Lake edge change detection method

Abstract

The invention provides a lake edge change detection method, and belongs to the technical field of environmental ecological remote sensing. The method comprises the following steps: firstly, acquiring Landsat time series data and MODIS time series data around a lake, and then filtering and fusing the two data sources by using a GF-SG method to obtain high-quality NDVI time series data; performing median synthesis on the data of the time period every year according to the lake melting period to obtain an NDVI annual curve; according to the characteristics of the water body on the NDVI, a region where the lake boundary may change is screened, and then LandTrendr change detection is performed on the region to obtain the lake edge change condition. According to the lake edge change detection method provided by the invention, the lake edge change condition can be effectively detected, a mode of combining two kinds of data is used, and the result is obviously superior to the detection result of a single data source.

Description

Technical Field

[0001] The invention relates to the technical field of environmental ecological remote sensing, and in particular to a lake edge change detection method. Background Art

[0002] Lakes are a vital component of global water resources. They are not only a key link in the natural water cycle, but also play an important role in storing, purifying and regulating water resources. More importantly, lakes are regarded as sensitive sentinels of climate change because they are extremely sensitive and responsive to climate and basin hydrological changes. The area, water level, water quality and other conditions of lakes and changes in ecosystems can directly or indirectly reflect the trend and impact of climate change, providing us with important monitoring and early warning information. As an important water resource, lakes are of great significance to local water supply and surrounding ecosystems. The reduction of lakes will affect the water supply of surrounding ecosystems and pose a threat to ecological balance. On the other hand, the expansion of lakes will increase the water level of lakes and increase the water storage capacity, which can increase soil moisture and promote the growth of vegetation in the non-flooded areas of lakes, which is beneficial to vegetation restoration and desertification control. At the same time, the expansion of lakes will change the morphology of lakes and change the hydrological connection between rivers and lakes, and the hydrological circulation will affect vegetation productivity, biodiversity, animal husbandry and many other aspects.

[0003] Lake extraction and change detection are beneficial to studying the expansion of lakes, but the remote and vast land and harsh natural environment make the measurement data scarce and expensive. Since remote sensing has the characteristics of large monitoring range, high timeliness and little influence from the ground environment, most of the existing detection methods are based on remote sensing images to extract and detect lake changes. They can be mainly divided into the following categories: visual interpretation, threshold method combined with spectral index, machine learning method, deep learning method, and change detection method based on time series. Visual interpretation is easily affected by experience and image quality, and has high requirements for the spatiotemporal resolution of images; the threshold method combined with spectral index generally calculates water indexes such as normalized difference water index (NDWI) and modified normalized difference water index (MNDWI), and determines its threshold to segment and extract lakes; the change detection method based on time series regards the time series as an overall trajectory, and obtains the difference in spatiotemporal information by detecting the changes in the trajectory. Commonly used methods mainly include BFAST method, DBEST method, LandTrendr method, etc.; machine learning methods and deep learning methods mainly achieve lake extraction effects through sample training models.

[0004] At present, the extraction of lakes is still mainly based on the extraction of the entire lake, and little attention is paid to the extraction of changes in the lake edge. Landsat and MODIS data are widely used in lake extraction work. However, these optical remote sensing images (such as Landsat and MODIS data) are difficult to extract the complete lake body and lake boundary due to the pollution of clouds and cloud shadows when monitoring lake areas, especially in areas with less optical image data (such as the Qinghai-Tibet Plateau). Landsat has a high spatial resolution, but during the melting period of lakes on the Qinghai-Tibet Plateau, Landsat remote sensing images with less than 10% cloud cover are even less than 30% of the total, which greatly limits the acquisition of simultaneous data, and the uncertainty of seasonal changes in lakes will be introduced into the results. Although MODIS with a lower spatial resolution is less affected by cloud pollution than Ladnsat, the low spatial resolution reduces MODIS's sensitivity to changes in lake details, making it difficult to detect changes in lake details. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a lake edge change detection method. By fusing Landsat and MODIS data, the method can effectively deal with the influence of inconsistency in remote sensing phases and poor quality of remote sensing images on lake edge extraction, and can more accurately extract the edge changes of the lake each year.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A lake edge change detection method is used to identify and extract lake edge change conditions, and is characterized by comprising the following steps:

[0008] S1. Obtain the Landsat time series data and MODIS time series data of the specified lake area over the years.

[0009] S2. Use the GF-SG (Gap Filling and Savitizky-Golay) method to filter and fuse the two data to obtain high-quality and high-resolution NDVI (Normalized Difference Vegetation Index) time series for each year.

[0010] S3. Intercept the NDVI time series in the lake melting period in each year's NDVI time series; in the intercepted NDVI time series, each pixel corresponds to a pixel time series, and the median of the pixel time series is taken as the data of the pixel for the year, and the data of all pixels for the year are combined to obtain a fused image of the year. The data of previous years are processed in the same way to obtain the fused images of previous years, forming an inter-annual NDVI time series.

[0011] S4. Based on the lake buffer area screening principle, the interannual NDVI time series corresponding to all pixels are screened to obtain pixels that meet the screening conditions. All screened pixels constitute the lake buffer mask; the lake buffer mask is spatially filtered to obtain the area where the lake boundary may change; the interannual NDVI time series corresponding to each pixel in the area where the lake boundary may change is obtained; the interannual NDVI time series corresponding to all pixels are used for lake edge change detection using the LandTrendr detection model to identify the changes in the lake edge.

[0012] Furthermore, in step S1, the Landsat and MODIS time series data are selected from the same year.

[0013] Furthermore, in step S2, first, the bicubic interpolation algorithm is used to sample MODIS data so that the sampled data has the same spatial resolution as the Landsat data; then, during the filtering and fusion process using the GF-SG method, the size of the local spatial adjacent window is set to 40 pixels × 40 pixels and the correlation coefficient is set to 0.8; because the research object is water bodies, which appear as negative values ​​on the NDVI, no threshold is set.

[0014] Furthermore, in step S3, the lake melting period is set to be the same every year.

[0015] Furthermore, in step S4, the lake buffer area screening principle is as follows: (1) the minimum value of the interannual NDVI time series corresponding to the pixel is less than 0; in this case, the pixel is regarded as a water body; (2) the average value of the three years after the maximum mutation point of the interannual NDVI time series corresponding to the pixel is less than 0, and the average value of the three years before the maximum mutation point is greater than 0; in this case, the pixel is considered to have changed.

[0016] Furthermore, in step S4, the lake buffer mask is spatially filtered by median filtering, and the filter window size of the median filter is set to 5 pixels × 5 pixels to obtain areas where the lake boundary may change.

[0017] Furthermore, in step S4, the parameters of the LandTrendr detection model are set as follows: the maximum number of segments is 6; the spike threshold is 0.9; the number of vertex overshoots is 3; preventing one-year recovery is false; the recovery threshold is 1; the p-value is 0.01; the best model ratio is 1; the minimum number of observations is 6; and the magnitude of change is 0.1.

[0018] Compared with the prior art, the present invention has the following beneficial effects: in step S2, by weighted fusion of long-time series massive images within the spatial neighborhood to restore missing image values, the reliability and continuity of the data are improved, allowing the present invention to demonstrate significant advantages in lake change detection with high timeliness; in step S4, the LandTrendr detection model is used as a detection method, the overall trend of the time series is considered, and the overall change trend of the lake is considered from the perspective of the time series, reducing noise interference in lake change detection. In addition, the present invention can detect disturbances of different amplitudes and times by setting different parameters of the LandTrendr detection model or changing the input data of the model, and has high flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a lake edge change detection method according to an embodiment of the present invention;

[0020] Figure 2 This is a detection map of edge changes of three lakes of different sizes at different locations according to an embodiment of the present invention;

[0021] Figure 3 Comparison between the edge changes of lakes from 2016 to 2020 obtained based on the method of the present invention and the actual images of Planet with high spatial resolution from 2016 to 2020;

[0022] Figure 4 It is the accuracy map of 100 randomly sampled changing lakes in the embodiment of the present invention;

[0023] Figure 5 It is a recall rate graph of 100 randomly sampled change lakes in an embodiment of the present invention;

[0024] Figure 6 It is the accuracy verification of a single representative year of Lake D, Lake E, and Lake F in the embodiment of the present invention and the comparison with the actual image of Planet with high spatial resolution;

[0025] Figure 7 1 is a comparison between the lake edge change detection result of Lake G in the embodiment of the present invention and the lake edge change detection result using the Landsat dataset and the MODIS dataset alone. DETAILED DESCRIPTION

[0026] In order to make the purpose and technical solution of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0027] Taking a lake in a certain area with an area greater than 1 square meter as of 2020 as an example, the method of the present invention is used to detect the edge changes of the lake from 2005 to 2020. In this embodiment, the process of the lake edge change detection method is as follows: Figure 1 As shown, the following steps are included:

[0028] S1. Obtain Landsat time series data and MODIS time series data for the lake area from 2005 to 2020.

[0029] S2. The GF-SG method is used to filter and fuse the Landsat time series data and MODIS time series data of each year to obtain high-quality and high-resolution ND VI time series with a spatial resolution of 30m and a time interval of 8 days.

[0030] S3. Intercept the NDVI time series in the lake melting period in each year's NDVI time series; in the intercepted NDVI time series, each pixel corresponds to a pixel time series, and the median of the pixel time series is taken as the data of the pixel for the year, and the data of all pixels for the year are combined to obtain a fused image of the year. The data of previous years are processed in the same way to obtain the fused images of previous years, forming an inter-annual NDVI time series.

[0031] S4. Based on the lake buffer area screening principle, the interannual NDVI time series corresponding to all pixels are screened to obtain pixels that meet the screening conditions. All screened pixels constitute the lake buffer mask; the lake buffer mask is spatially filtered to obtain the area where the lake boundary may change; the interannual NDVI time series corresponding to each pixel in the area where the lake boundary may change is obtained; the interannual NDVI time series corresponding to all pixels are used for lake edge change detection using the LandTrendr detection model to identify the changes in the lake edge.

[0032] In step S1, the lake area refers to lakes with an area greater than 1 square kilometer as of 2020, with a total of more than 1,300. It is necessary to obtain Landsat time series data and MODIS time series data from 2005 to 2020 within a 1,000m circumscribed rectangular area around each lake. MODIS has been providing data steadily since 2000, and Landsat has provided data from different satellites from 2005 to 2020. According to the year, Landsat 5 data was used from 2005 to 2013, and Landsat 7 or Landsat 8 data was selected from 2013 to 2020.

[0033] In step S2, MODIS data was sampled to the same 30m spatial resolution as Landsat using the bicubic interpolation algorithm, and the Landsat data and the sampled MODIS data were fused and filtered using GF-SG. In the process of filtering and fusion using the GF-SG method, the size of the local spatial adjacent window was set to 40 pixels × 40 pixels, and the correlation coefficient was set to 0.8; because the research object is water bodies, which appear as negative values ​​on NDVI, no threshold is set.

[0034] In step S3, the lake melt period is determined to be from July to September. The median NDVI value of each pixel in this time interval is synthesized to characterize the lake conditions of that year. Each pixel has a median lake melt period from 2005 to 2020 to characterize the characteristics of that year, forming an interannual time series from 2005 to 2020.

[0035] In step S4, the principles for selecting the lake buffer area are as follows: (1) the minimum value of the interannual NDVI time series corresponding to the pixel is less than 0; in this case, the pixel is considered to be a water body; (2) the average value of the three years after the maximum mutation point of the interannual NDVI time series corresponding to the pixel is less than 0, and the average value of the three years before the maximum mutation point is greater than 0; in this case, the pixel is considered to have changed.

[0036] After the pixels are screened, a lake buffer mask is formed. The lake buffer mask is spatially median filtered, and the filter window size of the median filter is set to 5 pixels × 5 pixels to obtain the area where the lake boundary may change. The interannual NDVI time series corresponding to all pixels in this area are used for lake edge change detection using the LandTrendr detection model. The detection years are set as input: 2005 to 2020. The parameters of the LandTrendr detection model are set as follows: maximum number of segments is 6; peak threshold is 0.9; number of vertex overshoots is 3; prevent one-year recovery is false; recovery threshold is 1; p value is 0.01; best model ratio is 1; minimum number of observations is 6; change magnitude is 0.1.

[0037] like Figure 2 As shown, a lake edge change detection method of the present invention was used to test Lake A, Lake B, and Lake C of different sizes and orientations, showing the changes in the overall picture of the lake and the local edge of the lake from 2005 to 2020. The edge changes are consistent with the lakeshore, and the years of lake expansion are clearly shown.

[0038] The following three experiments are provided to evaluate the effectiveness of a lake edge change detection method of the present invention.

[0039] The Planet 3m high spatial resolution remote sensing image dataset was used as a validation dataset. Because Planet only had data after 2015, the validation period was selected from 2016 to 2018. The Planet clear image data in August or September was selected to extract the range of changes by visual interpretation as the true value. The quantitative evaluation was measured using two indicators: precision and recall, both of which ranged from 0 to 1. Precision means: the proportion of samples that are truly positive among all samples predicted by the model to be positive. A high precision means that most of the samples predicted by the model to be positive are indeed positive, that is, the model's prediction results are relatively accurate. Recall means: the proportion of samples that are correctly predicted by the model to be positive among all samples that are truly positive. A high recall means that the model can find most of the samples that are truly positive, that is, the model's recall ability is strong.

[0040] The design and implementation results of each experiment are as follows:

[0041] Experiment 1: Combined with Planet high-resolution data to qualitatively evaluate the changes in lake edges; in order to verify the reliability of the detection results, we selected Planet's high-temporal and spatial resolution remote sensing image dataset to verify it. At the same time, in order to reduce the interference of lake water bodies with seasonal and other environmental factors, we selected Planet's clear image data in August for verification. Since the Planet image collection began to collect data in 2015, and some areas only have historical images after 2016, Figure 3 Only local changes from 2016 to 2020 are shown. Figure 3 The images with time stamps are lake images from Planet 3m high-resolution images. Figure 3 (a) The changes detected in the high-resolution Planet images of different years are integrated, and a unified reference standard is used as the baseline (border line). The dashed lines of different colors represent the parallel lines from the lakeshore to the baseline in different years, and the color bands of different colors represent the range of lake changes in that year. Compared with the high-resolution Planet images, the change trend in the detection results is highly consistent with it, and both show that the expansion trend of the lake was the most significant between 2017 and 2018.

[0042] Experiment 2: Combined with Planet high-resolution data to quantitatively evaluate the changes in lake edges; In order to quantitatively evaluate the results of lake change detection, we randomly selected 100 validation samples from the lake area. The samples were obtained from the high-resolution image set Planet, and the image range was cloud-free and clear from August to September 2016-2020. These samples use the lake shoreline in August or September each year as a reference, and then use ArcGIS spatial analysis tools to obtain the range of lake changes each year. The sample accuracy results of this sample are as follows: Figure 4 and Figure 5 We used the proposed method to detect 100 lake sample objects and compared the accuracy with the visual results of Pla net. The results showed that the precision was 0.854 and the recall was 0.723. The local accuracy evaluation results of Lake D, Lake E, and Lake F are also shown in Figure 6 , Figure 6 The first column is the result of lake change detection in this year, the second column is the true value of this year by visual interpretation of Planet images, and the third and fourth columns are remote sensing images of Planet at the corresponding time. The white dotted line and the black dotted line in the figure represent the reference lines before and after the change, respectively. It can be seen that the spatial structure of the detection results and the true value maintains a high consistency, and the accuracy exceeds 0.9.

[0043] Experiment three: verify the superiority of the method of the present invention compared with other data sets. The characteristics of water body changes in time series are very obvious, and data sets with the same resolution do not work well in foggy areas, and image quality is particularly severely impaired in spring and summer. The traditional approach generally uses neighboring images that are not disturbed by clouds and fog to fill in the missing images through simple interpolation and filtering. The accuracy of the image is affected by many factors, the stability is poor, and for land features with high timeliness (such as water bodies), large errors will be generated. The method of the present invention combines the advantages of Landsat and MODIS data, which is more advantageous than using the same data set alone. Figure 7 The local change detection of Lake G is shown. The data of pollution-free Lands at in 2019 in this area is seriously missing, and there is an obvious misdetection of the change year. Figure 7 (a)-(c) are Planet 3m spatial resolution images, showing the local change trend of the lake from 2017 to 2019; Figure 7 (d)-(f) are respectively the lake change detection results of the detection method of the present invention, the detection results using only Landsat data, and the detection results using only MO DIS data. Figure 7Compared with (d) and (e) and (f), the spatial details of lake changes are significantly increased, and the lakeshore is clearly layered, which is basically consistent with the lake outline under the Planet high-resolution image, indicating that high spatial resolution data has significant advantages in detecting lake changes. Figure 7 (a) and Figure 7 In the black rectangular area in (b), it can be seen that there are obvious boundary changes in the box, indicating that the lake changes in this area mainly occurred in 2018. Figure 7 (d) shows a high-resolution image Figure 7 (a) to Figure 7 (b) changes (2017-2018), and Figure 7 (e) The time of lake change in the detected area is 2019, which is inconsistent with the change in the high-resolution image, indicating that the detection method of the present invention can obtain more accurate results after fusing the two types of data.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A lake edge change detection method for identifying and extracting lake edge changes, characterized in that: The following steps are involved: S1. Obtain the Landsat time series data and MODIS time series data of the specified lake area over the years; S2, using the GF-SG method to filter and fuse the two data to obtain the NDVI time series of previous years; S3. Intercept the NDVI time series in each year's NDVI time series that is within the lake melting period; in the intercepted NDVI time series, each pixel corresponds to a pixel time series, and the median of the pixel time series is taken as the data of the pixel for that year, and the data of all pixels for that year are combined to obtain a fused image of that year. Perform the same processing on the data of previous years to obtain the fused images of previous years, forming an inter-annual NDVI time series; S4. Based on the lake buffer area screening principle, the interannual NDVI time series corresponding to all pixels are screened to obtain pixels that meet the screening conditions. All screened pixels constitute the lake buffer mask; the lake buffer mask is spatially filtered to obtain the area where the lake boundary may change; the interannual NDVI time series corresponding to each pixel in the area where the lake boundary may change is obtained; the interannual NDVI time series corresponding to all pixels are used for lake edge change detection using the LandTrendr detection model to identify the changes in the lake edge.

2. A lake edge change detection method as claimed in claim 1, characterized in that: In step S1, the Landsat time series data and MODIS time series data of the same year are selected.

3. A lake edge change detection method as claimed in claim 2, characterized in that: In step S2, first, the bicubic interpolation algorithm is used to sample MODIS data so that the sampled data and the Lands at data have the same spatial resolution; then, during the filtering and fusion process using the GF-SG method, the size of the local spatial adjacent window is set to 40 pixels × 40 pixels and the correlation coefficient is set to 0.

8.

4. A lake edge change detection method as claimed in claim 3, characterized in that: In step S3, the lake melting period is set to be the same every year.

5. A lake edge change detection method as claimed in claim 4, characterized in that: In step S4, the principles for selecting the lake buffer area are as follows: (1) the minimum value of the interannual NDVI time series corresponding to the pixel is less than 0; in this case, the pixel is considered to be a water body; (2) the average value of the three years after the maximum mutation point of the interannual NDVI time series corresponding to the pixel is less than 0, and the average value of the three years before the maximum mutation point is greater than 0; in this case, the pixel is considered to have changed.

6. A lake edge change detection method as claimed in claim 5, characterized in that: In step S4, the lake buffer mask is spatially filtered using median filtering, and the filter window size of the median filter is set to 5 pixels × 5 pixels to obtain areas where the lake boundary may change.

7. A lake edge change detection method as claimed in claim 6, characterized in that: In step S4, the parameters of the LandTrendr detection model are set as follows: maximum number of segments is 6; spike threshold is 0.9; number of vertex overshoots is 3; prevent one-year recovery is false; recovery threshold is 1; p-value is 0.01; best model proportion is 1; minimum number of observations is 6; and magnitude of change is 0.1.

Images