Coastal wetland extraction method based on remote sensing data and phenological characteristics

Abstract

The invention relates to the technical field of wetland classification, and particularly discloses a coastal wetland extraction method based on remote sensing data and phenological characteristics, and the method comprises the following steps: firstly, processing multiple years of multi-temporal remote sensing images based on a cloud computing platform; secondly, determining a coastal wetland potential distribution area; calculating the annual occurrence frequency of each pixel water body and vegetation in the research area, and extracting phenological characteristic data; further constructing and using a decision tree classification model to realize accurate identification and classification of the coastal wetland; and finally, verifying the precision of the wetland classification result. The method solves the problem that a traditional remote sensing wetland extraction method is large in space-time uncertainty, improves the space-time stability and accuracy of coastal wetland monitoring, and is suitable for long-term dynamic wetland monitoring and protection management in a large-scale area.

Description

Technical Field

[0001] This invention belongs to the field of wetland classification technology, specifically relating to a method for extracting coastal wetlands based on remote sensing data and phenological characteristics. Background Technology

[0002] Coastal wetlands, as an important component of the ecosystem, provide ecological services such as water conservation, biodiversity protection, flood control and regulation, and climate regulation. In recent years, affected by factors such as climate change and human activities, the area of ​​coastal wetlands has sharply decreased, and their ecological functions have severely degraded. Therefore, accurate identification and monitoring of coastal wetlands are of significant practical importance for ecological protection and restoration, watershed ecological management, and sustainable development.

[0003] Remote sensing technology, with its advantages of wide monitoring range, large information volume, and rapid update speed, has become an important means of wetland monitoring. However, traditional remote sensing extraction methods are usually based on single-temporal or a few composite images, neglecting factors such as the seasonality of water levels and vegetation phenological changes in coastal wetlands, resulting in significant uncertainty in wetland identification and unsatisfactory monitoring accuracy. With the rise of remote sensing big data platforms, such as Google Earth Engine and GEE, accurate extraction of coastal wetlands through remote sensing big data has become possible. Summary of the Invention

[0004] To address the shortcomings and problems of traditional remote sensing extraction methods in accurately identifying wetlands, which neglect factors such as seasonality of water levels and phenological changes in vegetation, this invention provides a method for extracting coastal wetlands based on remote sensing data and phenological characteristics.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for extracting coastal wetlands based on remote sensing data and phenological characteristics includes the following steps:

[0007] (1) Remote sensing data acquisition and preprocessing: Based on the remote sensing big data platform, Landsat series remote sensing images covering the study area for many years and multiple periods were acquired and the images were preprocessed.

[0008] (2) Water and vegetation identification based on preprocessed image data: Calculate the water body indices mNDWI and NDWI, and the vegetation indices NDVI, EVI, and LSWI, respectively, and then perform water and vegetation identification. The identification rules are as follows:

[0009] A mNDWI > 0.4 or NDWI > 0.2 indicates a body of water.

[0010] If EVI ≥ 0.1, NDVI ≥ 0.2, and LSWI > 0, the vegetation is identified.

[0011] (3) Determine the potential distribution area of ​​wetlands based on the identified water bodies: Calculate the slope using the digital elevation model (DEM) data of the study area to determine the potential distribution area of ​​coastal wetlands;

[0012] (4) Calculate phenological frequency characteristics: For the identified potential wetland area, calculate the annual water frequency WF and vegetation frequency VF for each pixel. For each pre-processed image, distinguish between water and vegetation, and calculate the water frequency Fsw and vegetation frequency Fgv for each pixel.

[0013] (5) Construct a decision tree classification model for coastal wetlands: Using the obtained water body frequency and vegetation frequency as features, a decision tree classification is adopted to analyze all effective observations of each pixel, calculate the probability of water and vegetation occurrence once every three years, and stably extract coastal wetland patches and classify their types.

[0014] The above-mentioned method for extracting coastal wetlands based on remote sensing data and phenological characteristics includes cloud masking, radiometric correction and geometric correction in step (1).

[0015] The aforementioned coastal wetland extraction method based on remote sensing data and phenological characteristics uses the QA band in the Landsat dataset for cloud removal during cloud masking. The cloud, cloud confidence, and cirrus cloud confidence information contained in this band are used as the basis for identifying cloud pixels. After removing the mask of pixels identified as clouds or cloud shadows, a remote sensing image without cloud interference is obtained.

[0016] The above-mentioned method for extracting coastal wetlands based on remote sensing data and phenological characteristics, in step (2), the calculation methods for water body indices mNDWI, NDWI, and vegetation indices NDVI, EVI, and LSWI are as follows:

[0017]

[0018] In the formula, B green The green band for Landsat satellites is band 2 in the Landsat 5 dataset and band 3 in the Landsat 8 dataset; B SWIR-1 This refers to the infrared bands of the Landsat satellite; it is band 5 in the Landsat 5 dataset and band 6 in the Landsat 8 dataset. Nir This is the near-infrared band of the Landsat satellite; it is band 4 in the Landsat 5 dataset and band 5 in the Landsat 8 dataset.

[0019]

[0020] In the formula: B blueThe blue band represents the Landsat satellite; it is band 1 in the Landsat 5 dataset and band 2 in the Landsat 8 dataset. (B) red This is the red band of the Landsat satellite, band 3 in the Landsat 5 dataset, and band 4 in the Landsat 8 dataset; B SWIR-1 This refers to the infrared bands of the Landsat satellite; it is band 5 in the Landsat 5 dataset and band 6 in the Landsat 8 dataset. Nir This is the near-infrared band of the Landsat satellite, which is band 4 in the Landsat 5 dataset and band 5 in the Landsat 8 dataset.

[0021] In the above-mentioned coastal wetland extraction method based on remote sensing data and phenological characteristics, the area with a DEM less than 5 and a slope threshold of 0-5° in step (3) is a potential coastal wetland area.

[0022] The above-mentioned method for extracting coastal wetlands based on remote sensing data and phenological characteristics uses the following formula for calculating the water body frequency Fsw of each pixel in step (4):

[0023]

[0024] Where: N water N represents the total number of times a pixel is identified as water. obs The total number of valid observations within a pixel over three years;

[0025] The formula for calculating the vegetation frequency Fgv for each pixel is:

[0026]

[0027] Where: N vegetation This represents the total number of times a pixel was identified as vegetation.

[0028] The above-mentioned method for extracting coastal wetlands based on remote sensing data and phenological characteristics, in step (5), the area where Fsw=0 is divided into inland areas, the area where 0<Fsw≤0.9 is divided into coastal wetlands, and the area where Fsw>0.9 is divided into open water areas;

[0029] Areas with Fgv ≤ 0.2 are classified as mudflats; areas with Fgv > 0.2 are classified as salt marshes.

[0030] The above-mentioned method for extracting coastal wetlands based on remote sensing data and phenological characteristics, step (5) further includes performing mode filtering on the extracted coastal wetlands to eliminate fragmented patches that are misidentified as coastal wetlands and improve the accuracy of wetland extraction, including the following:

[0031] a. Input a classified coastal wetland raster image and define the filter window size according to the size of the fragmentation.

[0032] b. Traverse the image, extract the pixel value within a specified window range around each pixel in the image, count the occurrence frequency of each pixel value, take the pixel value with the most occurrences as the new value, replace the original pixel value with the new value, and output the filtered image.

[0033] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention uses a buffer zone outside the coastline as a potential area, and utilizes long-term, multi-source remote sensing image datasets to comprehensively identify the extent and type of coastal wetlands by integrating the phenological frequency characteristics of water bodies and vegetation. It overcomes the spatiotemporal uncertainties inherent in single-image monitoring, significantly improving the accuracy and stability of large-scale wetland identification. This method significantly improves the spatiotemporal stability and monitoring accuracy of coastal wetland extraction, effectively overcoming the uncertainties of single-temporal remote sensing images during seasonal water level fluctuations and vegetation phenological changes. It is suitable for large-scale, long-term dynamic monitoring of coastal wetlands, providing scientific support for wetland protection and restoration. Attached Figure Description

[0034] Figure 1 This is a flowchart of the coastal wetland extraction method of the present invention.

[0035] Figure 2 This is a technical roadmap for the coastal wetland extraction method of the present invention.

[0036] Figure 3 This is a flowchart of the Landsat series remote sensing image preprocessing process of the present invention. Detailed Implementation

[0037] To address the shortcomings of existing technologies, this invention proposes a coastal wetland extraction method based on the concept of remote sensing big data. Utilizing long-term, multi-source remote sensing image datasets, and integrating the phenological frequency characteristics of water bodies and vegetation, this method accurately identifies the extent and type of coastal wetlands. It overcomes the spatiotemporal uncertainties inherent in single-image monitoring, significantly improving the accuracy and stability of large-scale wetland identification. The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0038] Example: This invention provides a method for extracting coastal wetlands based on remote sensing data and phenological characteristics. This example uses a typical watershed area as the study area and utilizes all Landsat satellite remote sensing images covering the study area from 1985 to 2020 on the Google Earth Engine cloud platform. First, cloud removal and radiometric correction are performed on the images to construct a year-round dataset. Next, potential coastal wetland areas are determined using DEM elevation and slope data. Then, the annual frequency of water bodies and vegetation occurrences in each pixel is calculated. Subsequently, a decision tree classification model is constructed, and wetland types are accurately extracted and classified according to phenological characteristic frequency rules. Finally, combined with high-precision validation data, the overall classification accuracy reaches over 90%, proving the method's stability and reliability. Figure 1 and Figure 2 As shown, this method mainly includes the following:

[0039] (1) Remote sensing data acquisition and preprocessing: Based on the remote sensing big data platform, Landsat series remote sensing images covering the study area over many years and in multiple periods were acquired and preprocessed. For example Figure 3 As shown, the preprocessing includes cloud masking, radiometric correction, and geometric correction. The cloud removal operation in the cloud masking uses the QA band in the Landsat dataset. The cloud, cloud confidence, and cirrus confidence information contained in this band are used as the basis for identifying cloud pixels. After removing the mask of pixels that are identified as clouds or cloud shadows, a remote sensing image without cloud interference is obtained.

[0040] (2) Water body and vegetation identification based on preprocessed image data: Water body indices mNDWI and NDWI and vegetation indices NDVI, EVI, and LSWI are calculated respectively.

[0041]

[0042] In the formula, B green The green band for Landsat satellites is band 2 in the Landsat 5 dataset and band 3 in the Landsat 8 dataset; B SWIR-1 This refers to the infrared bands of the Landsat satellite; it is band 5 in the Landsat 5 dataset and band 6 in the Landsat 8 dataset. Nir This is the near-infrared band of the Landsat satellite, which is band 4 in the Landsat 5 dataset and band 5 in the Landsat 8 dataset.

[0043]

[0044] In the formula: B blueThe blue band represents the Landsat satellite; it is band 1 in the Landsat 5 dataset and band 2 in the Landsat 8 dataset. (B) red This is the red band of the Landsat satellite, band 3 in the Landsat 5 dataset, and band 4 in the Landsat 8 dataset; B SWIR-1 This refers to the infrared bands of the Landsat satellite; it is band 5 in the Landsat 5 dataset and band 6 in the Landsat 8 dataset. Nir This is the near-infrared band of the Landsat satellite, which is band 4 in the Landsat 5 dataset and band 5 in the Landsat 8 dataset.

[0045] Then, based on the calculated water body indices mNDWI and NDWI, and vegetation indices NDVI, EVI, and LSWI, water body and vegetation identification are performed. The identification rules are as follows:

[0046] A value of mNDWI > 0.4 or NDWI > 0.2 indicates a body of water; a value of EVI ≥ 0.1, NDVI ≥ 0.2, and LSWI > 0 indicates a body of vegetation.

[0047] (3) Determine the potential distribution area of ​​wetlands based on the identified water bodies: Calculate the slope using the digital elevation model (DEM) data of the study area, and identify the areas with a DEM value less than 5 and a slope threshold of 0-5° as the potential distribution area of ​​coastal wetlands.

[0048] (4) Calculate phenological frequency characteristics: For the identified potential wetland area, calculate the annual water frequency WF and vegetation frequency VF for each pixel. For each pre-processed image, distinguish between water and vegetation, and calculate the water frequency Fsw and vegetation frequency Fgv for each pixel.

[0049]

[0050] Where: N water N represents the total number of times a pixel is identified as water. obs The total number of valid observations within a pixel over three years;

[0051] The formula for calculating the vegetation frequency Fgv for each pixel is:

[0052]

[0053] Where: N vegetation This represents the total number of times a pixel was identified as vegetation.

[0054] (5) Constructing a decision tree classification model for coastal wetlands: Using the obtained water body frequency and vegetation frequency as features, a decision tree classification is adopted to analyze all effective observations of each pixel, calculate the probability of water and vegetation occurrence once every three years, stably extract coastal wetland patches and classify their types; specifically, the area with Fsw=0 is classified as inland area, the area with 0<Fsw≤0.9 is classified as coastal wetland, the area with Fsw>0.9 is classified as open water, the area with Fgv≤0.2 is classified as mudflat, and the area with Fgv>0.2 is classified as salt marsh.

[0055] To eliminate fragmented patches that are misidentified as coastal wetlands and improve the accuracy of wetland extraction, a mode filtering process is applied to the extracted coastal wetlands. This process replaces fragmented, small pixels in the raster with the mode of the data values ​​of adjacent pixels. The steps include: a) Inputting a classified coastal wetland raster image and defining the filtering window size based on the size of the fragmented patches; b) Traversing the image, extracting the pixel values ​​within a specified window range around each pixel, counting the frequency of each pixel value, using the most frequent pixel value as the new value, replacing the original pixel value, and outputting the filtered image.

[0056] To validate and calibrate the wetland classification results, a confusion matrix method was used to evaluate the accuracy of the wetland classification results. High-resolution remote sensing imagery combined with field data was used as samples to validate the classification results, and the overall classification accuracy was calculated to evaluate the accuracy of the wetland classification results.

[0057]

[0058] In the formula, x ii N represents the number of correctly classified samples on the diagonal of the confusion matrix, and N is the total number of validation samples.

[0059] This embodiment validates the method using 972 sampling points. Finally, combined with the high-precision validation data of this typical watershed area, as shown in Table 1 below, the overall classification accuracy reaches over 90%, proving that the method is stable and reliable.

[0060] Table 1. Mixed-dish matrix and classification accuracy.

[0061]

[0062] This method utilizes long-term, multi-source remote sensing image datasets, integrating the phenological frequency characteristics of water bodies and vegetation, to accurately identify the extent and type of coastal wetlands. It overcomes the spatiotemporal uncertainties inherent in single-image monitoring, significantly improving the accuracy and stability of large-scale wetland identification. It is particularly suitable for the accurate identification and classification of coastal wetland extents over large-scale areas. This method significantly enhances the spatiotemporal stability and monitoring accuracy of coastal wetland extraction, effectively overcoming the uncertainties of single-temporal remote sensing images during seasonal water level fluctuations and vegetation phenological changes. It is suitable for dynamic monitoring of long-term coastal wetland series over large-scale areas, providing scientific support for wetland protection and restoration.

[0063] The above description is only a preferred embodiment of the present invention and does not limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting coastal wetlands based on remote sensing data and phenological features, characterized in that: The method comprises the following steps: (1) remote sensing data acquisition and preprocessing: based on a remote sensing big data platform, acquiring Landsat series remote sensing images covering the study area for multiple years and multiple periods, and preprocessing the images; (2) water body and vegetation identification based on the preprocessed image data: calculating water body indexes mNDWI, NDWI and vegetation indexes NDVI, EVI and LSWI, and then performing water body identification and vegetation identification, the identification rules being: mNDWI > 0.4 or NDWI > 0.2 is identified as a water body; EVI > 0.1 and NDVI > 0.2 and LSWI > 0 are identified as vegetation; (3) determining a potential distribution area of a wetland based on the identified water body: calculating the slope by using digital elevation model (DEM) data of the study area to determine a potential distribution area of a coastal wetland; (4) calculating a phenology frequency feature: calculating the water body frequency WF and the vegetation frequency VF of each pixel in the determined potential wetland area, identifying water and vegetation for each preprocessed image, and calculating the water body frequency Fsw and the vegetation frequency Fgv of each pixel; (5) constructing a decision tree classification model of the coastal wetland: taking the obtained water body frequency and vegetation frequency as features, using decision tree classification to analyze all effective observation values of each pixel, calculating the water and vegetation appearance probability every three years, and stably extracting the coastal wetland map patches and dividing the types thereof.

2. The method for coastal wetland extraction based on remote sensing data and phenological features according to claim 1, characterized in that: The preprocessing in step (1) comprises cloud masking, radiation correction and geometric correction.

3. The method of claim 2, wherein the method comprises: In the cloud masking, the cloud removal operation is performed by using the QA band in the Landsat data set, and the cloud, cloud confidence and cirrus confidence information contained in the band are used as the basis for identifying cloud pixels. After the cloud or cloud shadow pixels identified are removed, the remote sensing image without cloud layer interference is obtained.

4. The method of claim 1, wherein the method comprises: In step (2), the calculation methods of the water body indexes mNDWI, NDWI and the vegetation indexes NDVI, EVI and LSWI are as follows: wherein B green is the green band of the Landsat satellite, band 2 in Landsat 5 data and band 3 in Landsat 8 data; B SWIR-1 is the infrared band of the Landsat satellite, band 5 in Landsat 5 data and band 6 in Landsat 8 data; B Nir is the near infrared band of the Landsat satellite, band 4 in Landsat 5 data and band 5 in Landsat 8 data; where: B blue is the blue band of the Landsat satellite, band 1 in Landsat 5 data and band 2 in Landsat 8 data; B red is the red band of the Landsat satellite, band 3 in Landsat 5 data and band 4 in Landsat 8 data; B SWIR-1 is the infrared band of the Landsat satellite, band 5 in Landsat 5 data and band 6 in Landsat 8 data; B Nir is the near-infrared band of the Landsat satellite, band 4 in Landsat 5 data and band 5 in Landsat 8 data.

5. The method of claim 1, wherein the method is characterized by: In step (3), the area with DEM less than 5 and a slope threshold of 0-5° is the potential distribution area of the coastal wetland.

6. The method of claim 1, wherein the method further comprises: In step (4), the calculation formula of the water body frequency Fsw of each pixel is as follows: where: N water is the total number of times a pixel is identified as water, N obs is the total number of times a pixel has a valid observation within three years; The calculation formula of the vegetation frequency Fgv of each pixel is as follows: In the formula: N vegetation is the total number of times the pixel is identified as vegetation.

7. The method of claim 6, wherein the method further comprises: In step (5), the area with Fsw = 0 is divided into an inland area, the area with 0 < Fsw < 0.9 is divided into a coastal wetland, and the area with Fsw > 0.9 is divided into an open water area; The area with Fgv < 0.2 is divided into a mudflat, and the area with Fgv > 0.2 is divided into a salt marsh.

8. The method of claim 7, wherein the method further comprises: Step (5) further comprises performing mode filtering processing on the extracted coastal wetland to eliminate the scattered patches misidentified as the coastal wetland and improve the accuracy of wetland extraction, including the following contents: a. inputting the classified completed coastal wetland raster image, defining the filter window size according to the size of the scattered patches; b. traversing the image, extracting the pixel values in the specified window range around each pixel in the image, counting the number of occurrences of each pixel value, taking the pixel value with the most occurrences as the new value, replacing the original pixel value with the new value, and outputting the filtered image.

Images