Coastal wetland damage dynamic identification method and system based on remote sensing technology

Abstract

The invention relates to the technical field of wetland damage identification, in particular to a coastal wetland damage dynamic identification method and system based on a remote sensing technology, and provides the following scheme: calling multispectral remote sensing data through a cloud computing platform, performing cloud masking, atmospheric correction and resolution resampling, and generating a high-quality remote sensing image data set; collecting ground feature samples, and constructing a random forest classification model based on spectrums, vegetation indexes, water body indexes and texture features; applying the classification model to a remote sensing image, generating a coastal wetland classification chart, and distinguishing a natural wetland from a damaged area; through time sequence analysis, ground feature changes of the same spatial position are identified, and dynamic evolution characteristics, including change amplitude, rate and conversion relation, of the damaged area of the wetland are extracted; and evaluating the precision of the classification model by using the confusion matrix, and optimizing the model according to a verification result. According to the invention, automation and precision of dynamic monitoring of wetland damage are improved, and scientific support is provided for wetland protection and management.

Description

Technical Field

[0001] The present invention relates to the technical field of wetland damage identification, and particularly to a method and system for dynamically identifying coastal wetland damage based on remote sensing technology. Background Art

[0002] Currently, as an important ecosystem, coastal wetlands have ecological functions such as maintaining biodiversity, regulating climate, and purifying water quality. However, affected by natural changes and human activities, wetlands face problems such as degradation and damage. Traditional wetland monitoring methods rely on on-site investigations, which have deficiencies such as long data acquisition cycles, high costs, and limited spatial coverage, and are difficult to meet the needs of large-scale wetland dynamic monitoring. With the development of remote sensing technology, high-resolution multispectral data and cloud computing platforms provide new technical means for wetland monitoring. However, how to effectively use remote sensing data for wetland damage dynamic identification, change analysis, and accuracy verification still faces challenges such as low classification accuracy and inaccurate extraction of dynamic evolution characteristics. Therefore, there is an urgent need for an efficient, automated, and highly accurate method for dynamically identifying coastal wetland damage based on remote sensing technology to provide reliable data support and decision-making basis for wetland protection and management.

[0003] For example, the Chinese patent with the authorization announcement number CN109886067B provides a method and device for remotely sensing and identifying wetland damage. The method includes: obtaining land use data of a target area within a first preset historical time period, obtaining wetland type data, the spatial distribution of wetland damage, and the damage type according to the land use data, obtaining the maximum wetland range according to the wetland type data, and dividing the maximum wetland range into multiple sub-regions; obtaining image data of each sub-region within a second preset historical time period, dividing the second preset historical time period into multiple sub-time periods, obtaining multiple wetland elements of each sub-region in each sub-time period according to the image data of each sub-region in each sub-time period, and calculating the damage degree of each wetland element of each sub-region in each sub-time period; determining the damage mode of each sub-region according to the wetland element with the maximum damage degree in each sub-region in each sub-time period, the spatial distribution of wetland damage, and the damage type. The embodiments of the present invention improve the accuracy and precision of wetland damage identification.

[0004] The above patents all have the problems proposed in this background art: low data acquisition efficiency, limited spatial coverage, and insufficient accuracy in identifying dynamic changes. To solve the above problems, the present application designs a method and system for dynamically identifying coastal wetland damage based on remote sensing technology. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for dynamically identifying damaged coastal wetlands based on remote sensing technology in view of the deficiencies of the prior art. By calling Sentinel-2 multispectral remote sensing data through a cloud computing platform, cloud masking, atmospheric correction, and resolution resampling are performed to generate a high-quality remote sensing image dataset; ground object samples are collected, and a random forest classification model is constructed based on spectral, vegetation index, water body index, and texture features; the classification model is applied to the remote sensing images to generate a coastal wetland classification map to distinguish natural wetlands from damaged areas; through time series analysis, the changes of ground objects at the same spatial position are identified, and the dynamic evolution characteristics of the damaged wetland areas are extracted, including the change amplitude, rate, and conversion relationship; a confusion matrix is used to evaluate the accuracy of the classification model, and the model is optimized according to the verification results. The present invention improves the automation and accuracy of dynamic monitoring of damaged wetlands and provides scientific support for wetland protection and management.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for dynamically identifying damaged coastal wetlands based on remote sensing technology, the method comprising the following steps:

[0008] S1: Call multispectral remote sensing data through a cloud computing platform to generate a remote sensing image dataset;

[0009] S2: Collect ground object samples of coastal wetlands and construct a ground object classification model for coastal wetlands according to the samples;

[0010] S3: Process the remote sensing image dataset through the ground object classification model for coastal wetlands to generate a coastal wetland classification map;

[0011] S4: Perform time series analysis on the changes of ground objects at the same spatial position according to the coastal wetland classification map to identify the damaged areas of coastal wetlands and calculate the dynamic evolution characteristics of the damaged areas;

[0012] S5: Construct a confusion matrix, verify the accuracy of the classification model according to the dynamic evolution characteristics, and update the classification model according to the verification results.

[0013] The specific steps of S1 are as follows:

[0014] S1.1: Call Sentinel-2 series remote sensing image data through the Google Earth Engine cloud computing platform, including multispectral data provided by Sentinel-2A and Sentinel-2B satellites;

[0015] S1.2: Perform cloud masking on the Sentinel-2 remote sensing image data, identify and remove cloud pixels and cirrus cloud pixels using the QA60 band to generate a cloud-free observation image;

[0016] S1.3: Perform atmospheric correction on the cloudless observation images, and convert the apparent reflectance of the images into surface reflectance;

[0017] S1.4: Perform resolution resampling on the images after atmospheric correction, and resample the band data with resolutions of 20m and 60m to 10m resolution through bicubic interpolation;

[0018] S1.5: Screen the images with cloud cover less than 10% and covering the coastal wetland area, select multi-temporal images during the wetland vegetation growth period for median synthesis, and construct a continuous remote sensing image dataset.

[0019] The coastal wetland ground object samples include:

[0020] Coastal wetland natural ground object samples, including reed, suaeda glauca, Spartina alterniflora and bare flat ground object samples;

[0021] Damaged area ground object samples, including wetland damaged ground object samples caused by human activities such as aquaculture ponds, cultivated land, construction land, and bare land;

[0022] Collect the ground object samples through on-site investigation, and use the global positioning system to record the geographic coordinates and ground object type information of the samples;

[0023] For areas that are difficult for humans to reach, conduct visual interpretation through high-resolution historical remote sensing images to supplement the remaining ground object samples.

[0024] Constructing the coastal wetland ground object classification model according to the samples includes:

[0025] S2.1: Divide the collected coastal wetland ground object sample data into a training set and a validation set. The training set accounts for 70% of the total sample size and is used for training the classification model. The validation set accounts for 30% of the total sample size and is used for evaluating the performance of the classification model;

[0026] S2.2: Adopt the random forest classification algorithm, and randomly extract training set samples for model training;

[0027] S2.3: Use the validation set to evaluate the accuracy of the trained random forest model, calculate the confusion matrix of the classification results, and obtain the evaluation indicators;

[0028] S2.4: Optimize and adjust the model parameters according to the evaluation indicators, specifically including adjusting the number and maximum depth of decision trees.

[0029] Processing the remote sensing image dataset includes:

[0030] S3.1: Extract the reflectance values of spectral bands from the preprocessed remote sensing image dataset. Among them, the reflectance values of the spectral bands include blue light (B2), green light (B3), red light (B4), red edge bands (B5, B6, B7), near-infrared (B8), red edge 4 (B8A), and short-wave infrared (B11, B12).

[0031] S3.2: Calculate vegetation indices based on the reflectance values of the spectral bands. Among them, the vegetation indices include the Normalized Difference Vegetation Index (NDVI). Enhanced Vegetation Index (EVI). and Red Edge Normalized Difference Vegetation Index (RENDVI). Ratio Vegetation Index (RVI). Difference Vegetation Index DV = B8 - B4, where θ represents a constant greater than zero.

[0032] S3.3: Calculate water body indices based on the reflectance values of the spectral bands. Among them, the water body indices include the Normalized Difference Water Index (NDWI). Modified Normalized Difference Water Index (MNDWI). Surface Water Index (SWI). where θ represents a constant greater than zero;

[0033] S3.4: Based on the Gray Level Cooccurrence Matrix (GLCM), call the "glcmTexture" function on the GEE platform to calculate the texture features of the sample area, including: mean, variance, homogeneity, contrast, and entropy;

[0034] S3.5: Perform rasterization processing on the extracted spectral, index, and texture features. Among them, the spatial resolution of the unified raster is 10m. Stitch the data within different rasters according to the pixel coordinates to form a feature dataset as the input data for the classification model;

[0035] S3.6: Apply the trained coastal wetland land cover classification model to the feature dataset, input the feature variables of the pixels, and combine the decision tree ensemble classification mechanism to automatically classify the land cover categories of each pixel;

[0036] S3.7: According to the classification results, map different land cover categories onto the raster map of the coastal wetland area to generate a coastal wetland classification map. The coastal wetland classification map distinguishes various land cover types by setting color identifiers, including:

[0037] Natural wetland land cover types based on Phragmites australis, Suaeda salsa, Spartina alterniflora, and bare flats;

[0038] Damaged area land cover types based on aquaculture ponds, cultivated land, construction land, and bare land.

[0039] The time series analysis includes:

[0040] S4.1: Match the pixel positions and land cover classes of each classified map, sort the coastal wetland classified maps generated in different periods in chronological order to form a time series classification dataset;

[0041] S4.2: Perform pixel-by-pixel difference operation on the time series classification dataset to calculate the land cover class changes of each pixel in different periods;

[0042] S4.3: According to the land cover class change results, classify the pixel changes into "unchanged" and "changed" regions, and further mark the specific land cover change types;

[0043] S4.4: Classify and count the damaged areas by land cover class, calculate the area of the land cover class change region, and calculate the change amplitude A of the damaged area c = N c × R 2 ;

[0044] where N c represents the total number of pixels identified as land cover change pixels, and R represents the spatial resolution of a single pixel;

[0045] S4.5: Based on the classified maps of different periods, count the change areas of various land covers, generate a land cover change trend chart, and calculate the change rate according to the land cover change trend chart

[0046] where represent the times of t1 and t2, and represent the areas of the land cover at time t1 and time t2, and △T represents the time interval;

[0047] S4.6: Combine the change trend and the change rate to analyze the dynamic evolution characteristics of the damaged natural wetlands, where the dynamic evolution characteristics include:

[0048] The spatial distribution of the wetland degradation trend;

[0049] The dynamic change law of the expansion or contraction of the damaged area;

[0050] The mutual conversion situation between natural wetlands and human activity interference areas.

[0051] The construction of the confusion matrix includes:

[0052] Construct a confusion matrix by comparing the classification results with the actual land cover classes in the validation set pixel by pixel;

[0053] The rows of the confusion matrix represent the land cover classes predicted by the model, the columns represent the actual land cover classes, and the elements in the matrix represent the number of pixels.

[0054] Verifying the accuracy of the classification model according to the dynamic evolution characteristics includes:

[0055] S5.1: Calculate the proportion of correctly identified changed pixels in the dynamically changing area to the total changed pixels:

[0056] S5.2: Measure the temporal consistency of the classification model in the dynamically evolving area:

[0057] S5.3: Based on the verification results, analyze the classification errors in the dynamically evolving area, where the classification errors specifically include:

[0058] Spectral confusion of ground objects, including spectral similarities between bare flats and bare lands, and between wetland vegetation and farmlands;

[0059] Misclassification or missed classification of changed areas due to the quality problems of time-series images;

[0060] The dynamic changes of ground objects are fast, and the sample data fails to fully cover their change characteristics;

[0061] S5.4: For the inconsistency between the classification errors and the dynamic evolution characteristics, perform the following optimizations:

[0062] Increase the ground object sample data in the dynamically changing area;

[0063] Adjust the parameters of the random forest classification model;

[0064] Introduce time-series characteristics in the dynamically evolving area to enhance the model's recognition ability for the dynamically changing area.

[0065] A dynamic identification system for damaged coastal wetlands based on remote sensing technology, the system includes a data acquisition module, a model construction module, a dynamic identification module, and an accuracy verification module;

[0066] The data acquisition module is used to call multi-spectral remote sensing image data, preprocess the remote sensing images, including cloud masking, atmospheric correction, and resolution resampling, to generate a remote sensing image dataset;

[0067] The model construction module is used to train a random forest classification model based on coastal wetland ground object samples and generate a high-precision classification model suitable for coastal wetland ground object classification;

[0068] The dynamic identification module is used to process remote sensing image data using the classification model, generate coastal wetland classification maps for different periods, and identify ground object changes through time-series analysis, and extract the dynamic change amplitude, change rate, and evolution characteristics of the wetland damaged area;

[0069] The accuracy verification module is used to evaluate the accuracy of the classification model, analyze the causes of classification errors, and optimize and adjust the classification model based on the verification results through the confusion matrix and the dynamic change accuracy index.

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] 1. By combining remote sensing technology with the random forest classification model, the present invention realizes the efficient and automatic dynamic identification of damaged areas of coastal wetlands, has the advantages of rapid data acquisition, high classification accuracy, accurate extraction of change features, etc., and provides scientific and reliable data support and decision-making basis for wetland protection and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments read in conjunction with the accompanying drawings:

[0073] Figure 1 It is a schematic diagram of the steps of the method for dynamically identifying damaged coastal wetlands based on remote sensing technology according to Embodiment 1 of the present invention;

[0074] Figure 2 It is a schematic flow diagram of the method for dynamically identifying damaged coastal wetlands based on remote sensing technology according to Embodiment 1 of the present invention;

[0075] Figure 3 It is a system module diagram of Embodiment 2 of the present invention.

[0076] Figure 4 It is a result display diagram of the implementation of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0077] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0078] Embodiment 1:

[0079] Please refer to Figure 1 , a schematic diagram of the steps of the method for dynamically identifying damaged coastal wetlands based on remote sensing technology according to Embodiment 1 of the present invention. The basis of the entire process is to use the remote sensing image data provided by the Sentinel-2 satellite. These data have high temporal and spatial resolutions and are important data sources for large-scale wetland monitoring.

[0080] Step 1 is image preprocessing. The main purpose of this step is to improve the image quality and lay a foundation for subsequent analysis. It includes:

[0081] Correcting the boundary is used to correct the image coordinates to ensure the spatial consistency of each image;

[0082] Cropping and mosaicking, used to retain only the part of the study area;

[0083] QA60 cloud removal processing, used to identify and remove cloud and cirrus pixels using the QA60 band;

[0084] Atmospheric correction, used to convert apparent reflectance to true surface reflectance;

[0085] Resampling, used to unify the resolution of all bands (e.g., from 20m to 10m);

[0086] Median synthesis, used to synthesize multi-temporal cloud-free images and reduce the influence of noise.

[0087] Step 2 is feature screening and dataset construction. By extracting different features, a dataset for classification is constructed, including:

[0088] Spectral features (Spectral), including ND, DV, EV, RN, RV, LSW, NDW, MNDWI. These are common vegetation indices, water body indices, and wetland indices, used to reflect the spectral characteristics of ground objects;

[0089] Texture features (Texture), including homogeneity, contrast, entropy, variance, etc.: Texture information extracted through the gray-level co-occurrence matrix (GLCM), used to distinguish ground objects with similar spectra but different textures;

[0090] Topographic features, including DEM (Digital Elevation Model), used to understand the terrain undulation and also have an auxiliary effect on wetland classification.

[0091] Step 3 is the random forest classification algorithm, used to input training samples into the random forest classifier and perform model training in combination with the above features;

[0092] The classification model conducts "voting" through multiple decision trees to improve the classification stability and accuracy;

[0093] Validation samples are used for model evaluation.

[0094] Training samples come from:

[0095] GPS surveys and remote sensing interpretation results;

[0096] Classification labels obtained on-site are used for supervised learning.

[0097] Step 4 is accuracy evaluation, specifically including:

[0098] Using the confusion matrix to verify the accuracy of the model results;

[0099] Comparing the prediction results with the validation samples to evaluate classification accuracy rate, recall rate and other indicators;

[0100] Guide subsequent model optimization.

[0101] Step 5 is for classified mapping, which is used to create a classified map from the classification results;

[0102] It distinguishes natural wetland features (such as reeds, suaeda glauca, bare beaches, Spartina alterniflora) from damaged area features (such as fish ponds, cultivated land, construction land, bare land); each type of feature will have a dedicated color coding in the map. Further analysis of the classification results specifically includes:

[0103] Time series analysis, comparing the classified maps at multiple time points;

[0104] Differential operation to identify which pixels have changed;

[0105] Dynamic analysis of the damage to coastal wetlands, analyzing the expansion, transformation, and degradation trends of the damaged areas.

[0106] Please refer to Figure 2 , an embodiment provided by the present invention: a method for dynamically identifying the damage to coastal wetlands based on remote sensing technology, the method comprising the following steps:

[0107] S1: Invoke multi-spectral remote sensing data through a cloud computing platform to generate a remote sensing image dataset;

[0108] In this step, the cloud computing platform can quickly obtain remote sensing image data with a large range, high spatio-temporal resolution, reduce the data acquisition cost, improve the data processing efficiency and quality, and provide accurate and continuous data support for subsequent analysis.

[0109] S2: Collect samples of coastal wetland features and construct a classification model for coastal wetland features based on the samples;

[0110] In this step, make full use of the measured samples and high-resolution images to improve the accuracy and reliability of the classification model. At the same time, enhance the robustness of the model through the random forest algorithm to solve the problem of complex spectral characteristics of wetland features.

[0111] S3: Process the remote sensing image dataset through the classification model for coastal wetland features to generate a coastal wetland classification map;

[0112] In this step, based on the classification model, automated processing of the remote sensing images can improve the efficiency and accuracy of wetland feature classification, quickly generate the classification results of large-scale wetlands, and provide spatial data support for subsequent change analysis.

[0113] S4: Conduct time series analysis on the changes of features at the same spatial location according to the coastal wetland classification map, identify the damaged areas of coastal wetlands, and calculate the dynamic evolution characteristics of the damaged areas;

[0114] In this step, time series analysis can accurately identify the spatio-temporal changes of wetland features, quantitatively evaluate the dynamic evolution characteristics of damaged areas, provide a scientific basis for wetland protection and management, and reveal the spatial distribution law of wetland damage.

[0115] S5: Construct a confusion matrix, verify the accuracy of the classification model according to the dynamic evolution characteristics, and update the classification model according to the verification results;

[0116] In this step, evaluating the accuracy of the classification model through a confusion matrix and conducting precision verification and model optimization in combination with dynamic evolution characteristics helps improve the classification model's ability to identify wetland dynamic changes and ensure the reliability and applicability of classification results.

[0117] The specific steps of the said S1 are as follows:

[0118] S1.1: Through the Google Earth Engine cloud computing platform, call the Sentinel-2 series remote sensing image data, including the multi-spectral data provided by the Sentinel-2A and Sentinel-2B satellites; among them, the Sentinel-2 series images provide remote sensing data with high spatio-temporal resolution, having 13 spectral bands, covering visible light, red edge, near-infrared and short-wave infrared bands, with spatial resolutions of 10m, 20m and 60m respectively, which can meet the high-precision requirements for coastal wetland feature classification and dynamic identification. By calling the image data of Sentinel-2A and Sentinel-2B in this step, the time coverage density of the data can be significantly improved, continuous monitoring during the wetland vegetation growth cycle can be ensured, and the spatio-temporal consistency and comprehensiveness of the data can be guaranteed, laying a foundation for subsequent feature classification and dynamic change analysis.

[0119] S1.2: Perform cloud masking processing on the Sentinel-2 remote sensing image data, identify and remove cloud pixels and cirrus pixels using the QA60 band, and generate cloud-free observation images;

[0120] Specifically, use the quality control band of the Sentinel-2 image, which contains cloud masking information, set a masking threshold to remove the pixels marked as cloud-covered, and ensure that the image quality is not affected by clouds and the atmosphere. Through effective cloud masking processing in this step, the usability of the image can be significantly improved, the misjudgment problem caused by cloud interference can be reduced, and the data can be made more reliable. Especially for coastal wetland areas where cloud cover is relatively common, this operation can ensure the continuity and integrity of the data and improve the subsequent classification accuracy.

[0121] S1.3: Perform atmospheric correction on the cloud-free observation images to convert the apparent reflectance of the images into surface reflectance;

[0122] The specific principle of atmospheric correction is to eliminate the influence of atmospheric scattering and absorption effects on remote sensing data. Common methods include correction methods based on radiative transfer models, such as the DOS method or the 6S model correction method. Finally, the apparent reflectance data observed in the image is converted into the true surface reflectance. This step can reduce the errors in the reflectance data caused by changes in atmospheric composition and observation conditions, enhance the comparability of image data in different periods and regions, and provide high-quality reflectance feature data for wetland land cover classification.

[0123] S1.4: Perform resolution resampling on the atmospherically corrected image, and resample the 20m and 60m resolution band data to 10m resolution through bicubic interpolation;

[0124] The purpose of resolution resampling is to make the spectral bands with different resolutions in the Sentinel-2 image consistent in spatial resolution, facilitating data fusion and subsequent analysis and processing. The bicubic interpolation method used calculates the weighted average of the 16 pixels surrounding the target pixel to obtain a smoother and more delicate resampling result. This step can make the spatial resolution of all band data reach 10m, facilitating subsequent multi-band feature extraction and the training of classification models, and improving the classification accuracy and the spatial detail performance of dynamic change detection.

[0125] S1.5: Screen images with cloud cover less than 10% and covering the coastal wetland area, and select multi-temporal images during the growth period of wetland vegetation for median synthesis to construct a continuous remote sensing image dataset;

[0126] By setting image screening conditions, including cloud cover ratio (less than 10%), spatial coverage (matching the study area ROI), and time range (usually the peak growing season of wetland vegetation, such as April - October), a multi-temporal dataset is extracted from high-quality images to form high-quality time-series image data with temporal continuity and spatial integrity. This step reduces the influence of factors such as clouds and shadows on data quality to the greatest extent by strictly screening high-quality images, and at the same time ensures continuous observation data of wetland vegetation in the study area during the growing season, providing a complete spatio-temporal data basis for time-series analysis and facilitating the accurate identification of the dynamic evolution process of wetland land cover.

[0127] The coastal wetland land cover samples include:

[0128] Coastal wetland natural land cover samples, including reed, suaeda, Spartina alterniflora, and bare flat land cover samples;

[0129] Damaged area land cover samples, including wetland damaged land cover samples caused by human activities such as aquaculture ponds, cultivated land, construction land, and bare land;

[0130] For the ground object samples collected through field surveys, the geographic coordinates and ground object type information of the samples are recorded using the Global Positioning System;

[0131] For areas that are difficult for humans to reach, visual interpretation is carried out through high-resolution historical remote sensing images to supplement the remaining ground object samples.

[0132] The specific steps of S2 are as follows:

[0133] S2.1: Divide the collected ground object sample data of coastal wetlands into a training set and a validation set. The training set accounts for 70% of the total sample size and is used for training the classification model. The validation set accounts for 30% of the total sample size and is used to evaluate the performance of the classification model;

[0134] First, organize and preprocess the collected ground object sample data of coastal wetlands to ensure the integrity, balance, and effectiveness of the sample data. By checking the spatial distribution of the samples, ensure that the distribution of ground object categories in the training set and the validation set is balanced, covering all natural wetland categories and damaged area categories. When dividing, use the principle of combining random sampling and stratified sampling to avoid too few or unbalanced sample data of specific categories and ensure the generalization ability of the classification model.

[0135] S2.2: Adopt the random forest classification algorithm and randomly select training set samples for model training;

[0136] Random forest is an ensemble learning method that uses the Bootstrap method to randomly and with replacement draw samples from the training set to construct multiple decision trees. Each decision tree randomly selects some feature variables for classification when splitting nodes. The final classification result is determined by voting or averaging the results of multiple decision trees. In terms of technical implementation, by setting parameters such as the number of decision trees and the maximum depth of the tree, the classification model has sufficient complexity and can prevent overfitting problems.

[0137] S2.3: The validation set evaluates the accuracy of the trained random forest model, calculates the confusion matrix of the classification result, and obtains the evaluation indicators;

[0138] Use the validation set data to evaluate the accuracy of the trained random forest classification model, and evaluate the classification performance of the model by constructing a confusion matrix. The confusion matrix can intuitively reflect the comparison between the classification result of the model and the actual validation data. The diagonal elements represent the number of correctly classified samples, and the non-diagonal elements represent the number of misclassified samples.

[0139] S2.4: According to the evaluation indicators, optimize and adjust the model parameters, specifically including adjusting the number of decision trees and the maximum depth;

[0140] Gradually test the impact of different parameter combinations on the model performance through cross-validation methods to find the optimal parameter settings. For example, appropriately increasing the number of decision trees can improve the stability of the model, while restricting the maximum tree depth can reduce the calculation time and improve the generalization ability. The advantage of doing this is to reduce the model misclassification phenomenon through parameter optimization, improve the classification accuracy, and ensure the applicability and accuracy of the classification model for different wetland land cover classes. At the same time, parameter optimization can effectively improve the computational efficiency of the model to meet the processing requirements for large-scale data in practical applications.

[0141] The specific steps of S3 are as follows:

[0142] S3.1: Extract the reflectance values of spectral bands from the preprocessed remote sensing image dataset, where the reflectance values of the spectral bands include blue (B2), green (B3), red (B4), red edge bands (B5, B6, B7), near-infrared (B8), red edge 4 (B8A), and short-wave infrared (B11, B12);

[0143] The extraction of these bands is based on the differences in the reflection characteristics of different land covers to the electromagnetic spectrum. For example, vegetation land covers have a relatively high reflectance in the near-infrared band, while water bodies have a relatively low reflectance in the short-wave infrared band. This spectral difference can effectively distinguish natural and damaged land covers in wetlands. By extracting the reflectance values of these multi-spectral bands, the most basic and direct land cover information is provided, laying a data foundation for subsequent index calculation and classification.

[0144] S3.2: Calculate vegetation indices based on the reflectance values of the spectral bands, where the vegetation indices include the Normalized Difference Vegetation Index Enhanced Vegetation Index and Red Edge Normalized Difference Vegetation Index Ratio Vegetation Index Difference Vegetation Index DV = B8 - B4, where θ represents a constant greater than zero;

[0145] These indices enhance the discrimination ability between vegetation and other land covers through mathematical transformation of spectral data, effectively improving the recognition accuracy of natural wetland land covers (such as reeds and suaeda salsa).

[0146] S3.3: Calculate water body indices based on the reflectance values of the spectral bands, where the water body indices include the Normalized Difference Water Index Modified Normalized Difference Water Index Surface Water Index where θ represents a constant greater than zero;

[0147] Through the calculation of water body indices, the water areas and damaged wetland water body areas in wetlands can be accurately located, especially for the situation of reduced water body area, with a high recognition effect.

[0148] S3.4: Further, based on the Gray-Level Co-Occurrence Matrix (GLCM), the "glcmTexture" function is called on the GEE platform to calculate the texture features of the sample area, including parameters such as mean, variance, homogeneity, contrast, and entropy. GLCM extracts the texture information of the image by analyzing the gray value distribution relationship between pixels. For example, there are obvious differences in texture features between ground objects such as wetland vegetation and bare flats. Texture mean and variance reflect the central tendency and dispersion degree of gray values, homogeneity reflects the smoothness of the image, contrast represents the difference degree between pixel gray values, and entropy measures the complexity of image texture. These texture features make up for the deficiencies of spectral features in identifying complex ground objects and provide more comprehensive information support for wetland ground object classification.

[0149] S3.5: The extracted spectral, vegetation index, water body index, and texture features are rasterized, and the spatial resolution of the unified raster is set to 10m. The data of different feature variables are stitched according to the pixel coordinates to form a unified feature dataset. Rasterization ensures that all feature variables are consistent in spatial position and resolution, which helps the classification model process data more efficiently and avoids classification errors caused by inconsistent resolution or coordinates.

[0150] S3.6: Apply the trained coastal wetland ground object classification model to the said feature dataset, input the spectral, index, and texture features of the pixels, and combine the decision tree integration mechanism of the random forest classification algorithm to automatically determine the ground object category of each pixel. The random forest algorithm determines the final classification result through the voting mechanism of multiple decision trees, has strong anti-noise performance and high-precision performance, and can effectively handle the problem of complex ground object spectral confusion in wetlands.

[0151] S3.7: According to the classification results, map different ground object categories onto the raster map of the coastal wetland area to generate the final coastal wetland classification map, and distinguish different ground object categories by setting color identifiers, including natural wetland ground objects (such as reeds, Suaeda salsa, Spartina alterniflora, bare flats) and damaged area ground objects (such as aquaculture ponds, cultivated land, construction land, bare land). The classification map can visually display the spatial distribution of the wetland and provide clear spatial data support for the identification and dynamic monitoring of wetland damaged areas.

[0152] The specific steps of S4 are as follows:

[0153] S4.1: Match the pixel positions and ground object categories of each classification map, sort the coastal wetland classification maps generated in different periods in chronological order to form a time series classification dataset;

[0154] S4.2: Perform pixel-by-pixel difference operation on the said time series classification dataset to calculate the change of ground object categories of each pixel in different periods.

[0155] S4.3: Classify the pixel changes into "unchanged" and "changed" areas according to the results of the land cover type changes, and further mark the specific types of land cover changes;

[0156] S4.4: Classify and count the damaged areas by land cover type, calculate the area of the land cover type change area, and calculate the change amplitude A of the damaged area c = N c × R 2 ;

[0157] where N c represents the total number of pixels identified as land cover change pixels, and R represents the spatial resolution of a single pixel;

[0158] S4.5: Based on the classification maps of different periods, count the change areas of various land covers, generate a land cover change trend map, and calculate the change rate according to the land cover change trend map

[0159] where represent the times of t1 and t2, and represent the areas of the land cover at times t1 and t2, and △T represents the time interval;

[0160] S4.6: Combine the change trend and the change rate to analyze the dynamic evolution characteristics of the damaged natural wetlands. Among them, the dynamic evolution characteristics include:

[0161] The spatial distribution of the wetland degradation trend;

[0162] The dynamic change law of the expansion or contraction of the damaged area;

[0163] The mutual transformation between natural wetlands and human activity interference areas.

[0164] The construction of the confusion matrix includes:

[0165] Construct a confusion matrix by comparing the classification results with the actual land cover types in the validation set pixel by pixel;

[0166] The rows of the confusion matrix represent the land cover types predicted by the model, the columns represent the actual land cover types, and the elements in the matrix represent the number of pixels.

[0167] The specific steps of S5 are as follows:

[0168] S5.1: Calculate the proportion of correctly identified change pixels in the total change pixels in the dynamic change area:

[0169] S5.2: Measure the temporal consistency of the classification model in the dynamic evolution area:

[0170] S5.3: Analyze the classification errors within the dynamically evolving region based on the verification results. Specifically, the classification errors include:

[0171] Spectral confusion of ground objects, including spectral similarities between light beaches and bare lands, and between wetland vegetation and farmlands;

[0172] Misclassification or missed classification of changed regions due to the quality issues of time-series images;

[0173] Rapid dynamic changes of ground objects, and the sample data fails to fully cover their change characteristics;

[0174] S5.4: For the inconsistency between the classification errors and the dynamic evolution characteristics, perform the following optimizations:

[0175] Increase the ground object sample data of the dynamically changing region;

[0176] Adjust the parameters of the random forest classification model;

[0177] Introduce time-series features in the dynamically evolving region to enhance the model's recognition ability for the dynamically changing region.

[0178] Example 2:

[0179] Please refer to Figure 2 , the present invention provides an example: a dynamic identification system for damaged coastal wetlands based on remote sensing technology. The system includes a data acquisition module, a model construction module, a dynamic identification module, and an accuracy verification module;

[0180] The data acquisition module is used to call multi-spectral remote sensing image data, preprocess the remote sensing images, including cloud masking, atmospheric correction, and resolution resampling, to generate a remote sensing image dataset;

[0181] The model construction module is used to train a random forest classification model based on coastal wetland ground object samples and generate a high-precision classification model suitable for coastal wetland ground object classification;

[0182] The dynamic identification module is used to process the remote sensing image data using the classification model, generate coastal wetland classification maps for different periods, and identify ground object changes through time-series analysis, and extract the dynamic change amplitude, change rate, and evolution characteristics of the wetland damaged area;

[0183] The accuracy verification module is used to evaluate the accuracy of the classification model through a confusion matrix and dynamic change accuracy indicators, analyze the causes of classification errors, and optimize and adjust the classification model based on the verification results.

[0184] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for dynamically identifying damaged coastal wetlands based on remote sensing technology, characterized in that, The method includes the following steps: S1: Invoke multi-spectral remote sensing data through a cloud computing platform to generate a remote sensing image dataset; S2: Collect coastal wetland ground object samples and construct a coastal wetland ground object classification model based on the samples; S3: Process the remote sensing image dataset through the coastal wetland ground object classification model to generate a coastal wetland classification map; S4: Conduct time series analysis on the ground object changes at the same spatial location based on the coastal wetland classification map, identify the damaged areas of the coastal wetland, and calculate the dynamic evolution characteristics of the damaged areas; S5: Construct a confusion matrix, verify the accuracy of the classification model based on the dynamic evolution characteristics, and update the classification model according to the verification results.

2. The dynamic identification method for damaged coastal wetlands based on remote sensing technology according to claim 1, wherein The specific steps of S1 are as follows: S1.1: Through the Google Earth Engine cloud computing platform, invoke Sentinel-2 series remote sensing image data, including multi-spectral data provided by Sentinel-2A and Sentinel-2B satellites; S1.2: Perform cloud masking on the Sentinel-2 remote sensing image data, identify and remove cloud pixels and cirrus cloud pixels using the QA60 band to generate cloud-free observation images; S1.3: Perform atmospheric correction on the cloud-free observation images to convert the apparent reflectance of the images into surface reflectance; S1.4: Resample the resolution of the images after atmospheric correction, and resample the band data with 20m and 60m resolutions to 10m resolution through bicubic interpolation; S1.5: Screen images with cloud cover less than 10% and covering the coastal wetland area, select multi-temporal images during the wetland vegetation growth period for median synthesis, and construct a continuous remote sensing image dataset.

3. The method for dynamically identifying damaged coastal wetlands based on remote sensing technology according to claim 2, characterized in that The coastal wetland ground object samples include: Coastal wetland natural ground object samples, including reed, suaeda salsa, Spartina alterniflora, and bare flat ground object samples; Damaged area ground object samples, including wetland damaged ground object samples caused by human activities such as aquaculture ponds, cultivated land, construction land, and bare land; Ground object samples collected through field surveys, and use the global positioning system to record the geographical coordinates and ground object type information of the samples; For areas that are difficult for humans to reach, conduct visual interpretation through high-resolution historical remote sensing images to supplement the remaining ground object samples.

4. The method for dynamically identifying damaged coastal wetlands based on remote sensing technology according to claim 3, wherein Constructing the coastal wetland ground object classification model according to the samples includes: S2.1: Divide the collected coastal wetland ground object sample data into a training set and a validation set. The training set accounts for 70% of the total sample volume and is used for training the classification model. The validation set accounts for 30% of the total sample volume and is used for evaluating the performance of the classification model; S2.2: Adopt the random forest classification algorithm and randomly extract training set samples for model training; S2.3: The validation set evaluates the accuracy of the trained random forest model, calculates the confusion matrix of the classification results, and obtains evaluation indicators; S2.4: Optimize and adjust the model parameters according to the evaluation indicators, specifically including adjusting the number and maximum depth of decision trees.

5. The method for dynamically identifying damaged coastal wetlands based on remote sensing technology according to claim 4, characterized in that, Processing the remote sensing image dataset includes: S3.1: Extract the reflectance values of spectral bands from the preprocessed remote sensing image dataset, where the reflectance values of the spectral bands include blue light (B2), green light (B3), red light (B4), red edge bands (B5, B6, B7), near-infrared (B8), red edge 4 (B8A), and shortwave infrared (B11, B12); S3.2: Calculate the vegetation index based on the reflectance values of the spectral bands, where the vegetation index includes the Normalized Difference Vegetation Index Enhanced Vegetation Index and the Red Edge Normalized Difference Vegetation Index Ratio Vegetation Index , and the Difference Vegetation Index DV = B8 - B4, where θ represents a constant greater than zero S3.3: Calculate the water body index based on the reflectance values of the spectral bands, where the water body index includes the Normalized Difference Water Index Modified Normalized Difference Water Index Surface Water Body Index where θ represents a constant greater than zero; S3.4: Based on the Gray Level Cooccurrence Matrix (GLCM), call the "glcmTexture" function on the GEE platform to calculate the texture features of the sample area, including: mean, variance, homogeneity, contrast, and entropy; S3.5: Perform rasterization processing on the extracted spectral, index, and texture features. Among them, the spatial resolution of the unified raster is 10m, and the data in different rasters are stitched according to the pixel coordinates to form a feature dataset as the input data for the classification model; S3.6: Apply the trained coastal wetland land cover classification model to the feature dataset, input the feature variables of the pixels, and combine the decision tree ensemble classification mechanism to automatically classify the land cover categories of each pixel; S3.7: According to the classification results, map different land cover categories onto the raster map of the coastal wetland area to generate a coastal wetland classification map. The coastal wetland classification map distinguishes various land cover types by setting color identifiers, including: Natural wetland land cover types based on reeds, Suaeda salsa, Spartina alterniflora, and tidal flats; Damaged area land cover types based on aquaculture ponds, cultivated land, construction land, and bare land.

6. The method for dynamically identifying damaged coastal wetlands based on remote sensing technology according to claim 5, characterized in that, The time series analysis includes: S4.1: Match the pixel positions and land cover categories of each classification map, sort the coastal wetland classification maps generated in different periods in chronological order to form a time series classification dataset; S4.2: Perform pixel-by-pixel difference operations on the time series classification dataset to calculate the changes in land cover categories of each pixel in different periods; S4.3: According to the results of land cover category changes, classify the pixel changes into "unchanged" and "changed" areas, and further mark the specific types of land cover changes; S4.4: Classify and statistically analyze the damaged area according to the land cover types, calculate the area of the land cover type change area, and calculate the change amplitude A of the damaged area c = N c × R 2 ; Among them, N c represents the total number of pixels recognized as land cover change pixels, and R represents the spatial resolution of a single pixel; S4.5: Based on the classification maps of different periods, calculate the change areas of various ground features, generate a ground feature change trend map, and calculate the change rate according to the ground feature change trend map where, \(t_1\) and \(t_2\) represent time, \(A_{t2}\) and \(A_{t1}\) represent the areas of land cover at time \(t_1\) and \(t_2\), and \(\Delta T\) represents the time interval; S4.6: Combine the change trend and change rate to analyze the dynamic evolution characteristics of natural wetland damage, where the dynamic evolution characteristics include: The spatial distribution of wetland degradation trends; The dynamic change laws of the expansion or contraction of damaged areas; The mutual transformation between natural wetlands and human activity interference areas.

7. The method for dynamically identifying damaged coastal wetlands based on remote sensing technology according to claim 6, wherein The construction of the confusion matrix includes: Construct a confusion matrix by comparing the classification results with the actual land cover categories in the validation set pixel by pixel; The rows of the confusion matrix represent the land cover categories predicted by the model, the columns represent the actual land cover categories, and the elements in the matrix represent the number of pixels.

8. The method for dynamically identifying damaged coastal wetlands based on remote sensing technology according to claim 7, wherein The verification of the accuracy of the classification model according to the dynamic evolution characteristics includes: S5.1: Calculate the proportion of correctly identified changed pixels in the total changed pixels in the dynamic change area: S5.2: Measure the time series consistency of the classification model in the dynamic evolution area: S5.3: Based on the verification results, analyze the classification errors in the dynamically evolving area. Specifically, the classification errors include: Spectral confusion of ground objects, including spectral similarities between bare flats and bare land, and between wetland vegetation and farmland; Misclassification or missed classification of changed areas due to the quality problems of time-series images; Rapid dynamic changes of ground objects, and the sample data fails to fully cover their change characteristics; S5.4: For the inconsistency between the classification errors and the dynamic evolution characteristics, perform the following optimizations: Increase the ground object sample data in the dynamically changing area; Adjust the parameters of the random forest classification model; Introduce time-series features in the dynamically evolving area to enhance the model's recognition ability for the dynamically changing area.

9. A dynamic identification system for damaged coastal wetlands based on remote sensing technology, which is implemented based on the dynamic identification method for damaged coastal wetlands based on remote sensing technology according to any one of claims 1-8, characterized in that, The system includes a data acquisition module, a model construction module, a dynamic recognition module, and an accuracy verification module; The data acquisition module is used to call multi-spectral remote sensing image data, preprocess the remote sensing images, including cloud masking, atmospheric correction, and resolution resampling, to generate a remote sensing image dataset; The model construction module is used to train a random forest classification model based on the ground object samples of coastal wetlands and generate a high-precision classification model suitable for the classification of coastal wetland ground objects; The dynamic recognition module is used to process the remote sensing image data using the classification model, generate coastal wetland classification maps for different periods, and identify ground object changes through time-series analysis, and extract the dynamic change amplitude, change rate, and evolution characteristics of the damaged wetland areas; The accuracy verification module is used to evaluate the accuracy of the classification model through the confusion matrix and dynamic change accuracy indicators, analyze the causes of classification errors, and optimize and adjust the classification model based on the verification results.

Images