Automatic extraction method of submerged macrophyte based on fusion of sentinel-1 sar and sentinel-2 msi
By fusing Sentinel-1 SAR and Sentinel-2 MSI images and employing principal component analysis and Gaussian mixture model clustering, the problem of submerged vegetation monitoring in traditional methods was solved, enabling rapid, real-time, large-scale, and high-precision automatic extraction of submerged vegetation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING INST OF GEOGRAPHY & LIMNOLOGY
- Filing Date
- 2024-04-12
- Publication Date
- 2026-07-21
AI Technical Summary
Traditional methods are difficult to use for rapid, real-time, and large-area monitoring of the distribution of submerged vegetation in shallow lakes, especially due to interference from algal blooms and floating-leaved vegetation. Furthermore, classification threshold methods that rely on human intervention are not applicable.
By fusing Sentinel-1 SAR and Sentinel-2 MSI images, submerged vegetation was automatically extracted through principal component analysis and Gaussian mixture model clustering. Submerged vegetation areas were obtained by using dual-polarization radar vegetation index and Gaussian mixture model clustering.
It enables rapid, real-time, large-scale, and high-precision automatic monitoring of submerged vegetation, reducing human intervention and improving monitoring efficiency and accuracy.
Smart Images
Figure CN118379610B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of remote sensing applications and relates to an automatic extraction method for submerged vegetation in lakes based on the fusion of Sentinel-1 SAR and Sentinel-2 MSI. Background Technology
[0002] Submerged vegetation is one of the main primary producers in lake ecosystems, playing a crucial role in lake ecological and environmental functions. It reduces nutrient concentrations and inhibits algal growth through oxygen production, nutrient retention, and denitrification, while also maintaining lake biodiversity as habitat and shelter for fish and zooplankton. These functions make submerged vegetation a key element in maintaining clear water homeostasis and preventing the transition to an algae-dominated turbid water homeostasis in lakes. Therefore, as an important regulator and indicator of lake ecosystems, real-time and rapid monitoring of submerged vegetation is of guiding significance for assessing lake ecosystem transformation and ecosystem service functions.
[0003] Traditional methods for monitoring submerged vegetation include manual quadrat and transect surveys, which are time-consuming, labor-intensive, have limited monitoring range and sample size, and are difficult to monitor (submerged vegetation is located underwater and is not easy to collect), making it impossible to quickly, in real time, and over a large area obtain the distribution status of submerged vegetation. Remote sensing has the advantage of rapid, real-time, and large-area monitoring of surface information, but in shallow lakes, especially eutrophic lakes, the difficulties and challenges of monitoring submerged vegetation are: (1) the extraction of submerged vegetation is affected by various other land features (such as algal blooms and floating-leaved vegetation); (2) traditional classification trees and thresholding methods require human intervention. Summary of the Invention
[0004] The purpose of this invention is to overcome the difficulties in remote sensing extraction of submerged vegetation in shallow lakes, especially eutrophic lakes, in the prior art, and to provide an automatic extraction method for submerged vegetation in lakes based on the fusion of Sentinel-1 SAR and Sentinel-2 MSI.
[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:
[0006] An automatic extraction method for submerged vegetation in lakes based on the fusion of Sentinel-1 SAR and Sentinel-2 MSI includes:
[0007] Principal component analysis was performed on the sensitive bands related to aquatic vegetation in Sentinel-2 MSI TOA images of the study area;
[0008] Acquire and process Sentinel-1 SAR GRD images with the closest time frame to the Sentinel-2 MSI TOA images, and calculate the dual-polarization radar vegetation index (RVI) from the Sentinel-1 SAR GRD images. dual Using RVI dual The first principal component in the principal component analysis is replaced, and the result of the replaced principal component analysis is inversely transformed to obtain the fused bands. Band 1 of the fused bands is used for submerged vegetation extraction.
[0009] Gaussian mixture model clustering was performed on the band 1 image. The corresponding clustering results were obtained based on the optimal number of clusters. The mean value of band 1 for each class was calculated. The class with the highest mean value is the submerged vegetation.
[0010] In a preferred embodiment, the sensitive bands related to aquatic vegetation are the B2, B3, B4, B8, B11 and B12 bands of the Sentinel-2 MSI TOA image.
[0011] As a preferred implementation, Sentinel-1 SAR GRD images within 10 days of the Sentinel-2 MSI TOA image time are obtained for calculation, and the VV and VH polarization bands of the Sentinel-1 SAR GRD image are added to the Sentinel-2 MSI TOA image.
[0012] As a preferred implementation, the optimal number of clusters is determined based on the following method:
[0013] For different numbers of clusters, calculate the corresponding Davidson-Bourdin index (DBI) and Bayesian Information Criterion (BIC), and then standardize them.
[0014] The weighted sum of the standardized DBI and standardized BIC values is denoted as the CI value.
[0015] Find the number of clusters that minimizes the CI value, i.e., the optimal number of clusters.
[0016] Preferably, the standardized DBI index and the standardized BIC index values have equal weights when weighted and summed.
[0017] Preferably, the optimal number of clusters is obtained within the range of 3-8.
[0018] As a preferred implementation, the Sentinel-1 SAR GRD and Sentinel-2 MSI TOA images are acquired from the GEE cloud platform.
[0019] As a preferred embodiment, the method also includes preprocessing the Sentinel-1 SAR GRD and Sentinel-2 MSI TOA images, including filtering the Sentinel-1 SAR GRD images, removing clouds from the Sentinel-2 MSI TOA images, and cropping the study area from the Sentinel-1 SAR GRD and Sentinel-2 MSI TOA images for subsequent calculations.
[0020] As a preferred implementation method, the Lee-Sigma method is used for filtering.
[0021] The cloud removal process is as follows: use the QA60 band and set the values of cloudBitMask and cirrusBitMask to 0 before completing the masking.
[0022] As a preferred implementation method, after the submerged vegetation has been extracted, the extraction accuracy is evaluated using producer accuracy, user accuracy, F-score, and overall accuracy.
[0023] This invention first fuses Sentinel-1 SAR GRD and Sentinel-2 MSI TOA images using the principal component replacement method to obtain SSAVI, and then uses a Gaussian mixture model for clustering to obtain submerged vegetation regions. This overcomes the problems of removing various interfering features and determining classification thresholds when extracting submerged vegetation. The method of this invention can automatically, in real time, over a large area, and with high accuracy identify, monitor, and map the spatial distribution of lake submerged vegetation, which has important practical significance for monitoring and tracing the spatiotemporal distribution of lake submerged vegetation and for lake ecological restoration and management. Attached Figure Description
[0024] Figure 1 This is a flowchart of the technical solution of the present invention.
[0025] Figure 2 The classification results for different polarization indices.
[0026] Figure 3 The CI curves corresponding to the clustering parameters in the embodiments of the method described in this invention are shown.
[0027] Figure 4 The above is the remote sensing extraction result of submerged vegetation in Taihu Lake on July 28, 2019, in an embodiment of the method described in this invention. Detailed Implementation
[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0029] An automatic extraction method for submerged vegetation in lakes based on Sentinel-1 SAR and Sentinel-2 MSI fusion is presented as a case study using Taihu Lake, a typical large shallow lake in China (average depth 1.9 meters). Taihu Lake's latitude and longitude ranges from 119.55–120.34E and 30.55–31.32N. Surveys and data show that large areas of submerged vegetation and emergent / floating-leaved plants are distributed in the eastern part of Taihu Lake, while large-scale algal blooms frequently occur in the central and northwestern parts. Real-time and rapid monitoring of the distribution area and area of submerged vegetation in Taihu Lake using remote sensing technology has important guiding significance for lake ecological restoration. The following is a case study using Taihu Lake as the research area, with the specific process as follows: Figure 1 As shown, it includes:
[0030] Step 1: Based on the acquisition time of the measured sample points, download the Sentinel-2 MSI TOA image ("COPERNICUS / S2") and the Sentinel-1 SAR GRD image ("COPERNICUS / S1_GRD") covering Taihu Lake, synchronized with the measured sample points, from Google Earth Engine (GEE) (https: / / code.earthengine.google.com / ). Declouding and cropping are performed on the Sentinel-2 MSI TOA image using the GEE cloud platform, and filtering and cropping are performed on the Sentinel-1 SAR GRD image. The filtering is performed using the Lee-Sigma method; cloud declouding is achieved by setting the values of cloudBitMask and cirrusBitMask to 0 using the QA60 band before masking.
[0031] Step 2: Perform principal component analysis on the six bands B2 (Blue), B3 (Green), B4 (Red), B8 (NIR), B11 (SWIR1), and B12 (SWIR2) of the Sentinel-2 MSI TOA image from July 28, 2018, and calculate the polarization index of the Sentinel-1 SAR GRD image from August 5, 2018.
[0032] This application tested the classification performance of different polarization indices, and the results are as follows: Figure 2 As shown, the vegetation index RVI is obtained using dual-polarization radar. dual The Dual-Pol radar vegetation index effectively distinguishes submerged vegetation from other categories, and its calculation method is as follows:
[0033]
[0034] In the formula, σ VH and σ VV Represents the VH backscattering coefficient and the VV backscattering coefficient.
[0035] Using RVI dual The first principal component with the most information content is replaced, and the inverse transformation of principal component analysis is performed on the replaced principal component analysis results to obtain six fused bands containing optical and SAR information. Among them, the fused band 1 is named SSAVI (Sentinel-based submerged aquatic vegetation index) and is used to extract submerged vegetation.
[0036] Specifically, the general formula for SSAVI is:
[0037] SSAVI = RVI dual ×X 11 +[R Blue ,R Green ,R Red ,R NIR ,R SWIR1 ,R SWIR2 ]×Y
[0038] Y = [X] ij ]×[X 1j ] T
[0039] Where Y is the weighting coefficient matrix when constructing SSAVI using bands and fusion indices; R represents band reflectance; X is the 6*6 transformation matrix in PCA transformation; and the subscripts are i = (1,2,3,4,5,6) and j = (2,3,4,5,6), respectively. Step 3: Considering that SSAVI is obtained by weighting the spectra of each band, multiplicative noise (such as atmospheric effects) cannot be eliminated, and it is very sensitive to changes in the atmospheric or aquatic environment, traditional thresholding methods struggle to determine the threshold or preset parameters. Therefore, this invention introduces Gaussian mixture model clustering to obtain submerged vegetation in SSAVI. This method can automatically obtain submerged vegetation areas based on the CI index through different SSAVI image features, that is: adaptively obtaining different optimal clustering parameters according to different SSAVI images. The CI index is constructed based on the DBI index (Davies-bouldin Index) and the BIC index (Bayesian Information Criterion), and is a weighted sum of the standardized values of the DBI and BIC indices.
[0040] In this implementation, the DBI and BIC indicators have equal weights, and the CI indicator is calculated using the following formula:
[0041]
[0042] Among them, DBI min It is the minimum DBI value corresponding to a given cluster; DBI max It is the maximum DBI value corresponding to a given cluster; BIC min It is the minimum BIC value corresponding to a given cluster; BIC max It is the maximum BIC value corresponding to a given cluster.
[0043] The BIC and DBI indices are calculated using existing methods, as follows:
[0044]
[0045] BIC = k ln(n) - 2ln(L)
[0046] In the formula, k represents the number of clusters; C x The center of cluster x; σ i Representing cluster x to C x The distance; d(c i ,c j ) represents c i and c j The distance is L; the likelihood of the given model data is L; the number of model parameters is k; and the sample size is n.
[0047] The steps for determining the optimal clustering parameters based on CI are as follows:
[0048] 1) Perform min-max standardization on the DBI and BIC indices respectively;
[0049] 2) Perform a weighted summation of DBI and BIC to obtain the CI value corresponding to the clustering parameter (number of clusters k) 3-8;
[0050] 3) Obtain the clustering parameters corresponding to the minimum CI value and set them as the optimal parameters.
[0051] The results are as follows Figure 3 As shown, in this embodiment, k = 5.
[0052] The method for determining the submerged vegetation area is as follows:
[0053] Based on the acquired SSAVI, the corresponding clustering results are obtained using the optimal clustering parameters (k=5). The mean SSAVI value for each class is calculated, and the class with the largest value is submerged vegetation.
[0054] Step 4: Graph the classification results in ArcGIS, as shown below. Figure 4Based on the measured sample points and classification results, the accuracy of submerged vegetation extraction is evaluated using producer accuracy, user accuracy, F-score, and overall accuracy. The calculation formula is as follows:
[0055]
[0056]
[0057]
[0058]
[0059] Wherein, TP stands for True positive (TP), meaning the predicted value is positive and the actual value is also positive, indicating a correct prediction; FP stands for False positive (FP), meaning the predicted value is positive but the actual value is negative, indicating an incorrect prediction; FN stands for False negative (FN), meaning the predicted value is negative but the actual value is positive, indicating an incorrect prediction; TN stands for True negative (TN), meaning the predicted value is negative and the actual value is also negative, indicating a correct prediction; PA stands for Producer Precision; UA stands for User Precision; and OA stands for Overall Precision.
[0060] The classification results are evaluated using producer accuracy, user accuracy, F-score, and overall accuracy, as shown in Table 1.
[0061] Table 1 is an accuracy evaluation table of the automatic extraction results of submerged vegetation in Taihu Lake obtained by the method described in the embodiments of the present invention. By comparing the monitoring results with the measured results, it can be seen that the monitoring accuracy of the method of the present invention for submerged vegetation in shallow eutrophic lakes (Taihu Lake) is 89.22% for producers, 96.81% for users, and the F-score is 0.92, with an overall accuracy of 88.23%, which meets most application requirements. Figure 4 This is a schematic diagram illustrating the automatic extraction results of submerged vegetation in Taihu Lake obtained by the method described in this embodiment of the invention. It can be seen that the potential and advantages of this invention compared to existing traditional methods are: it does not rely on actual measured sample points and is not limited by classification thresholds; it can quickly and accurately obtain the distribution area of submerged vegetation in lakes using only satellite imagery.
[0062] Table 1. Accuracy Evaluation of Submerged Vegetation Extraction Results in Taihu Lake
[0063]
Claims
1. An automatic extraction method for submerged vegetation in lakes based on the fusion of Sentinel-1 SAR and Sentinel-2 MSI, characterized in that, include: Principal component analysis was performed on the sensitive bands related to aquatic vegetation in the Sentinel-2 MSI TOA image of the study area; the sensitive bands related to aquatic vegetation are the B2, B3, B4, B8, B11 and B12 bands of the Sentinel-2 MSI TOA image. Acquire and process Sentinel-1 SAR GRD images with the closest time frame to the Sentinel-2 MSI TOA images, and calculate the dual-polarization radar vegetation index (RVI) from the Sentinel-1 SAR GRD images. dual Using RVI dual The first principal component in the principal component analysis is replaced, and the result of the replaced principal component analysis is inversely transformed to obtain the fused bands. Band 1 in the fused bands is used for submerged vegetation extraction. Band 1 is characterized based on the following formula: ； ； Among them, SSAVI is band 1; Y is the weighting coefficient matrix used to construct SSAVI using bands and fusion indices; R represents the band reflectivity; X is the 6x6 transformation matrix in PCA transformation; i = 1, 2, 3, 4, 5, 6; j = 2, 3, 4, 5, 6: Gaussian mixture model clustering was performed on the Band 1 images. The corresponding clustering results were obtained based on the optimal number of clusters. The mean value of Band 1 for each cluster was calculated, and the cluster with the highest mean value was identified as submerged vegetation. The optimal number of clusters was determined as follows: For different numbers of clusters, calculate the corresponding Davidson-Bourdin index (DBI) and Bayesian Information Criterion (BIC), and then standardize them. The weighted sum of the standardized DBI and standardized BIC values is denoted as the CI value. Find the number of clusters that minimizes the CI value, i.e., the optimal number of clusters.
2. The method according to claim 1, characterized in that, Using the Sentinel-2 MSI TOA image time as a reference, Sentinel-1 SAR GRD images with a time difference of 10 days were acquired for calculation; and the VV and VH polarization bands of the Sentinel-1 SAR GRD images were added to the Sentinel-2 MSI TOA images.
3. The method according to claim 1, characterized in that, The standardized DBI and standardized BIC values have equal weights when weighted and summed.
4. The method according to claim 1, characterized in that, Find the optimal number of clusters within the range of 3-8.
5. The method according to claim 1, characterized in that, The Sentinel-1 SAR GRD and Sentinel-2 MSI TOA images were acquired from the GEE cloud platform.
6. The method according to claim 1, characterized in that, It also includes preprocessing of Sentinel-1 SAR GRD and Sentinel-2 MSI TOA images, including filtering the Sentinel-1 SAR GRD image, removing clouds from the Sentinel-2 MSI TOA image, and cropping the study area from the Sentinel-1 SAR GRD and Sentinel-2 MSI TOA images for subsequent calculations.
7. The method according to claim 6, characterized in that, Filtering is performed using the Lee-Sigma method; The cloud removal process is as follows: use the QA60 band and set the values of cloudBitMask and cirrusBitMask to 0 before completing the masking.
8. The method according to claim 1, characterized in that, After the submerged vegetation was extracted, the extraction accuracy was evaluated using producer accuracy, user accuracy, F-score, and overall accuracy.