Mangrove extraction method and device based on multi-source data step-by-step correction

Abstract

The application discloses a mangrove extraction method and device based on multi-source data step-by-step correction, and belongs to the technical field of remote sensing monitoring. The method comprises the following steps: acquiring a Sentinel-2 multispectral image, a GF-2 multispectral image, an on-site unmanned aerial vehicle remote sensing image and DEM data; performing preliminary extraction on mangroves based on the spectral index of the Sentinel-2 multispectral image, and performing error correction based on the on-site unmanned aerial vehicle remote sensing image; based on the GF-2 multispectral image, correcting the boundary of the initially extracted mangrove region by adopting a machine learning method, and performing error correction based on the on-site unmanned aerial vehicle remote sensing image; constructing a terrain correction index based on the DEM data and the GF-2 multispectral image, and removing potential water accumulation regions from the corrected mangrove region to obtain a final mangrove distribution region. The application adopts a progressive framework of "spectral extraction-geometric correction-terrain optimization", solves the problem of low mangrove extraction precision caused by spectral confusion from the perspective of terrain hydrology, and improves the extraction precision of mangroves.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring technology, specifically to a method and apparatus for extracting mangrove forests based on multi-source data with step-by-step correction. Background Technology

[0002] Mangroves are precious ecosystems distributed in the intertidal zones of tropical and subtropical coasts, playing vital ecological roles such as wind and wave protection, water purification, carbon sequestration, and biodiversity maintenance. Timely and accurate understanding of their spatial distribution and dynamic changes is fundamental to mangrove protection, restoration, and management. Traditional methods of in-situ monitoring of mangroves provide accurate sample data, laying the foundation for research. However, due to the specific distribution areas of mangroves, in-situ sampling is time-consuming and labor-intensive, and the resulting data exhibits poor spatial continuity.

[0003] Remote sensing technology, capable of long-distance detection, is an irreplaceable and crucial tool for mangrove monitoring due to its macroscopic and rapid advantages. With the increasing abundance of remote sensing data sources and continuous advancements in analytical methods, mangrove information extraction technology is evolving from reliance on single data sources and simple algorithms to the integration of multi-source data and complex models. However, in this development process, the severe spectral confusion phenomenon (i.e., "different objects, same spectrum") between mangroves located in the flat intertidal zone and adjacent stagnant water areas remains a core challenge and difficulty for high-precision extraction. To overcome this problem, existing technologies have undergone the following development paths, each of which, while improving accuracy, also reveals its inherent limitations.

[0004] Single-data-source methods based on spectral features offer strong universality but are limited in accuracy and detail. These methods form the basis of mangrove remote sensing monitoring, primarily relying on spectral information from single-type medium-resolution optical remote sensing images (such as the Landsat series and Sentinel-2). Extraction is achieved by calculating spectral indices (such as Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI)) and setting empirical thresholds. The advantages of this method are its strong universality and high computational efficiency, making it suitable for large-scale rapid surveys. However, its inherent limitation lies in the pixel mixing effect of medium-resolution images, which leads to blurred boundaries and loss of detail in the extracted results. This is particularly pronounced in spectrally confused areas where mangroves and water bodies are interspersed, where misclassification is significant, limiting extraction accuracy. Even utilizing richer band information from some images struggles to effectively distinguish fine boundaries at the land-water interface in flat areas.

[0005] With the rapid development of remote sensing technology, the spatial resolution of images has gradually improved. Using high spatial resolution imagery (such as GF-2 and WorldView series) and employing machine learning algorithms to comprehensively mine the spectral and spatial texture features of the images has significantly optimized the geometric contours of mangrove patches, achieving more refined identification than medium-resolution imagery. However, the application of high-resolution imagery often involves the construction of high-dimensional feature sets, which easily introduces feature redundancy. Feature redundancy not only fails to improve model performance but also dilutes the contribution of the truly crucial spectral features for distinguishing mangroves from water bodies, affecting the model's generalization ability. Furthermore, high-resolution imagery is mostly multispectral data with a wide spectral band width and a limited number of bands, making it difficult to fully utilize spectral information for accurate mangrove extraction. More importantly, even though the spatial resolution of images has improved to sub-meter level, the problem of spectral confusion between mangroves and stagnant water areas due to "different objects sharing the same spectrum" has not been well resolved.

[0006] The limitations of single data sources have made multi-source data fusion the primary method in mangrove research. Current multi-source data fusion strategies commonly combine DEM data with optical imagery. However, existing fusion methods remain at the level of "feature overlay," meaning that DEM and spectral features are input into the classifier together, failing to fully utilize the information provided by the DEM data. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a mangrove extraction method and apparatus based on multi-source data with stepwise correction. This method solves the problem of low mangrove extraction accuracy caused by spectral confusion from the perspective of topographic and hydrological factors, thereby improving the extraction accuracy of mangroves.

[0008] The technical solution provided by this invention is as follows:

[0009] A method for extracting mangrove forests based on multi-source data with step-by-step correction, the method comprising:

[0010] S1: Obtain multi-source data for the same time period in the study area;

[0011] The multi-source data includes medium-resolution Sentinel-2 multispectral imagery, high-resolution GF-2 multispectral imagery, on-site UAV remote sensing imagery, and DEM data.

[0012] S2: Preprocess the Sentinel-2 multispectral image, GF-2 multispectral image, and on-site UAV remote sensing image;

[0013] S3: Based on the spectral indices of the preprocessed Sentinel-2 multispectral image, the mangrove forest area is initially extracted, and the error is corrected based on the on-site UAV remote sensing image to obtain the initially extracted mangrove forest area.

[0014] S4: Based on the preprocessed GF-2 multispectral image, machine learning methods are used to correct the boundaries of the initially extracted mangrove area, and error correction is performed based on the on-site UAV remote sensing image to obtain the corrected mangrove area.

[0015] S5: Construct a topographic correction index based on DEM data and preprocessed GF-2 multispectral image, and remove potential water accumulation areas from the corrected mangrove area according to the topographic correction index to obtain the final mangrove distribution area;

[0016] The formula for calculating the terrain correction index is as follows:

[0017]

[0018] S represents the terrain correction index, and RE and ESD represent the slope, relative elevation, and standard deviation of elevation obtained from the DEM data, respectively. The normalized water index is calculated based on the preprocessed GF-2 multispectral image, where a, b, c, and d are the weights of slope, relative elevation, elevation standard deviation, and normalized water index, respectively.

[0019] Furthermore, S3 includes:

[0020] S31: Use the boundary vector of the study area to crop the preprocessed Sentinel-2 multispectral image to obtain the medium-resolution image data of the study area;

[0021] S32: Calculate the spectral index of each pixel in the study area based on the medium-resolution image data of the study area;

[0022] The spectral indices include one or more of the following: red-edge normalized vegetation index, red-edge location index, and water stress index, and the calculation formula is as follows:

[0023]

[0024]

[0025]

[0026] RE_NDVI is the red-edge normalized vegetation index, REP is the red-edge position index, MSI is the water stress index, and B4, B5, B7, B8A, and B11 represent the remote sensing reflectance of bands 4, 5, 7, 8A, and 11 of the preprocessed Sentinel-2 multispectral image, respectively.

[0027] S33: Compare each spectral index with the set index thresholds to obtain the mangrove spectral extraction results.

[0028] Furthermore, S33 includes:

[0029] S331: For each pixel in the study area, the red-edge normalized vegetation index of the pixel is compared with the set vegetation index threshold. Pixels with a red-edge normalized vegetation index greater than the vegetation index threshold are classified as mangroves, and the first spectral extraction result is obtained.

[0030] S332: For each pixel in the study area, compare the red edge position index of the pixel with the set position index threshold, and classify the pixels with the red edge position index greater than the position index threshold as mangroves to obtain the second spectral extraction result;

[0031] S333: For each pixel in the study area, the water stress index of the pixel is compared with the set water stress index threshold. Pixels with a water stress index less than the water stress index threshold are classified as mangroves, and the third spectral extraction result is obtained.

[0032] S334: Take the union of the first spectral extraction result, the second spectral extraction result, and the third spectral extraction result to obtain the mangrove spectral extraction result.

[0033] Furthermore, S3 also includes:

[0034] S34: Obtain real points of mangrove forests based on on-site drone remote sensing images, and use them as verification sample points;

[0035] S35: Compare the verification sample points with the mangrove spectral extraction results to obtain the first misclassification set;

[0036] The first misclassification set includes misclassification points that classify non-mangrove forests as mangrove forests and omission points that classify mangrove forests as non-mangrove forests;

[0037] S36: When the proportion of misclassified points in the first misclassification set exceeds the set proportion threshold, adjust the index threshold and re-extract the mangrove spectral extraction results;

[0038] Specifically, when adjusting the index thresholds, the values ​​of the vegetation index threshold and the location index threshold are increased, while the value of the water stress index threshold is decreased.

[0039] S37: Construct a buffer zone outward from the omission points in the first misclassification set to obtain the omission distribution area;

[0040] S38: Based on the spectral characteristics of the underscored distribution area, redetermine the index threshold, and use the redetermined index threshold to extract mangroves in the underscored distribution area to obtain the underscored area extraction result;

[0041] S39: Take the union of the mangrove spectral extraction results and the missing region extraction results to obtain the initially extracted mangrove region.

[0042] Furthermore, S4 includes:

[0043] S41: Based on the boundary of the initially extracted mangrove area, a buffer zone with a set distance is constructed on both the inner and outer sides of the boundary to obtain the boundary area of ​​the initially extracted mangrove area;

[0044] S42: Based on the boundary vector of the initially extracted mangrove boundary region, the preprocessed GF-2 multispectral image is cropped to obtain the boundary region to be corrected.

[0045] S43: The random forest-based machine learning model is trained using a selected training set containing positive and negative samples, and the mangrove forests in the boundary area to be corrected are extracted using the trained machine learning model to obtain the mangrove forest extraction result in the boundary area.

[0046] S44: Compare the verification sample points with the mangrove extraction results of the boundary area to obtain the second misclassification set;

[0047] The second misclassification set includes misclassification points that classify non-mangrove forests as mangrove forests and omission points that classify mangrove forests as non-mangrove forests;

[0048] S45: The misclassified points and omissions in the second misclassification set are used as negative samples and positive samples, respectively, to fine-tune the machine learning model trained on the training set. The fine-tuned machine learning model is then used to extract the mangroves in the boundary area to be corrected, and the mangrove boundary correction result is obtained.

[0049] S46: Replace the initially extracted mangrove boundary region within the initially extracted mangrove region with the mangrove boundary correction result to obtain the corrected mangrove region.

[0050] Furthermore, S5 includes:

[0051] S51: Based on DEM data, calculate the slope, relative elevation, and elevation standard deviation of each pixel in the study area using geographic information system software;

[0052] S52: Calculate the normalized water index for each pixel in the study area based on the preprocessed GF-2 multispectral image;

[0053]

[0054] Wherein, NDWI is the normalized water index, and Green and NIR represent the remote sensing reflectance of the green band and near-infrared band of the preprocessed GF-2 multispectral image, respectively.

[0055] S53: The terrain correction index for each pixel within the study area is constructed using the following formula:

[0056]

[0057] S54: Compare the terrain correction index of each pixel in the study area with the determined segmentation threshold, and classify the pixels with terrain correction index greater than the segmentation threshold as potential water accumulation areas.

[0058] S55: Remove the potential waterlogged area from the modified mangrove area to obtain the final mangrove distribution area.

[0059] Furthermore, the weights of the slope, relative elevation, elevation standard deviation, and normalized water index are determined as follows:

[0060] The subjective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the analytic hierarchy process (AHP).

[0061] The objective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the CRITIC method.

[0062] The subjective and objective weights are linearly weighted to obtain the weights of slope, relative elevation, elevation standard deviation, and normalized water index.

[0063] Furthermore, the segmentation threshold is determined in the following manner:

[0064] The dataset of the terrain correction index was subjected to unsupervised clustering fitting using a Gaussian mixture model, resulting in two Gaussian distributions.

[0065] The value of the terrain correction index corresponding to the intersection of the probability density functions of the two Gaussian distributions is used as the segmentation threshold.

[0066] A mangrove forest extraction device based on multi-source data stepwise correction, the device comprising:

[0067] The data acquisition module is used to acquire multi-source data of the same time period in the study area;

[0068] The multi-source data includes medium-resolution Sentinel-2 multispectral imagery, high-resolution GF-2 multispectral imagery, on-site UAV remote sensing imagery, and DEM data.

[0069] The preprocessing module is used to preprocess the Sentinel-2 multispectral image, GF-2 multispectral image, and on-site UAV remote sensing image.

[0070] The initial extraction module is used to perform preliminary extraction of mangroves based on the spectral indices of the preprocessed Sentinel-2 multispectral image, and to perform error correction based on the on-site UAV remote sensing image to obtain the initially extracted mangrove area.

[0071] The boundary correction module is used to correct the boundaries of the initially extracted mangrove area based on the preprocessed GF-2 multispectral image using machine learning methods, and to correct errors based on the on-site UAV remote sensing image to obtain the corrected mangrove area.

[0072] The terrain correction module is used to construct a terrain correction index based on DEM data and preprocessed GF-2 multispectral imagery, and remove potential water accumulation areas from the corrected mangrove area according to the terrain correction index to obtain the final mangrove distribution area.

[0073] The formula for calculating the terrain correction index is as follows:

[0074]

[0075] S represents the terrain correction index, and RE and ESD represent the slope, relative elevation, and standard deviation of elevation obtained from the DEM data, respectively. The normalized water index is calculated based on the preprocessed GF-2 multispectral image, where a, b, c, and d are the weights of slope, relative elevation, elevation standard deviation, and normalized water index, respectively.

[0076] Furthermore, the initial extraction module includes:

[0077] The first cropping unit is used to crop the preprocessed Sentinel-2 multispectral image using the boundary vector of the study area to obtain medium-resolution image data of the study area.

[0078] The spectral index calculation unit is used to calculate the spectral index of each pixel in the study area based on the medium-resolution image data of the study area.

[0079] The spectral indices include one or more of the following: red-edge normalized vegetation index, red-edge location index, and water stress index, and the calculation formula is as follows:

[0080]

[0081]

[0082]

[0083] RE_NDVI is the red-edge normalized vegetation index, REP is the red-edge position index, MSI is the water stress index, and B4, B5, B7, B8A, and B11 represent the remote sensing reflectance of bands 4, 5, 7, 8A, and 11 of the preprocessed Sentinel-2 multispectral image, respectively.

[0084] The spectral index segmentation unit is used to compare each spectral index with the set index thresholds to obtain the mangrove spectral extraction results.

[0085] Furthermore, the spectral index segmentation unit includes:

[0086] The first segmentation subunit is used to compare the red-edge normalized vegetation index of each pixel in the study area with the set vegetation index threshold, and classify the pixels with the red-edge normalized vegetation index greater than the vegetation index threshold as mangroves to obtain the first spectral extraction result.

[0087] The second segmentation subunit is used to compare the red edge position index of each pixel in the study area with a set position index threshold, and classify the pixels with the red edge position index greater than the position index threshold as mangroves to obtain the second spectral extraction result.

[0088] The third segmentation subunit is used to compare the water stress index of each pixel in the study area with the set water stress index threshold, and classify the pixels with water stress index less than the water stress index threshold as mangroves to obtain the third spectral extraction result.

[0089] The mangrove spectral extraction result determination subunit is used to take the union of the first spectral extraction result, the second spectral extraction result, and the third spectral extraction result to obtain the mangrove spectral extraction result.

[0090] Furthermore, the initial extraction module also includes:

[0091] The verification sample point determination unit is used to obtain real points of mangroves based on on-site UAV remote sensing images, which are then used as verification sample points.

[0092] The first misclassification set determination unit is used to compare the verification sample points with the mangrove spectral extraction results to obtain the first misclassification set.

[0093] The first misclassification set includes misclassification points that classify non-mangrove forests as mangrove forests and omission points that classify mangrove forests as non-mangrove forests;

[0094] The re-extraction unit is used to adjust the index threshold and re-extract the mangrove spectral extraction result when the proportion of misclassified points in the first misclassification set exceeds the set proportion threshold.

[0095] Specifically, when adjusting the index thresholds, the values ​​of the vegetation index threshold and the location index threshold are increased, while the value of the water stress index threshold is decreased.

[0096] The omission distribution area determination unit is used to construct a buffer zone outward from the omission points in the first misclassification set to obtain the omission distribution area;

[0097] The missing region extraction unit is used to redetermine the index threshold based on the spectral characteristics of the missing distribution area, and use the redetermined index threshold to extract mangroves in the missing distribution area to obtain the missing region extraction result;

[0098] The initial mangrove region determination unit is used to take the union of the mangrove spectral extraction results and the missing region extraction results to obtain the initial mangrove region.

[0099] Furthermore, the boundary correction module includes:

[0100] A buffer unit is used to construct buffer zones at a set distance on both the inner and outer sides of the boundary of the initially extracted mangrove area to obtain the boundary area of ​​the initially extracted mangrove area.

[0101] The second cropping unit is used to crop the preprocessed GF-2 multispectral image based on the boundary vector of the initially extracted mangrove boundary region to obtain the boundary region to be corrected.

[0102] The first classification unit is used to train a random forest-based machine learning model using a selected training set containing positive and negative samples, and to extract mangroves in the boundary area to be corrected using the trained machine learning model, thereby obtaining the mangrove extraction result of the boundary area.

[0103] The second misclassification set determination unit is used to compare the verification sample points with the mangrove extraction results of the boundary area to obtain the second misclassification set.

[0104] The second misclassification set includes misclassification points that classify non-mangrove forests as mangrove forests and omission points that classify mangrove forests as non-mangrove forests;

[0105] The second classification unit is used to fine-tune the machine learning model trained on the training set by taking the misclassified points and omissions of the second misclassification set as negative samples and positive samples, respectively, and to extract the mangroves in the boundary area to be corrected by the fine-tuned machine learning model to obtain the mangrove boundary correction result.

[0106] The boundary correction unit is used to replace the initially extracted mangrove boundary region within the initially extracted mangrove region with the mangrove boundary correction result to obtain the corrected mangrove region.

[0107] Furthermore, the terrain correction module includes:

[0108] The topographic factor calculation unit is used to calculate the slope, relative elevation, and elevation standard deviation of each pixel in the study area based on DEM data and using geographic information system software.

[0109] The water index calculation unit is used to calculate the normalized water index of each pixel in the study area based on the preprocessed GF-2 multispectral image.

[0110]

[0111] Wherein, NDWI is the normalized water index, and Green and NIR represent the remote sensing reflectance of the green band and near-infrared band of the preprocessed GF-2 multispectral image, respectively.

[0112] The terrain correction index determination unit is used to construct the terrain correction index for each pixel within the study area using the following formula:

[0113]

[0114] The potential water accumulation area determination unit is used to compare the terrain correction index of each pixel in the study area with a determined segmentation threshold, and classify the pixels with the terrain correction index greater than the segmentation threshold as potential water accumulation areas.

[0115] The mangrove distribution area determination unit is used to remove the potential waterlogged areas from the modified mangrove area to obtain the final mangrove distribution area.

[0116] Furthermore, the weights of the slope, relative elevation, elevation standard deviation, and normalized water index are determined as follows:

[0117] The subjective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the analytic hierarchy process (AHP).

[0118] The objective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the CRITIC method.

[0119] The subjective and objective weights are linearly weighted to obtain the weights of slope, relative elevation, elevation standard deviation, and normalized water index.

[0120] Furthermore, the segmentation threshold is determined in the following manner:

[0121] The dataset of the terrain correction index was subjected to unsupervised clustering fitting using a Gaussian mixture model, resulting in two Gaussian distributions.

[0122] The value of the terrain correction index corresponding to the intersection of the probability density functions of the two Gaussian distributions is used as the segmentation threshold.

[0123] The present invention has the following beneficial effects:

[0124] The present invention aims to solve the problem of low extraction accuracy of mangroves in existing methods, especially the low extraction accuracy caused by the homonym of sparse mangroves in intertidal areas and foreign objects in water pixels.

[0125] This invention employs a progressive framework of "spectral extraction - geometric correction - terrain optimization," following the research approach of "compensating for weaknesses by leveraging strengths," and comprehensively utilizing the unique advantages of multi-source data to obtain progressively corrected, high-precision mangrove extraction results. First, Sentinel-2 multispectral imagery is used for rapid preliminary extraction of potential mangrove distribution areas, with error correction based on UAV data. Next, sub-meter-level GF-2 imagery is introduced to refine the geometric boundaries of the preliminary extraction results, with further error correction based on UAV data. Finally, based on digital elevation model (DEM) data, a comprehensive terrain correction index is constructed to accurately extract mangroves, correcting spectral confusion between sparse mangrove areas and water bodies and mangroves.

[0126] This invention overcomes the problem of low extraction accuracy of sparse mangroves in areas with small topographic relief caused by severe spectral confusion between mangroves and adjacent stagnant water areas. It solves the problem of low extraction accuracy of mangroves due to spectral confusion from the perspective of topographic and hydrological factors, reduces the impact of water bodies on the accurate identification of mangroves in sparse mangrove areas that even high spatial resolution imagery cannot achieve, alleviates the problem of spectral confusion caused by water interference, improves the extraction accuracy of sparse mangroves, and provides an effective technical solution for the accurate extraction and dynamic monitoring of coastal mangrove distribution.

[0127] Furthermore, this invention does not simply input multi-source data in parallel. Instead, it employs a progressive framework of "spectral extraction - geometric correction - terrain optimization," using a step-by-step optimization method to correct and improve each step in the mangrove extraction process. This fully leverages the advantages of each data type, maximizing its strengths and minimizing its weaknesses. UAV survey data is used for this step-by-step correction and improvement, clearly defining the functional roles and logical connections between the multi-source data. Medium-resolution imagery, high-resolution imagery, and digital elevation model (DEM) data sequentially utilize their core advantages: medium-resolution imagery handles rapid surveying, high-resolution imagery performs preliminary geometric boundary optimization, and the DEM, based on topographic and hydrological mechanisms, addresses spectral confusion that is difficult to resolve even with high-resolution imagery. This step-by-step, targeted correction logic ensures that the advantages of each data type are maximized, the correction process is clear and transparent, the correction effect is good, and the accuracy of mangrove extraction is improved. Attached Figure Description

[0128] Figure 1 The flowchart shows the mangrove extraction method based on multi-source data stepwise correction according to the present invention.

[0129] Figure 2 This is a schematic diagram of the mangrove extraction device based on multi-source data stepwise correction according to the present invention. Detailed Implementation

[0130] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0131] This invention provides a mangrove extraction method based on multi-source data with stepwise correction. This method aims to solve the problem that sparse mangroves in flat intertidal areas are difficult to accurately identify due to the severe spectral confusion between mangroves located in flat intertidal zones and adjacent stagnant water areas.

[0132] like Figure 1 As shown, the method includes:

[0133] S1: Obtain multi-source data for the same time period in the study area;

[0134] The multi-source data includes medium-resolution Sentinel-2 multispectral imagery, high-resolution GF-2 multispectral imagery, on-site UAV remote sensing imagery, and DEM (Digital Elevation Model) data.

[0135] Sentinel-2 and GF-2 multispectral images are large-scale remote sensing images covering the entire study area. However, due to the difficulty in acquiring on-site UAV remote sensing images, it is impossible to cover the entire study area. Therefore, several representative small areas (or points) were selected.

[0136] S2: Preprocessing Sentinel-2 multispectral imagery, GF-2 multispectral imagery, and on-site UAV remote sensing imagery.

[0137] Preprocessing includes radiometric correction, atmospheric correction, cloud removal, land-water separation, and image mosaicking, resulting in remote sensing images of different resolutions at the same time.

[0138] S3: Based on the spectral indices of the preprocessed Sentinel-2 multispectral image, the mangrove forest area is initially extracted, and error correction is performed based on the on-site UAV remote sensing image to obtain the initially extracted mangrove forest area.

[0139] This step is used for rapid preliminary extraction of mangrove data, and one implementation method includes:

[0140] S31: Use the boundary vector of the study area to crop the preprocessed Sentinel-2 multispectral image to obtain the medium-resolution image data of the study area.

[0141] For GF-2 multispectral imagery, on-site UAV remote sensing imagery, and DEM data, the boundary vector of the study area can be uniformly cropped to obtain data for the study area, which will then be used for subsequent data processing and analysis.

[0142] S32: Calculate the spectral index of each pixel in the study area based on the medium-resolution image data of the study area.

[0143] Based on the unique red-edge band of Sentinel-2 imagery, the spectral indices include one or more of the following: Red Edge Normalized Difference Vegetation Index (RE_NDVI), Red Edge Position Index (REP), and Moisture Stress Index (MSI). The calculation formulas for each spectral index are as follows:

[0144]

[0145]

[0146]

[0147] RE_NDVI is the red-edge normalized vegetation index, REP is the red-edge position index, MSI is the water stress index, and B4, B5, B7, B8A, and B11 represent the remote sensing reflectance of bands 4, 5, 7, 8A, and 11 of the preprocessed Sentinel-2 multispectral image, respectively.

[0148] S33: Compare each spectral index with the set index thresholds to obtain the mangrove spectral extraction results.

[0149] After calculating each spectral index, this invention uses a threshold limitation method to quickly extract the mangrove range and obtain the mangrove spectral extraction result, denoted as A1.

[0150] Specifically, including:

[0151] S331: For each pixel in the study area, the red-edge normalized vegetation index of the pixel is compared with the set vegetation index threshold. Pixels with a red-edge normalized vegetation index greater than the vegetation index threshold are classified as mangroves, and the first spectral extraction result is obtained.

[0152] The higher the red border normalized vegetation index, the more lush the mangroves are. Therefore, areas with a vegetation index value greater than the threshold are classified as mangroves.

[0153] S332: For each pixel in the study area, the red edge position index of the pixel is compared with the set position index threshold. Pixels with a red edge position index greater than the position index threshold are classified as mangroves, and the second spectral extraction result is obtained.

[0154] The red edge location index represents the growth of mangroves. The larger the value, the more vigorous the mangrove growth. Therefore, areas with a location index value greater than the threshold are classified as mangroves.

[0155] S333: For each pixel in the study area, the water stress index of the pixel is compared with the set water stress index threshold. Pixels with a water stress index less than the water stress index threshold are classified as mangroves, and the third spectral extraction results are obtained.

[0156] The water stress index characterizes the water content of mangrove leaves. The healthier the mangrove growth, the higher the water content of its leaves, the stronger the absorption, and the lower the water stress index. Therefore, areas with water stress index values ​​below the threshold are classified as mangroves.

[0157] S334: Take the union of the first spectral extraction result, the second initial spectral extraction result, and the third spectral extraction result to obtain the mangrove spectral extraction result, i.e., A1.

[0158] This invention extracts mangrove data using multiple spectral indices, and the combined results from multiple spectral indices are more accurate.

[0159] After obtaining the spectral extraction results of mangroves, error correction can be performed based on in-situ UAV remote sensing imagery, as follows:

[0160] S34: Obtain real points of mangroves from on-site drone remote sensing images and use them as verification sample points.

[0161] The true locations of mangroves can be obtained through visual interpretation of on-site drone remote sensing images. Since on-site drone remote sensing images represent several small, representative areas, the true locations of mangroves obtained are a limited number of points, which can be used as verification sample points.

[0162] S35: Compare the validation sample points with the mangrove spectral extraction results to obtain the first misclassification set.

[0163] The first misclassification set includes misclassification points that classify non-mangrove forests as mangrove forests and omission points that classify mangrove forests as non-mangrove forests.

[0164] The first error classification set uses validation sample points from in-situ UAV remote sensing imagery to verify the accuracy of the mangrove spectral extraction result A1, identifying the errors in the extraction results from the medium-resolution Sentinel-2 imagery and statistically analyzing these errors. The first error classification set is denoted as E1, and the misclassified points it includes are denoted as E... 11 The point of failure is recorded as E. 12 .

[0165] S36: When the proportion of misclassified points in the first misclassification set exceeds the set proportion threshold, adjust the index threshold and re-extract the mangrove spectral extraction results through S33.

[0166] When adjusting the index threshold, the spectral index can be recalculated based on the spectral characteristics of the misclassified points to determine the index threshold.

[0167] Alternatively, when adjusting the index thresholds, the values ​​of the vegetation index and location index thresholds can be increased, while the value of the water stress index threshold can be decreased. This is because a higher red-edge normalized vegetation index and red-edge location index strongly indicate mangrove forests; therefore, to reduce misclassification, the vegetation index and location index thresholds need to be increased. Conversely, a lower water stress index value strongly indicates mangrove forests, so the water stress index threshold needs to be decreased.

[0168] The mangrove spectral extraction result obtained by re-extraction is denoted as A. 11 A 11 This is the result of re-extracting the Sentinel-2 image, which replaces the previous mangrove spectral extraction result A1.

[0169] S37: Construct a buffer zone outward from the omission points in the first misclassification set to obtain the omission distribution area.

[0170] Specifically, the missing point E can be identified. 12 The pixel boundaries are buffered at 10 times the image resolution to obtain the missing distribution area.

[0171] S38: Based on the spectral characteristics of the missed distribution area, the index threshold is re-determined, and the re-determined index threshold is used to extract mangroves in the missed distribution area, obtaining the missed area extraction result, denoted as A. 12 .

[0172] A 12 This is the extraction result of the part that was missed in the mangrove spectral extraction result A1, and it is a supplement to the part that was missed in A1.

[0173] S39: Take the union of the mangrove spectral extraction results and the missing region extraction results to obtain the initially extracted mangrove region.

[0174] The mangrove spectral extraction results mentioned in this step refer to the aforementioned A1 or A... 11 If S36 is executed, it refers to A. 11 Otherwise, refer to A1. (The last part, "A," appears to be a typo and can be left as is.) 11 Or A1 and A 12 The initial extracted mangrove region is denoted as A by taking the union of the sets. 13 .

[0175] S4: Based on the preprocessed GF-2 multispectral image, machine learning methods are used to correct the boundaries of the initially extracted mangrove area, and error correction is performed based on the on-site UAV remote sensing image to obtain the corrected mangrove area.

[0176] This step is used to refine and complete the initially extracted mangrove boundaries using high-resolution GF-2 multispectral imagery. In one example, this step is implemented as follows:

[0177] S41: Based on the boundary of the initially extracted mangrove area, a buffer zone with a set distance is constructed on both the inner and outer sides of the boundary to obtain the boundary area of ​​the initially extracted mangrove area.

[0178] The buffer distance D can be determined based on the image resolution or other possible factors; for example, D can be 10 pixels. Initially, the mangrove region A is extracted. 13 After constructing a buffer zone at a distance D on both sides of the boundary, the initial extracted mangrove boundary region is denoted as B.

[0179] S42: Based on the initial extraction of the boundary vector of the mangrove boundary region B, the preprocessed GF-2 multispectral image is cropped to obtain the boundary region to be corrected.

[0180] S43: The random forest-based machine learning model is trained using a selected training set containing positive and negative samples, and the mangrove forests in the boundary area to be corrected are extracted using the trained machine learning model to obtain the mangrove forest extraction result of the boundary area.

[0181] The machine learning method of this invention can employ a random forest classifier. First, a training set containing sufficient positive and negative samples is selected to train the random forest model. Then, the boundary region to be corrected is input into the trained random forest model for classification, resulting in the extraction result of mangrove forests in the boundary region based on medium-resolution imagery, denoted as A2.

[0182] S44: Compare the verification sample points with the mangrove extraction results in the boundary area to obtain the second misclassification set.

[0183] Similar to the first misclassification set, the second misclassification set also includes misclassifications of non-mangrove forests as mangrove forests and omissions of mangrove forests as non-mangrove forests.

[0184] The second misclassification set is used to verify the accuracy of the mangrove extraction result A2 in the boundary area using validation sample points from in-situ UAV remote sensing imagery, and to statistically analyze the extraction results from the GF-2 imagery. The second misclassification set is denoted as E2, and the misclassified points it includes are denoted as E... 21 The point of failure is recorded as E. 22 .

[0185] S45: The misclassified and missed classification points in the second misclassification set are used as negative and positive samples, respectively, to fine-tune the machine learning model trained on the training set. The fine-tuned machine learning model is then used to extract mangroves within the boundary area to be corrected, resulting in the mangrove boundary correction result, denoted as A. 21 .

[0186] S46: Replace the initially extracted mangrove boundary region within the initially extracted mangrove region with the mangrove boundary correction result to obtain the corrected mangrove region.

[0187] Specifically, through Boolean operations, using A 21 Replace A 13 In the corresponding part within the boundary region B, fill in A. 13 Missing patches in the middle, and correct A 13 The boundary error is used to obtain the corrected mangrove area, denoted as A3.

[0188] S5: Based on DEM data and preprocessed GF-2 multispectral images, a topographic correction index is constructed, and potential water accumulation areas are removed from the corrected mangrove areas according to the topographic correction index to obtain the final mangrove distribution area.

[0189] This step is used to collaboratively identify and remove water bodies in sparse mangrove areas using GF-2 imagery and DEM data. The specific steps are as follows:

[0190] S51: Based on DEM data, calculate the slope, relative elevation, and elevation standard deviation of each pixel in the study area using geographic information system software.

[0191] This step is used to construct terrain correction factors such as slope, relative elevation, and elevation standard deviation. These factors can be calculated using the terrain analysis tools built into ArcGIS.

[0192] S52: Calculate the Normalized Difference Water Index (NDWI) for each pixel in the study area based on the preprocessed GF-2 multispectral image.

[0193]

[0194] Wherein, NDWI is the normalized water index, and Green and NIR represent the remote sensing reflectance of the green band and near-infrared band of the preprocessed GF-2 multispectral image, respectively.

[0195] S53: Construct the Potential Waterlogging Index (PWI) for each pixel within the study area using the following formula:

[0196]

[0197] in, S represents the terrain correction index, and RE and ESD represent the slope, relative elevation, and standard deviation of elevation obtained from DEM data, respectively. The normalized water index is calculated based on the preprocessed GF-2 multispectral image, where a, b, c, and d are the weights of slope, relative elevation, elevation standard deviation, and normalized water index, respectively.

[0198] Weights a, b, c, and d are assigned through a comprehensive weighting strategy, specifically determined as follows:

[0199] 1. The Analytic Hierarchy Process (AHP) was used to determine the subjective weights of slope, relative elevation, elevation standard deviation, and normalized water index that reflect expert experience.

[0200] 2. The CRITIC method is used to determine the objective weights of slope, relative elevation, elevation standard deviation, and normalized water index to reflect the inherent laws of the data.

[0201] 3. The subjective and objective weights are linearly weighted to obtain the weights of slope, relative elevation, elevation standard deviation and normalized water index.

[0202] The weights for linearly weighting subjective and objective weights are set according to needs or experience; for example, the weights for subjective and objective weights are each set to 0.5.

[0203] The constructed topographic correction index includes both topographic data (slope, relative elevation, elevation standard deviation) and water index. This topographic correction index comprehensively reflects the potential impact of micro-topography on water accumulation and can represent absolute low-lying areas and water-bearing areas in the study area, i.e. potential water accumulation areas. These water accumulation areas are not areas where mangroves grow and need to be excluded. This solves the problem of spectral confusion between mangroves and water bodies from the perspective of topographic and hydrological causes.

[0204] After constructing the terrain correction index, it is necessary to calculate the terrain correction index values ​​of all pixels in the study area to form a dataset.

[0205] S54: Compare the terrain correction index of each pixel in the study area with the determined segmentation threshold, and classify the pixels with terrain correction index greater than the segmentation threshold as potential water accumulation areas.

[0206] The segmentation threshold can be determined using a Gaussian mixture model. First, an unsupervised clustering fit is performed on the terrain correction index dataset using a Gaussian mixture model to obtain two Gaussian distributions. Then, the value of the terrain correction index corresponding to the intersection of the probability density functions of the two Gaussian distributions is used as the segmentation threshold.

[0207] The probability density function of a Gaussian mixture model is a weighted sum of multiple Gaussian distribution components, and its mathematical expression is as follows:

[0208]

[0209] Where x represents a data point, and in this invention, it represents the terrain correction index. K is the number of Gaussian distributions (also known as components). In this invention, K=2, meaning that this invention distinguishes between two types of areas: potentially flooded areas and non-flooded areas. It is the mixture weight of the k-th (k=1,2) Gaussian distribution, satisfying ; It is the k-th Gaussian distribution with a mean of Covariance Matrix .

[0210] The specific operation process is as follows:

[0211] Model initialization: The number of Gaussian distributions is preset to K. In this invention, to distinguish between potentially flooded areas and non-flooded areas, K=2 is set.

[0212] Model Training: The Gaussian mixture model is trained using the expectation-maximization algorithm. This process involves iteratively performing the following steps until the parameters of the Gaussian mixture model converge.

[0213] Expectation step: Based on the current parameters, calculate the probability that each PWI value data point belongs to each Gaussian distribution component.

[0214] Maximization step: Based on the probabilities calculated in the previous step, re-estimate the parameters of each Gaussian distribution component, including its mean, variance, and mixing weights.

[0215] Threshold Determination: After training, the Gaussian mixture model fits the overall distribution of PWI values ​​into two Gaussian distributions, representing the "potentially flooded area" and the "non-flooded area," respectively. The segmentation threshold is automatically determined as the PWI value corresponding to the intersection of the probability density functions of these two Gaussian distributions.

[0216] After determining the segmentation threshold, pixels with PWI values ​​greater than the threshold are identified as "potential water accumulation areas" and output as result A4.

[0217] S55: Remove the potential waterlogged area A4 from the corrected mangrove area A3 to achieve further optimization and obtain the final mangrove distribution area, thus achieving accurate extraction of mangroves.

[0218] The present invention aims to solve the problem of low extraction accuracy of mangroves in existing methods, especially the low extraction accuracy caused by the homonym of sparse mangroves in intertidal areas and foreign objects in water pixels.

[0219] This invention employs a progressive framework of "spectral extraction - geometric correction - terrain optimization," following the research approach of "compensating for weaknesses by leveraging strengths," and comprehensively utilizing the unique advantages of multi-source data to obtain progressively corrected, high-precision mangrove extraction results. First, Sentinel-2 multispectral imagery is used for rapid preliminary extraction of potential mangrove distribution areas, with error correction based on UAV data. Next, sub-meter-level GF-2 imagery is introduced to refine the geometric boundaries of the preliminary extraction results, with further error correction based on UAV data. Finally, based on digital elevation model (DEM) data, a comprehensive terrain correction index is constructed to accurately extract mangroves, correcting spectral confusion between sparse mangrove areas and water bodies and mangroves.

[0220] This invention overcomes the problem of low extraction accuracy of sparse mangroves in areas with small topographic relief caused by severe spectral confusion between mangroves and adjacent stagnant water areas. It solves the problem of low extraction accuracy of mangroves due to spectral confusion from the perspective of topographic and hydrological factors, reduces the impact of water bodies on the accurate identification of mangroves in sparse mangrove areas that even high spatial resolution imagery cannot achieve, alleviates the problem of spectral confusion caused by water interference, improves the extraction accuracy of sparse mangroves, and provides an effective technical solution for the accurate extraction and dynamic monitoring of coastal mangrove distribution.

[0221] Furthermore, this invention does not simply input multi-source data in parallel. Instead, it employs a progressive framework of "spectral extraction - geometric correction - terrain optimization," using a step-by-step optimization method to correct and improve each step in the mangrove extraction process. This fully leverages the advantages of each data type, maximizing its strengths and minimizing its weaknesses. UAV survey data is used for this step-by-step correction and improvement, clearly defining the functional roles and logical connections between the multi-source data. Medium-resolution imagery, high-resolution imagery, and digital elevation model (DEM) data sequentially utilize their core advantages: medium-resolution imagery handles rapid surveying, high-resolution imagery performs preliminary geometric boundary optimization, and the DEM, based on topographic and hydrological mechanisms, addresses spectral confusion that is difficult to resolve even with high-resolution imagery. This step-by-step, targeted correction logic ensures that the advantages of each data type are maximized, the correction process is clear and transparent, the correction effect is good, and the accuracy of mangrove extraction is improved.

[0222] This invention also provides a mangrove extraction device based on multi-source data step-by-step correction, such as... Figure 2 As shown, the device includes:

[0223] Data acquisition module 1 is used to acquire multi-source data of the same time period in the study area.

[0224] The multi-source data includes medium-resolution Sentinel-2 multispectral imagery, high-resolution GF-2 multispectral imagery, on-site UAV remote sensing imagery, and DEM data.

[0225] Preprocessing module 2 is used to preprocess Sentinel-2 multispectral images, GF-2 multispectral images, and on-site UAV remote sensing images.

[0226] The initial extraction module 3 is used to initially extract mangroves based on the spectral indices of the preprocessed Sentinel-2 multispectral image, and to correct errors based on the on-site UAV remote sensing image to obtain the initially extracted mangrove area.

[0227] Boundary correction module 4 is used to correct the boundaries of the initially extracted mangrove area based on the preprocessed GF-2 multispectral image using machine learning methods, and to correct errors based on the on-site UAV remote sensing image to obtain the corrected mangrove area.

[0228] The terrain correction module 5 is used to construct a terrain correction index based on DEM data and preprocessed GF-2 multispectral imagery, and remove potential water accumulation areas from the corrected mangrove areas according to the terrain correction index to obtain the final mangrove distribution area.

[0229] The formula for calculating the terrain correction index is as follows:

[0230]

[0231] S represents the terrain correction index, and RE and ESD represent the slope, relative elevation, and standard deviation of elevation obtained from DEM data, respectively. The normalized water index is calculated based on the preprocessed GF-2 multispectral image, where a, b, c, and d are the weights of slope, relative elevation, elevation standard deviation, and normalized water index, respectively.

[0232] As an example, the initial extraction module includes:

[0233] The first cropping unit is used to crop the preprocessed Sentinel-2 multispectral image using the boundary vector of the study area to obtain medium-resolution image data of the study area.

[0234] The spectral index calculation unit is used to calculate the spectral index of each pixel in the study area based on the medium-resolution image data of the study area.

[0235] Among them, the spectral indices include one or more of the following: red-edge normalized vegetation index, red-edge location index, and water stress index, and the calculation formula is as follows:

[0236]

[0237]

[0238]

[0239] RE_NDVI is the red-edge normalized vegetation index, REP is the red-edge position index, MSI is the water stress index, and B4, B5, B7, B8A, and B11 represent the remote sensing reflectance of bands 4, 5, 7, 8A, and 11 of the preprocessed Sentinel-2 multispectral image, respectively.

[0240] The spectral index segmentation unit is used to compare each spectral index with the set index thresholds to obtain the mangrove spectral extraction results.

[0241] Specifically, the spectral index segmentation unit includes:

[0242] The first segmentation subunit is used to compare the red-edge normalized vegetation index of each pixel in the study area with the set vegetation index threshold. Pixels with a red-edge normalized vegetation index greater than the vegetation index threshold are classified as mangroves, thus obtaining the first spectral extraction result.

[0243] The second segmentation subunit is used to compare the red edge position index of each pixel in the study area with the set position index threshold. Pixels with a red edge position index greater than the position index threshold are classified as mangroves, thus obtaining the second spectral extraction result.

[0244] The third segmentation subunit is used to compare the water stress index of each pixel in the study area with the set water stress index threshold. Pixels with a water stress index less than the water stress index threshold are classified as mangroves, thus obtaining the third spectral extraction result.

[0245] The mangrove spectral extraction result determination subunit is used to take the union of the first, second, and third spectral extraction results to obtain the mangrove spectral extraction result.

[0246] The initial extraction module also includes:

[0247] The verification sample point determination unit is used to obtain real points of mangroves based on on-site UAV remote sensing images, which are then used as verification sample points.

[0248] The first misclassification set determination unit is used to compare the verification sample points with the mangrove spectral extraction results to obtain the first misclassification set.

[0249] The first misclassification set includes misclassification points that classify non-mangrove forests as mangrove forests and omission points that classify mangrove forests as non-mangrove forests.

[0250] The re-extraction unit is used to adjust the index threshold and re-extract the mangrove spectral extraction result when the proportion of misclassified points in the first misclassification set exceeds the set proportion threshold.

[0251] Specifically, when adjusting the index thresholds, the values ​​of the vegetation index threshold and the location index threshold are increased, while the value of the water stress index threshold is decreased.

[0252] The omission distribution area determination unit is used to construct a buffer zone outward from the omission points in the first misclassification set to obtain the omission distribution area.

[0253] The missing region extraction unit is used to redetermine the index threshold based on the spectral characteristics of the missing distribution area, and use the redetermined index threshold to extract mangroves from the missing distribution area to obtain the missing region extraction result.

[0254] The initial mangrove region determination unit is used to take the union of the mangrove spectral extraction results and the missing region extraction results to obtain the initial mangrove region.

[0255] As one implementation, the boundary correction module includes:

[0256] The buffer unit is used to construct buffer zones at a set distance on both the inner and outer sides of the initially extracted mangrove area, based on the boundary of the initially extracted mangrove area, to obtain the boundary area of ​​the initially extracted mangrove area.

[0257] The second cropping unit is used to crop the preprocessed GF-2 multispectral image based on the boundary vector of the initially extracted mangrove boundary region to obtain the boundary region to be corrected.

[0258] The first classification unit is used to train a random forest-based machine learning model using a selected training set containing positive and negative samples, and to extract mangroves in the boundary area to be corrected using the trained machine learning model, thereby obtaining the mangrove extraction result of the boundary area.

[0259] The second misclassification set determination unit is used to compare the verification sample points with the mangrove extraction results of the boundary area to obtain the second misclassification set.

[0260] The second misclassification set includes misclassification points that classify non-mangrove forests as mangrove forests and omission points that classify mangrove forests as non-mangrove forests.

[0261] The second classification unit is used to fine-tune the machine learning model trained on the training set by taking the misclassified and omitted points of the second misclassification set as negative and positive samples, respectively, and to extract the mangroves in the boundary area to be corrected by the fine-tuned machine learning model to obtain the mangrove boundary correction result.

[0262] The boundary correction unit is used to replace the initially extracted mangrove boundary region within the initially extracted mangrove region with the mangrove boundary correction result, thus obtaining the corrected mangrove region.

[0263] As an improvement to this embodiment of the invention, the terrain correction module includes:

[0264] The topographic factor calculation unit is used to calculate the slope, relative elevation, and elevation standard deviation of each pixel in the study area based on DEM data and using geographic information system software.

[0265] The water index calculation unit is used to calculate the normalized water index of each pixel in the study area based on the preprocessed GF-2 multispectral image.

[0266]

[0267] Wherein, NDWI is the normalized water index, and Green and NIR represent the remote sensing reflectance of the green band and near-infrared band of the preprocessed GF-2 multispectral image, respectively.

[0268] The terrain correction index determination unit is used to construct the terrain correction index for each pixel within the study area using the following formula:

[0269]

[0270] The potential water accumulation area determination unit is used to compare the terrain correction index of each pixel in the study area with a determined segmentation threshold, and classify pixels with terrain correction indices greater than the segmentation threshold as potential water accumulation areas.

[0271] The mangrove distribution area determination unit is used to remove potential waterlogged areas from the modified mangrove area to obtain the final mangrove distribution area.

[0272] The weights of the aforementioned slope, relative elevation, elevation standard deviation, and normalized water index can be determined as follows:

[0273] 1. The subjective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the analytic hierarchy process.

[0274] 2. The objective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the CRITIC method.

[0275] 3. The subjective and objective weights are linearly weighted to obtain the weights of slope, relative elevation, elevation standard deviation and normalized water index.

[0276] In this invention, the segmentation threshold can be determined in the following way:

[0277] An unsupervised clustering fit was performed on the terrain correction index dataset using a Gaussian mixture model, resulting in two Gaussian distributions.

[0278] The value of the terrain correction index corresponding to the intersection of the probability density functions of the two Gaussian distributions is used as the segmentation threshold.

[0279] The apparatus provided in this embodiment of the invention operates on the same principle and produces the same technical effects as the aforementioned method embodiments. For the sake of brevity, any parts not mentioned in the apparatus embodiments can be referred to the corresponding content in the aforementioned method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the apparatus and units described above can all be referred to the corresponding processes in the aforementioned method embodiments, and will not be repeated here.

[0280] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A method for extracting mangrove forests based on multi-source data with step-by-step correction, characterized in that, The method includes: S1: Obtain multi-source data for the same time period in the study area; The multi-source data includes medium-resolution Sentinel-2 multispectral imagery, high-resolution GF-2 multispectral imagery, on-site UAV remote sensing imagery, and DEM data. S2: Preprocess the Sentinel-2 multispectral image, GF-2 multispectral image, and on-site UAV remote sensing image; S3: Based on the spectral indices of the preprocessed Sentinel-2 multispectral image, the mangrove forest area is initially extracted, and the error is corrected based on the on-site UAV remote sensing image to obtain the initially extracted mangrove forest area. S4: Based on the preprocessed GF-2 multispectral image, machine learning methods are used to correct the boundaries of the initially extracted mangrove area, and error correction is performed based on the on-site UAV remote sensing image to obtain the corrected mangrove area. S5: Construct a topographic correction index based on DEM data and preprocessed GF-2 multispectral image, and remove potential water accumulation areas from the corrected mangrove area according to the topographic correction index to obtain the final mangrove distribution area; The formula for calculating the terrain correction index is as follows: S represents the terrain correction index, and RE and ESD represent the slope, relative elevation, and standard deviation of elevation obtained from the DEM data, respectively. The normalized water index is calculated based on the preprocessed GF-2 multispectral image, where a, b, c, and d are the weights of slope, relative elevation, elevation standard deviation, and normalized water index, respectively. S5 includes: S51: Based on DEM data, calculate the slope, relative elevation, and elevation standard deviation of each pixel in the study area using geographic information system software; S52: Calculate the normalized water index for each pixel in the study area based on the preprocessed GF-2 multispectral image; Wherein, NDWI is the normalized water index, and Green and NIR represent the remote sensing reflectance of the green band and near-infrared band of the preprocessed GF-2 multispectral image, respectively. S53: The terrain correction index for each pixel within the study area is constructed using the following formula: S54: Compare the terrain correction index of each pixel in the study area with the determined segmentation threshold, and classify the pixels with terrain correction index greater than the segmentation threshold as potential water accumulation areas. S55: Remove the potential waterlogged area from the modified mangrove area to obtain the final mangrove distribution area; The weights of the slope, relative elevation, elevation standard deviation, and normalized water index are determined as follows: The subjective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the analytic hierarchy process (AHP). The objective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the CRITIC method. The subjective and objective weights are linearly weighted to obtain the weights of slope, relative elevation, elevation standard deviation, and normalized water index. The segmentation threshold is determined in the following manner: The dataset of the terrain correction index was subjected to unsupervised clustering fitting using a Gaussian mixture model, resulting in two Gaussian distributions. The value of the terrain correction index corresponding to the intersection of the probability density functions of the two Gaussian distributions is used as the segmentation threshold.

2. The mangrove extraction method based on multi-source data stepwise correction according to claim 1, characterized in that, S3 includes: S31: Use the boundary vector of the study area to crop the preprocessed Sentinel-2 multispectral image to obtain the medium-resolution image data of the study area; S32: Calculate the spectral index of each pixel in the study area based on the medium-resolution image data of the study area; The spectral indices include one or more of the following: red-edge normalized vegetation index, red-edge location index, and water stress index, and the calculation formula is as follows: RE_NDVI is the red-edge normalized vegetation index, REP is the red-edge position index, MSI is the water stress index, and B4, B5, B7, B8A, and B11 represent the remote sensing reflectance of bands 4, 5, 7, 8A, and 11 of the preprocessed Sentinel-2 multispectral image, respectively. S33: Compare each spectral index with the set index thresholds to obtain the mangrove spectral extraction results.

3. The mangrove extraction method based on multi-source data stepwise correction according to claim 2, characterized in that, S33 includes: S331: For each pixel in the study area, the red-edge normalized vegetation index of the pixel is compared with the set vegetation index threshold. Pixels with a red-edge normalized vegetation index greater than the vegetation index threshold are classified as mangroves, and the first spectral extraction result is obtained. S332: For each pixel in the study area, compare the red edge position index of the pixel with the set position index threshold, and classify the pixels with the red edge position index greater than the position index threshold as mangroves to obtain the second spectral extraction result; S333: For each pixel in the study area, the water stress index of the pixel is compared with the set water stress index threshold. Pixels with a water stress index less than the water stress index threshold are classified as mangroves, and the third spectral extraction result is obtained. S334: Take the union of the first spectral extraction result, the second spectral extraction result, and the third spectral extraction result to obtain the mangrove spectral extraction result.

4. The mangrove forest extraction method based on multi-source data stepwise correction according to claim 3, characterized in that, S3 further includes: S34: Obtain real points of mangrove forests based on on-site drone remote sensing images, and use them as verification sample points; S35: Compare the verification sample points with the mangrove spectral extraction results to obtain the first misclassification set; The first misclassification set includes misclassification points that classify non-mangrove forests as mangrove forests and omission points that classify mangrove forests as non-mangrove forests; S36: When the proportion of misclassified points in the first misclassification set exceeds the set proportion threshold, adjust the index threshold and re-extract the mangrove spectral extraction results; Specifically, when adjusting the index thresholds, the values ​​of the vegetation index threshold and the location index threshold are increased, while the value of the water stress index threshold is decreased. S37: Construct a buffer zone outward from the omission points in the first misclassification set to obtain the omission distribution area; S38: Based on the spectral characteristics of the underscored distribution area, redetermine the index threshold, and use the redetermined index threshold to extract mangroves in the underscored distribution area to obtain the underscored area extraction result; S39: Take the union of the mangrove spectral extraction results and the missing region extraction results to obtain the initially extracted mangrove region.

5. The mangrove forest extraction method based on multi-source data stepwise correction according to claim 4, characterized in that, S4 includes: S41: Based on the boundary of the initially extracted mangrove area, a buffer zone with a set distance is constructed on both the inner and outer sides of the boundary to obtain the boundary area of ​​the initially extracted mangrove area; S42: Based on the boundary vector of the initially extracted mangrove boundary region, the preprocessed GF-2 multispectral image is cropped to obtain the boundary region to be corrected. S43: The random forest-based machine learning model is trained using a selected training set containing positive and negative samples, and the mangrove forests in the boundary area to be corrected are extracted using the trained machine learning model to obtain the mangrove forest extraction result in the boundary area. S44: Compare the verification sample points with the mangrove extraction results of the boundary area to obtain the second misclassification set; The second misclassification set includes misclassification points that classify non-mangrove forests as mangrove forests and omission points that classify mangrove forests as non-mangrove forests; S45: The misclassified points and omissions in the second misclassification set are used as negative samples and positive samples, respectively, to fine-tune the machine learning model trained on the training set. The fine-tuned machine learning model is then used to extract the mangroves in the boundary area to be corrected, and the mangrove boundary correction result is obtained. S46: Replace the initially extracted mangrove boundary region within the initially extracted mangrove region with the mangrove boundary correction result to obtain the corrected mangrove region.

6. A mangrove forest extraction device based on multi-source data with step-by-step correction, characterized in that, The device includes: The data acquisition module is used to acquire multi-source data of the same time period in the study area; The multi-source data includes medium-resolution Sentinel-2 multispectral imagery, high-resolution GF-2 multispectral imagery, on-site UAV remote sensing imagery, and DEM data. The preprocessing module is used to preprocess the Sentinel-2 multispectral image, GF-2 multispectral image, and on-site UAV remote sensing image. The initial extraction module is used to perform preliminary extraction of mangroves based on the spectral indices of the preprocessed Sentinel-2 multispectral image, and to perform error correction based on the on-site UAV remote sensing image to obtain the initially extracted mangrove area. The boundary correction module is used to correct the boundaries of the initially extracted mangrove area based on the preprocessed GF-2 multispectral image using machine learning methods, and to correct errors based on the on-site UAV remote sensing image to obtain the corrected mangrove area. The terrain correction module is used to construct a terrain correction index based on DEM data and preprocessed GF-2 multispectral imagery, and remove potential water accumulation areas from the corrected mangrove area according to the terrain correction index to obtain the final mangrove distribution area. The formula for calculating the terrain correction index is as follows: S represents the terrain correction index, and RE and ESD represent the slope, relative elevation, and standard deviation of elevation obtained from the DEM data, respectively. The normalized water index is calculated based on the preprocessed GF-2 multispectral image, where a, b, c, and d are the weights of slope, relative elevation, elevation standard deviation, and normalized water index, respectively. The terrain correction module includes: The topographic factor calculation unit is used to calculate the slope, relative elevation, and elevation standard deviation of each pixel in the study area based on DEM data and using geographic information system software. The water index calculation unit is used to calculate the normalized water index of each pixel in the study area based on the preprocessed GF-2 multispectral image. Wherein, NDWI is the normalized water index, and Green and NIR represent the remote sensing reflectance of the green band and near-infrared band of the preprocessed GF-2 multispectral image, respectively. The terrain correction index determination unit is used to construct the terrain correction index for each pixel within the study area using the following formula: The potential water accumulation area determination unit is used to compare the terrain correction index of each pixel in the study area with a determined segmentation threshold, and classify the pixels with the terrain correction index greater than the segmentation threshold as potential water accumulation areas. The mangrove distribution area determination unit is used to remove the potential water accumulation area from the modified mangrove area to obtain the final mangrove distribution area. The weights of the slope, relative elevation, elevation standard deviation, and normalized water index are determined as follows: The subjective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the analytic hierarchy process (AHP). The objective weights of slope, relative elevation, elevation standard deviation, and normalized water index were determined using the CRITIC method. The subjective and objective weights are linearly weighted to obtain the weights of slope, relative elevation, elevation standard deviation, and normalized water index. The segmentation threshold is determined in the following manner: The dataset of the terrain correction index was subjected to unsupervised clustering fitting using a Gaussian mixture model, resulting in two Gaussian distributions. The value of the terrain correction index corresponding to the intersection of the probability density functions of the two Gaussian distributions is used as the segmentation threshold.

Images