Mangrove plant kandelia candel extraction method based on Sentinel-1 / 2 time sequence remote sensing image

Abstract

The invention discloses a mangrove plant kandelia candel extraction method based on Sentinel-1 / 2 time sequence remote sensing images, and relates to satellite remote sensing image processing. The method comprises the following steps: generating year-by-year quantile data of an original spectral band and a spectral index by utilizing a Sentinel-1 / 2 time sequence image, and generating 10m elevation, gradient and slope direction data by adopting 30m global elevation data; establishing a research area classification system, and generating a training sample set according to a high-resolution image provided by Google Earth software; and generating a training model based on a random forest classification algorithm to obtain a mangrove plant kandelia candel distribution result. The mangrove plant kandelia candel extraction method based on percentile is designed by using Sentinel-1 / 2 time sequence remote sensing data, and high-precision extraction of the mangrove plant kandelia candel with plaque fragmentation and large scale is realized. The result shows that the spatial distribution condition of the kandelia candel community provides data support for mangrove forest ecological restoration, blue carbon evaluation, ecological system service evaluation and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of satellite remote sensing image processing, and particularly relates to a method for extracting the mangrove plant Kandelia obovata based on Sentinel-1 / 2 time series remote sensing images. Background Art

[0002] The mangrove ecological restoration project plays an important role in mitigating global climate change. Due to its high survival rate and cold tolerance, Kandelia obovata has become an important mangrove plant species in the mangrove ecological restoration project along the southeast coast of China and is widely distributed in various regions of mangroves in China. Rapid and effective extraction of Kandelia obovata community information on a large scale (national scale) is of great significance for the restoration, management, and evaluation of the mangrove ecosystem. Remote sensing, due to its ability to monitor surface information over a large area and dynamically at high speed, has become the main means to obtain the above information. Due to the lack of a large-scale field sampling sample library of mangrove species, the scattered distribution of Kandelia obovata among other mangrove plant species, the high spectral and texture similarity among mangrove plant species, and the high extraction difficulty, current research on Kandelia obovata extraction mainly focuses on the regional scale, and there are few research results on a large scale and with high precision. Summary of the Invention

[0003] The purpose of the present invention is to overcome the deficiencies of existing technologies and methods, and provide a method for extracting the mangrove plant Kandelia obovata based on Sentinel-1 / 2 time series remote sensing images, which fully considers the spatio-temporal information required for the extraction of the mangrove plant Kandelia obovata and has the ability to extract the mangrove plant Kandelia obovata at the national scale with high precision.

[0004] Using the original spectral bands and annual percentile data of spectral indices of Sentinel-1 / 2 time series images, and generating elevation, slope, and aspect data by using 30-meter global elevation data; generating a training sample set based on Google Earth high-resolution images; obtaining the distribution result of the mangrove plant Kandelia obovata based on the training model generated by the random forest classification algorithm. On the basis of using Sentinel-1 / 2 time series remote sensing data with high spatio-temporal resolution, a method for extracting the mangrove plant Kandelia obovata based on percentiles is designed to achieve high-precision extraction of the mangrove plant Kandelia obovata with patch fragmentation and large scale; the research results reveal the spatial distribution of the Kandelia obovata community in China and provide data support for mangrove ecological restoration, blue carbon assessment, ecosystem service assessment, and sustainable management.

[0005] A method for extracting the mangrove plant Kandelia obovata based on Sentinel-1 / 2 time series remote sensing images includes the following steps:

[0006] 1) Data acquisition: Based on the Google Earth Engine platform (https: / / code.earthengine.google.com / ), all Sentinel-1 SAR GRD data, Sentinel-2 Level-1C data along the coast of the whole country, and 30-meter elevation data obtained by space shuttle radar mapping are used to obtain the mangrove boundary data of China in that year;

[0007] 2) Data preprocessing: Based on the Google Earth Engine platform, Sentinel-1 SAR GRD, Sentinel-2 Level-1C, and elevation data are clipped and cloud-removed for preprocessing;

[0008] 3) Feature selection: Based on the Google Earth Engine platform, ① slope and aspect data are calculated using 30-meter elevation data; ② the VV and VH bands of Sentinel-1 SAR GRD and the original spectral bands (B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12 bands) of Sentinel-2 Level-1C are extracted; ③ spectral indices (Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Enhanced Vegetation Index (EVI), Green Normalized Difference Vegetation Index (GNDVI), Ratio Vegetation Index (RVI), and Atmospheric Resistant Vegetation Index (ARVI)) are calculated using Sentinel-2 Level-1C data; ④ texture information (Contrast (CON), Entropy (ENT), Correlation (COR), and Standard Deviation (STD)) is calculated using Sentinel-1 SAR GRD data;

[0009] 4) Quantile feature extraction based on time series: Based on the original band data, spectral index data, and texture information data obtained in step 3), on the Google Earth Engine platform, the pixels of each data are sorted in ascending order, and according to the sorting results, 90%, 75%, 50%, 25%, and 10% quantile features are extracted;

[0010] 5) Generate a classification sample set: Combining field research and Google Earth high-resolution remote sensing images, Kandelia obovata sample points and non-Kandelia obovata sample points are obtained, the longitude and latitude information of the sample points is recorded, and a classification sample set is established (there is no specific requirement for the number of sample points. In actual operation, the representativeness and diversity of the samples should be ensured as much as possible according to the size of the study area and the complexity of the actual ground objects);

[0011] 6) Random forest algorithm training: According to the training sample set and quantile features, a binary classification map is generated using the random forest classification algorithm;

[0012] 7) Conduct a preliminary experiment: Conduct a preliminary experiment to determine the number of eigenvalues used when importing the random forest algorithm.

[0013] 8) Obtain the results: After determining the number of eigenvalues, perform the random forest algorithm again to identify Kandelia obovata and export the results.

[0014] The method for extracting the mangrove plant Kandelia obovata based on Sentinel-1 / 2 time series remote sensing images provided by the present invention solves the problem that the existing research on extracting Kandelia obovata based on remote sensing images mainly focuses on small scales and low accuracies. As a dominant species among mangrove plants, extracting the spatial distribution area of Kandelia obovata nationwide can provide a scientific basis for the mangrove protection and restoration project in China. Obtaining the spatial distribution information of Kandelia obovata is also beneficial for estimating the biomass of mangroves, blue carbon assessment, and ecosystem value assessment, which has certain significance for the scientific management of mangrove wetlands. This method uses the unique spectral-temporal variability index of Kandelia obovata and other mangrove species as the identification feature for classification. Its main advantage is that it can be applied globally without modification for specific locations, and an effective spectral-temporal quantile feature for distinguishing Kandelia obovata in the mangrove plant community is explored, that is, the spectral-temporal variation characteristics of the blue light band and the red edge band can be used as an important feature combination for identifying Kandelia obovata, which has good practical value. Description of the Drawings

[0015] Figure 1 is the flow chart for extracting the mangrove plant Kandelia obovata in the embodiment of the present invention.

[0016] Figure 2 is the comparison chart of the classification accuracies for taking the top five, ten, fifteen, twenty, and fifty eigenvalues in the embodiment of the present invention.

[0017] Figure 3 is the spectral percentile chart of Kandelia obovata and other mangrove species in the embodiment of the present invention.

[0018] Figure 4 is the extraction result of the spatial distribution of the mangrove plant Kandelia obovata in China in 2018. Detailed Embodiment

[0019] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the following embodiments will further illustrate the present invention in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0020] Figure 1 The flow chart for extracting the mangrove plant Kandelia obovata in the embodiment of the present invention is given. The embodiment of the present invention specifically includes the following steps:

[0021] Step 1: Obtain the vector boundary of mangroves in China in 2018; here, the research results of Tao Zhang (2021) et al. are used. The data download address is: https: / / www.scidb.cn / en / detail?dataSetId=785913341146038272& version=V1. After downloading, the compressed file Mangrove China 2018.zip is obtained.

[0022] Step 2: Upload the mangrove boundary of China in 2018 to the Google Earth Engine platform: Open the Google Earth Engine platform (https: / / code.earthengine.google.com / ). In the platform interface, click the "Assets" button in the upper left corner, find the "new" option, and select "Shape file", indicating that the file to be uploaded is a vector data file; after clicking "SELECT", a file selection window will pop up. Locate the storage path of the Mangrove China 2018.zip downloaded in Step 1. In the "Asset Name" input box, enter the name of this data, which is set to Mangrove China2018 here for subsequent identification and management. The rest of the settings are default. Click "UPLOAD" to upload the data to the Google Earth Engine platform.

[0023] Step 3: Define the uploaded mangrove boundary of China in 2018 as the area of interest, and load and display it on the Google Earth Engine platform; in the Google Earth Engine platform, set the uploaded mangrove boundary data as the area of interest (AOI) to ensure that subsequent operations are only carried out on this specific area, improving the accuracy and efficiency of analysis; through the relevant tools and interfaces provided by the platform, load and display the area of interest on the platform interface for easy and intuitive viewing and confirmation of the accuracy and integrity of the data.

[0024] Step 4: Obtain the Sentinel-2 Level-1C images within the area of interest and perform preprocessing; using the relevant functions and methods provided by the Google Earth Engine platform, first use.filterDate to filter out the images with the date range between January 1, 2018 and December 31, 2018 to obtain the image data of the target year to meet the time requirements of the research; limit the image collection to a specific area (i.e., the area of interest) through.filterBounds(aoi); filter out the images with cloud cover exceeding 10% through the.filter function; apply the maskS2clouds function to each image in the image collection using.map(maskS2clouds) to mask out clouds and cirrus clouds.

[0025] Step 5: Obtain a digital elevation model with a resolution of 30 meters, and calculate slope, aspect, and elevation; load the SRTM (Shuttle Radar Topography Mission) digital elevation model dataset provided by the United States Geological Survey (USGS) (dataset name 'USGS / SRTMGL1_003'), which is high-precision terrain data obtained through radar technology; use the relevant algorithms and tools provided by the Google Earth Engine platform to calculate slope, aspect, and elevation based on this digital elevation model. These terrain parameters are used to understand the geographical environment for mangrove growth, and the calculated output variables are temporarily stored in the console.

[0026] Step 6: Calculate spectral indices: Define the add_RS_index function that accepts an image img as input and calculates multiple remote sensing indices; the multiple remote sensing indices include Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Enhanced Vegetation Index (EVI), Green Normalized Difference Vegetation Index (GNDVI), Ratio Vegetation Index (RVI), and Atmospheric Resistant Vegetation Index (ARVI); the formulas used for calculation are as follows:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] Among them, NDVI is the Normalized Difference Vegetation Index, NDWI is the Normalized Difference Water Index, EVI is the Enhanced Vegetation Index, GNDVI is the Green Normalized Difference Vegetation Index, RVI is the Ratio Vegetation Index, and ARVI is the Atmospheric Resistant Vegetation Index; BLUE is the blue light band, GREEN is the green light band, RED is the red light band, and NIR is the near-infrared band.

[0034] After calculating the spectral indices, to more intuitively understand the differences in spectral percentiles between Kandelia obovata and other mangrove species, Figure 3 a spectral percentile map of Kandelia obovata and other mangrove species in the embodiments of the present invention is given.

[0035] Step 7: Extract Sentinel-2 percentile features based on time series; ① Define the percentile_bands function, which uses the reduce() method. The principle is to perform statistical analysis on the data of each band in the image collection and calculate different percentiles (the 10th, 25th, 50th, 75th, and 90th percentiles) of each band in the image collection; ② Call percentile_bands to calculate the 10th, 25th, 50th, 75th, and 90th percentiles of the original spectral bands B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12 of Sentinel-2 Level-1C and each spectral index, and temporarily store the calculated output variables in the console.

[0036] Step 8: Obtain the images of Sentinel-1 SAR GRD within the region of interest, including VH and VV polarization data, and perform preprocessing; construct an image collection of Sentinel-1 SAR GRD, and filter the data according to time and the region of interest through the.filter function.

[0037] Step 9: Calculate texture features; use the glcmTexture({size: 3}) function to create a 3x3 window to calculate texture features; texture features are important parameters for describing the surface structure and features of ground objects, including contrast (CON), entropy (ENT), correlation (COR), and standard deviation (STD).

[0038] Step 10: Extract Sentinel-1 percentiles and texture features based on time series; call the percentile_bands function defined in Step 7 to calculate the 10th, 25th, 50th, 75th, and 90th percentiles of the VH and VV polarization data and texture features of Sentinel-2 Level-1C, and temporarily store the output variables in the console.

[0039] Step 11: Output the feature set; temporarily store the Sentinel-2 percentile features obtained in Step 7, the VH and VV polarization data obtained in Step 10, and the percentage features of the texture features in the feature set for subsequent analysis; this feature set contains rich image feature information about the mangrove area and is used for subsequent analysis and classification work.

[0040] Step 12: Generate a classification sample set and upload it to the Google Earth Engine platform: Combine on-site investigations and Google Earth high-resolution remote sensing images to obtain Kandelia obovata sample points and non-Kandelia obovata sample points, record the longitude and latitude information of the sample points, and establish a classification sample set; upload the classification sample set to the Google Earth Engine platform by the method in Step 2; there is no specific requirement for the number of sample points. In actual operation, according to the size of the study area and the complexity of the actual ground objects, try to ensure the representativeness and diversity of the samples; in this example, first establish 278 10m*10m quadrats in each mangrove experimental sample area across the country, and then investigate each mangrove species within the quadrat. If it is a pure Kandelia obovata or other pure species, record the longitude and latitude of the center point of the quadrat, and the recorded longitude and latitude are the positions of the sample points. Finally, 120 pure-species sample points are selected from the community survey sample data in 2018; in addition, the plots of pure mangrove species communities obtained by visual interpretation in the existing plots include: Fujian Ningde Fuding Mangrove Plot, Guangdong Zhongshan Tanzhou Mangrove Plot, Guangdong Shantou Mangrove Plot, Guangxi Beihai Mangrove Plot, Hainan Danzhou Mangrove Plot, Hainan Haikou Mangrove Plot. Finally, a total of 5382 sample points are selected for classification, with 1351 Kandelia obovata training samples and 4031 other mangrove species training sample points.

[0041] Step 13: Divide the sample set into a training set and a test set; use the split=0.7 code to define the ratio of the training set and the test set. 70% of the samples will be used to train the random forest classifier to enable it to learn the features and classification patterns of the samples; 30% of the samples are used for testing to evaluate the performance and accuracy of the classifier; in this embodiment, the random seed number is fixed at 20.

[0042] Step 14: Extract the features of the training set and the test set from the image data; use the sampleRegions function to extract the feature set of each training set and test set sample point; this function extracts the corresponding feature information from the previously processed image data according to the position of the sample point, providing specific input data for training and testing the classifier.

[0043] Step 15: Create a random forest classifier; use ee.Classifier.smileRandomForest to define the forest classifier parameters; in this embodiment, it is defined that: numberOfTrees: 20, bagFraction: 0.99, that is, the number of iterations of the random forest is selected as 20, and 99% of the samples are selected for training in each iteration.

[0044] Step 16: Calculate the importance of feature values; use the ee.Dictionary function to calculate the importance degree of the feature set; through this function, the importance of each feature in the classification process can be evaluated. Features with a higher importance degree have a greater impact on the classification result and should be given priority when selecting the feature set.

[0045] Step 17: Conduct a preliminary experiment to determine the importance degree of the feature set values used; call the random forest classifier created in Step 15 to perform random forest classification; call the random forest classifier created in Step 15, use the training set and test set in Step 13 to classify the feature set created in Step 14, and obtain the importance degree of the feature set in Step 16. Arrange the importance degrees from largest to smallest to intuitively understand the importance order of each feature.

[0046] Step 18: Determine the number of feature sets finally used for classification; sequentially select the top five, top ten, top fifteen, top twenty, and top fifty features obtained in Step 17, and call the random forest classifier created in Step 15 again to perform five random forest classifications. Figure 2 It is a comparison chart of the classification accuracies of taking the first five, ten, fifteen, twenty, and fifty eigenvalue classifications in the embodiment of the present invention. After each classification, use the test.errorMatrix function to calculate the confusion matrix, which is used to intuitively reflect the accuracy and error situation of the classification result; by comparing the Kappa coefficients of these five classification results, finally select the classification result with the highest classification accuracy as the final spatial distribution extraction result of Kandelia obovata in China in 2018 ( Figure 4 )

[0047] The present invention first fuses the Sentinel-2 multi-temporal percentile spectral features and the Sentinel-1 texture features, and constructs a unique feature set suitable for mangrove species classification through the feature optimization mechanism in Steps 17-18.

[0048] The present invention first generates annual quantile data of backscattering coefficients, original spectral bands, and spectral indices using Sentinel-1 / 2 time series images, and generates elevation, slope, and aspect data using 30-meter global elevation data. Then, a classification system for the study area is established, and a training sample set is generated based on high-resolution Google Earth images. Next, based on the training model generated by the random forest classification algorithm, the distribution result of the mangrove plant Kandelia obovata is obtained. Based on the Sentinel-1 / 2 time series remote sensing data with high spatio-temporal resolution, the present invention designs a method for extracting Kandelia obovata based on percentiles, achieving high-precision extraction of Kandelia obovata with patch fragmentation and large scale. The research results reveal the spatial distribution of the Kandelia obovata community in China, providing data support for mangrove ecological restoration, blue carbon assessment, ecosystem service assessment, and sustainable management.

[0049] The above embodiments are only preferred embodiments of the present invention and should not be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the application of the present invention should still fall within the scope covered by the patent of the present invention.

Claims

1. A method for extracting mangrove plants Kandelia candel based on Sentinel-1 / 2 time series remote sensing images, characterized in that: The steps include: Step 1: Obtain all Sentinel-1 SAR GRD data, Sentinel-2 Level-1C data, elevation data and research boundary data along the coast of China; Step 2: Preprocess the Sentinel-1 SAR GRD data, Sentinel-2 Level-1C data and elevation data; Step 3: Calculate spectral and texture features and extract percentile features based on time series; Step 4: Based on Google Earth high-resolution remote sensing images and field survey results, obtain the sample points and non-sample points of Kandelia candel in the study area, establish a classification sample set, and generate a training sample set; Step 5: Use the random forest algorithm to conduct a preliminary experiment to determine the number of feature sets that will be used for classification; Step 6: Based on the number of feature sets determined in step 5, combined with the training sample set in step 4 and the quantile features in step 3, a random forest classification algorithm is used to generate a binary classification map; Step 7: Identify Kandelia officinalis in the classification map and export the raster file.

2. The method for extracting mangrove plant Kandelia candel based on Sentinel-1 / 2 time series remote sensing images according to claim 1, characterized in that In step 2, the Sentinel-1 SAR GRD data, Sentinel-2 Level-1C data and elevation data are preprocessed, specifically including clipping, extracting required bands, removing unavailable pixels and resampling.

3. A method for extracting mangrove plant Kandelia candel based on Sentinel-1 / 2 time series remote sensing images according to claim 2, characterized in that The bands required for extraction include the VV and VH bands of Sentinel-1 SAR GRD, and the B2, B3, B4, B5, B6, B7, B8, B8A, B11 and B12 bands of Sentinel-2 Level-1C.

4. The method for extracting mangrove plant Kandelia candel based on Sentinel-1 / 2 time series remote sensing images according to claim 1, characterized in that In step 3, the spectrum is calculated, and specific spectrum indices include normalized vegetation index, normalized water index, enhanced vegetation index, green light normalized difference vegetation index, ratio vegetation index and anti-atmospheric vegetation index.

5. The method for extracting mangrove plant Kandelia candel based on Sentinel-1 / 2 time series remote sensing images according to claim 1, characterized in that In step 3, the texture features include contrast, entropy, correlation and standard deviation.

6. The method for extracting mangrove plant Kandelia candel based on Sentinel-1 / 2 time series remote sensing images according to claim 1, characterized in that In step 3, the calculation of spectrum and texture features, specifically generating and calculating spectrum index and texture index time series data, includes the following steps: (1) Calculate the spectral index based on the original spectral band data of Sentinel-2 Level-1C; (2) Calculate the texture index based on the backscatter coefficient band data of Sentinel-1 SAR GRD and the original spectral band and spectral index data of Sentinel-2 Level-1C; (3) The backscattering coefficient, original spectral band data, and the spectral index and texture index data generated above are superimposed together to form time series data.

7. The method for extracting mangrove plant Kandelia candel based on Sentinel-1 / 2 time series remote sensing images according to claim 1, characterized in that In step 3, the percentile feature extraction includes the following steps: (1) Sort each pixel in each band time series data in ascending order; (2) Based on the sorting results, extract the 90%, 75%, 50%, 25%, and 10% quantile features.

8. The method for extracting mangrove Kandelia candel based on Sentinel-1 / 2 time series remote sensing images according to claim 1, characterized in that In step 5, the preliminary experiment is performed, including the following steps: (1) Calculate the importance of the feature set used for random forest classification and sort it in ascending order; (2) Select the first 5, first 10, first 15, first 20, and first 50 features for random forest classification, determine the final number of feature sets used, and calculate the classification accuracy.

Images