and calculates, which is a commonly used method in the evaluation of remote sensing feature classification. The confusion matrix is shown in Table 2.

[0045] Table 2 Confusion Matrix

[0046]

[0047] The following accuracy evaluation indexes are calculated by using the confusion matrix, including misclassification rate (Misclassification Rate), leakage classification rate (Leakage Classification Rate), overall classification accuracy (Overall Classification Accuracy) and Kappa coefficient.

[0048] (1) Misclassification rate (MR)

[0049] The misclassification rate refers to the ratio of sample points that are actually non-water bodies but are incorrectly classified as water bodies to the total sample points, and the value range is 0-1.

[0050]

[0051] (2) The omission rate (LR)

[0052] The omission rate refers to the ratio of sample points that are actually water bodies but are not classified as water bodies to the total sample points, and the value range is 0-1.

[0053]

[0054] (3) Overall accuracy (OA)

[0055] The overall accuracy refers to the percentage of the number of all correctly classified samples to the total number of samples, and is an important indicator for representing the quality of land cover classification, with a value range of 0-1.

[0056]

[0057] (4) Kappa coefficient

[0058] The Kappa coefficient is an important indicator for representing the consistency of land cover classification, and it is generally considered that a value above 0.8 indicates good classification results.

[0059]

[0060] In the above formula, ∑ = (TP + FP) × (TP + FN) + (FN + TN) × (FP + TN), and T represents the total number of sample points.

[0061] Table 3 Accuracy evaluation of water body extraction results

[0062]

[0063] The accuracy index is calculated according to the above formula, and the accuracy result is shown in Table 3 above. From the statistical results of the misclassification rate and the omission rate in Table 3, for Banghu, the misclassification rate of the MNDWI threshold method is the highest, followed by the NDWI threshold method and the MNDWI_ISODATA method, and the misclassification rate of the NDWI_ISODATA method is the lowest. The omission rate of the NDWI_ISODATA method is the highest, followed by the NDWI method, the MNDWI_ISODATA method, and the omission rate of the MNDWI is the lowest. For Linchong Lake, the misclassification rate of the NDWI_ISODATA method is the highest, and the misclassification rates of the two water body index threshold methods are the same and the lowest. The omission rate of the NDWI is the highest, and the omission rate of the MNDWI_ISODATA method is the lowest. For Taiyang Lake, the misclassification rate of the NDWI threshold method is the lowest, and the misclassification rate of the water body index_ISODATA method is the same and the highest. The omission rate of the NDWI threshold method is the highest, and the omission rate of the MNDWI_ISODATA method is the lowest. From the statistical results of the overall accuracy and the Kappa coefficient, the overall accuracy of the four water body extraction methods for the extraction of the disc-shaped lake is more than 90%, among which the overall accuracy of the MNDWI_ISODATA method is the highest in the three disc-shaped lakes, which is 97.9675%, 98.4127% and 96.9925% respectively. For the Kappa coefficient, the Kappa coefficient of each extraction method is more than 0.89, among which the Kappa coefficient of the MNDWI_ISODATA method is the highest, and the Kappa coefficients of the three disc-shaped lakes are 0.9368, 0.9544 and 0.8707 respectively, so this method is suitable for the extraction of the water surface area of the disc-shaped lake.

[0064] Step four, for the months with clouds of Landsat images, establish a model, and carry out MNDWI-based downscaling, the specific process is as follows:

[0065] (1) Image screening. First, the Landsat 8 images covering the study area and having good quality in 2013-2020 are screened, and the date of each image is recorded; secondly, according to the image screening results, the MODIS images covering the study area, having good quality and being consistent or similar in date (within 2 days) in 2013-2020 are screened, part of which will be used for downscaling research, and the other part will be used for verification. In order to ensure the reliability and accuracy of the accuracy verification, the verification data set is selected according to the standard of being evenly distributed in each year and in the wet season and dry season, and the rest of the data is used as the modeling data set. The image dates are shown in Table 4 below; finally, select the images of the months in which Landsat 8 is missing in 2013-2020 from the MODIS data, and this part of data will be used for the construction of the monthly water surface area data set.

[0066] (2) Downscaling. MNDWI of each image of Landsat 8 and MODIS data obtained was calculated, for MODIS_MNDWI, nearest neighbor resampling method was used to resample to 30m resolution, using Matlab to carry out pixel by pixel fitting regression on Landsat_MNDWI(30m) and resampled MODIS_MNDWI(30m), to obtain slope image and intercept image. Finally, for MODIS_NDWI(500m) used to construct water surface area data set, using the obtained slope image and intercept image, combined with formula (1), to obtain Downscaled_MNWDI with resolution of 30m.

[0067] Step five, water surface extraction. Based on Landsat_MNDWI and Downscaled_MNWDI obtained by downsizing, ISODATA method based on MNDWI was used to extract water surface area.

[0068] Table 4 corresponding Landsat 8 and MODIS data date list

[0069]

[0070] Step six, accuracy verification. In terms of qualitative evaluation, the shape contour and spatial distribution of water surface extraction results based on downsized MODIS_MNDWI in different periods of the year (wet period and dry period) have been greatly improved compared with the original MODIS_MNDWI, and show good consistency with Landsat_MNDWI, such as Figures 6-7 , the dry period is March 14, 2014, and the wet period is September 14, 2017. Table 5 below is the accuracy evaluation data of MODIS and Downscaled_MODIS water surface extraction results. In terms of quantitative evaluation, the water extracted by downsized MODIS_MNDWI has obviously improved in the missed classification rate and the wrong classification rate compared with the original MODIS_MNDWI, and the overall accuracy and Kappa coefficient (87.39% and 0.7476 respectively) have also been significantly improved, such as Figure 8 In addition, the water surface area extracted based on downsized MODIS_MNDWI and Landsat_MNDWI has better fitting compared with the water surface area extracted by the original MODIS_MNDWI and Landsat_MNDWI, which is manifested as the determination coefficient R2 increases from 0.9612 to 0.9876, the RMSE decreases from 50.0299km 2 to 20.3167km 2 and the slope of the fitting straight line is closer to 1, such as Figures 9-10 .

[0071] Table 5 Accuracy assessment of water surface extraction results of MODIS and Downscaled_MODIS

[0072]

Claims

1. A method for extracting the time series of water bodies in a saucer-shaped sub-lake based on multi-source remote sensing data, characterized in that, Includes the following steps: Step 1: Acquire Landsat 8 remote sensing images and MODIS data within the study area, using Landsat and MODIS satellite images as the main data sources and GF-1 data as validation data. Step 2: Register all Landsat and MODIS images together. After registration, crop the images uniformly according to the study area using vector data. Step 3: For months with no clouds in the Landsat imagery, the water surface area is extracted using the ISODATA method based on MNDWI. After extraction, the accuracy is compared with the NDWI threshold method, NDWI-ISODATA method, and MNDWI threshold method to verify the accuracy. Step 4: For months with clouds in Landsat imagery, build a model and perform downscaling based on MNDWI. Step 5: The water surface area is extracted using the ISODATA method based on MNDWI to obtain the water bodies that reach Landsat resolution during the cloud and rain period, thus filling in the water body sequence during the missing periods of Landsat. In step four, MNDWI is calculated based on MODIS data and Landsat data to obtain MODIS_MNDWI and Landsat_MNDWI. MODIS_MNDWI is then resampled to match the resolution of Landsat_MNDWI. The resampling method is the nearest neighbor method, which establishes a linear regression model between the resampled MNDWI value and the corresponding high-resolution MNDWI value for each pixel: In the formula, and represents the MNDWI value of the high-resolution image and the MNDWI value of the low-resolution image at time t and pixel position (i, j), respectively. , This represents the relevant regression coefficient.

2. The method for extracting the time series of water bodies in a dish-shaped sub-lake based on multi-source remote sensing data as described in claim 1, characterized in that: In step one, the Landsat 8 remote sensing image has a resolution of 30m, and the MODIS data has a resolution of 500m. Preprocessing includes image registration and image cropping. The registration accuracy is a root mean square error of less than 0.5 pixels. The registered images are then uniformly cropped using vector data according to the study area.

3. The method for extracting the time series of water bodies in a dish-shaped sub-lake based on multi-source remote sensing data as described in claim 1, characterized in that: In step three, the selection of sample points for verifying accuracy is achieved by uniformly generating n×n grids within the rectangular areas of the three saucer-shaped lake boundaries using ArcGIS. 2 Points were selected, with those falling within the lake area serving as validation sample points. The extraction results of each water extraction method were compared with the reference values. The number of water bodies and non-water bodies falling within the three saucer-shaped lakes was counted, and a confusion matrix was established between the sample points and the reference values. The accuracy evaluation index was calculated using the confusion matrix.

4. The method for extracting the time series of water bodies in a dish-shaped sub-lake based on multi-source remote sensing data according to claim 3, characterized in that: The accuracy evaluation indicators include misclassification rate, omission rate, overall accuracy, and Kappa coefficient, to determine which water extraction method has the highest accuracy.

5. The method for extracting the time series of water bodies in a dish-shaped sub-lake based on multi-source remote sensing data as described in claim 4, characterized in that: The misclassification rate refers to the ratio of sample points that are actually non-water bodies but are incorrectly classified as water bodies to the total number of sample points, and its value ranges from 0 to 1.

6. The method for extracting the time series of water bodies in a dish-shaped sub-lake based on multi-source remote sensing data as described in claim 4, characterized in that: The omission rate refers to the ratio of sample points that are actually water bodies but were not classified as water bodies to the total number of sample points, and the value ranges from 0 to 1.

7. The method for extracting the time series of water bodies in a dish-shaped sub-lake based on multi-source remote sensing data as described in claim 4, characterized in that: The overall precision is the percentage of all correctly classified samples out of the total number of samples, and its value ranges from 0 to 1.

8. The method for extracting the time series of water bodies in a dish-shaped sub-lake based on multi-source remote sensing data as described in claim 4, characterized in that: The formula for calculating the Kappa coefficient is as follows: in, T represents the total number of sample points, TP represents the number of samples that were actually water bodies but were extracted as water bodies, FP represents the number of samples that were actually non-water bodies but were extracted as water bodies, FN represents the number of samples that were actually water bodies but were extracted as non-water bodies, and TN represents the number of samples that were actually non-water bodies but were extracted as non-water bodies.

Images