Automatic classification method and device of coastal ecosystem based on multi-source remote sensing fusion

Abstract

The invention provides a method and device for automatically classifying coastal ecosystems based on multi-source remote sensing fusion, and relates to the technical field of coastal wetland classification. The invention obtains Landsat satellite image data, MODIS image data and a multi-layer coastal wetland map of a target area, and uses GEE to filter the cloud layer and cloud threshold of the Landsat satellite image data using the NSPI algorithm to obtain cloud-free and clear Landsat satellite image data; adopts a regression-based spatiotemporal fusion model to fuse and reconstruct the filtered MODIS image data to obtain fused image data with less cloud and uniform distribution over time; extracts wetland vector data from the multi-layer coastal wetland map by layer according to the wetland type; based on GEE, randomly extracts a number of random points, calculates the mean and standard deviation, and reconstructs the spectral index curve using the time linear interpolation method; then identifies the tidal flat using the tide level and vegetation-free characteristics, and then distinguishes mangroves and salt marshes according to phenological characteristics, thereby obtaining the final coastal wetland classification result.

Description

Technical Field

[0001] The present invention relates to the technical field of coastal wetland classification, and in particular to a method and device for automatic classification of coastal ecosystems based on multi-source remote sensing fusion. Background Art

[0002] Composed of mangroves, salt marshes and unvegetated tidal flats, coastal wetlands play a unique role in connecting terrestrial and marine environments, providing important ecosystem services such as coastal protection, fisheries enhancement and carbon sequestration.

[0003] At present, coastal wetland monitoring mainly relies on remote sensing image interpretation, which provides an economical and efficient means for continuously assessing wetland composition over large areas and long periods of time. However, due to the highly dynamic variability of the intertidal environment, remote sensing detection methods for coastal wetland composition are more challenging than those for terrestrial landscapes. The acquisition of remote sensing data for coastal wetlands is greatly affected by tides. Therefore, previous studies usually use low tide images to capture the exposure moments of wetlands and assess the entire spatial range. On the other hand, coastal areas, especially in tropical and subtropical areas, are often covered by clouds, which reduces the possibility of obtaining clear low tide images. Even when using cloud-free low tide images, the spectral characteristics of coastal wetlands and terrestrial landscapes (such as terrestrial vegetation and bare soil) are relatively similar, resulting in inaccurate classification.

[0004] Therefore, exploring effective and generally applicable algorithms to solve the problems of tidal fluctuations, cloud pollution and spectral similarity is crucial for the detection of coastal wetland components. In view of this, the applicant has proposed this application after studying the existing technologies. Summary of the invention

[0005] The present invention aims to provide a method, device, equipment and medium for automatic classification of coastal ecosystems based on multi-source remote sensing fusion, so as to solve the shortcomings of existing methods that lead to inaccurate classification of coastal wetlands due to problems such as tidal fluctuations, cloud pollution and spectral similarity.

[0006] In order to solve the above technical problems, the present invention is implemented through the following technical solutions:

[0007] An automatic classification method for coastal ecosystems based on multi-source remote sensing fusion, including:

[0008] S1, obtain Landsat satellite image data, MODIS image data and multi-layer coastal wetland maps of the target area;

[0009] S2, using Google Earth Engine, performing NSPI algorithm cloud filtering and cloud threshold screening on the Landsat satellite image data to obtain cloud-free and clear Landsat satellite image data;

[0010] S3, using a regression-based spatiotemporal fusion model to fuse and reconstruct the cloud-free and clear Landsat satellite image data with the MODIS image data filtered by a cloud threshold, thereby obtaining fused image data with less cloud and uniform distribution over time;

[0011] S4, extracting wetland vector data from the multi-layer coastal wetland map in layers according to wetland type; randomly extracting a number of random points from the selected wetland type vector data based on Google Earth Engine, and obtaining the corresponding spectral index threshold by calculating the mean and standard deviation of the spectral index of the random points; and then reconstructing the spectral index curve using a time linear interpolation method;

[0012] S5, according to the fused image data and the spectral index curve, the tidal flat is identified by using the dynamic change of tide level and the feature of no vegetation, and the tidal flat distinction image is obtained;

[0013] S6, according to the tidal flat distinction image and the spectral index curve, using phenological characteristics to distinguish evergreen mangroves and deciduous salt marshes, so as to obtain a final classification result of the coastal wetland image.

[0014] Preferably, the steps of using the NSPI algorithm to perform cloud layer filtering and cloud amount threshold screening are:

[0015] Using Google Earth Engine, bit-by-bit operations are performed on the data recorded in the QA_PIXEL band of the Landsat satellite image data in turn to perform cloud threshold filtering processing;

[0016] During the cloud threshold filtering process, the first four bits of the QA_PIXEL band are compared with zero to identify pixel fill values, expansion clouds, high-confidence cirrus clouds, clouds, and cloud shadows;

[0017] Select pixels whose first four bits in the QA_PIXEL band are all zero to create a mask to block invalid pixels;

[0018] Select pixels with a QA_RADSAT band value of zero in the Landsat satellite image data to generate a saturation filter mask to exclude pixels with saturated spectral information in the image;

[0019] Applying the created mask and saturation filter mask to the Landsat satellite image data to obtain a corresponding Landsat cloud-free image;

[0020] According to the area of ​​Landsat cloud-free images, the cloud-free area ratio and cloud pollution area ratio are obtained, and the filtering threshold of Landsat cloud-free images is adjusted according to the cloud pollution area ratio;

[0021] According to the adjusted Landsat cloud-free image, similar pixels are identified using an adaptive moving window and a preset threshold, and the spectral values ​​of cloud-covered pixels are predicted based on a weighted linear model using similar pixels adjacent in space and time, so as to effectively remove the cloud amount in the Landsat cloud-free image, thereby obtaining cloud-free and clear Landsat satellite image data.

[0022] Preferably, the regression-based spatiotemporal fusion model integrates partial least squares regression (PLSR) into a regression-based spatiotemporal fusion method, and S3 is specifically:

[0023] Based on PLSR, the regression-based spatiotemporal fusion model is established and trained using MODIS image data filtered by cloud threshold, and the spatiotemporal fusion model corresponding to the regression coefficient under the optimal PLS component is selected to obtain the trained spatiotemporal fusion model.

[0024] Inputting the cloud-free and clear Landsat satellite image data into the trained spatiotemporal fusion model in sequence to predict the predicted Landsat image at the next moment;

[0025] Since the spatial resolutions of MODIS and Landsat images are quite different, the spatial interpolation residual method from coarse resolution to fine resolution may introduce errors and smooth spatial textures. Therefore, adaptive residual compensation is used to optimize the errors in the fusion process to obtain the PLSR predicted Landsat image.

[0026] The PLSR predicted Landsat image is merged with the cloud-free and clear Landsat satellite image data to obtain fused image data with less cloud cover and uniform distribution over time.

[0027] Preferably, the spectral index includes: Normalized Difference Vegetation Index NDVI, Normalized Difference Water Index NDWI, Mangrove Vegetation Index MVI and Enhanced Vegetation Index EVI, which are used for identifying and distinguishing surface water bodies and wetland vegetation.

[0028] Preferably, the S4 is specifically:

[0029] Extracting wetland vector data from the multi-layer coastal wetland map by wetland type; wherein the wetland types include mangroves, salt marshes and tidal flats;

[0030] Based on Google Earth Engine, a number of random points are randomly extracted from the wetland type vector data, and the classification threshold of the corresponding wetland type is obtained by calculating the mean and standard deviation of the random point index data;

[0031] According to the classification threshold of each wetland type, outliers in random points are removed, and the spectral index curve is reconstructed using the time linear interpolation method; wherein the spectral index curve includes the normalized vegetation index NDVI curve, the enhanced vegetation index EVI curve, the mangrove vegetation index MVI curve and the normalized water body index NDWI curve.

[0032] Preferably, the S5 is specifically:

[0033] Based on the fused image data, a high tide image with an NDWI band is generated by using a maximum synthesis method of Google Earth Engine, thereby obtaining a highest tide level image;

[0034] Based on the fused image data, synthesizing the Landsat low tide image by the ninety percentile method, thereby obtaining the lowest tide image to ensure the spatial range of subsequent processing steps;

[0035] The highest tidal water pixel of the highest tidal level image and the lowest tidal water pixel of the lowest tidal level image are identified as tidal flats to determine the intertidal zone boundary and obtain a tidal flat distinction image.

[0036] Preferably, the high tide image with NDWI band is defined as HETC image, and the low tide image of Landsat is defined as LETC image. When determining the boundary of the intertidal zone:

[0037] Selecting a vector boundary whose normalized difference water index NDWI value of the HETC image is greater than a first preset threshold as the maximum range of seawater at high tide, thereby obtaining a vector boundary of the maximum seawater range;

[0038] Selecting a vector boundary whose normalized water index NDWI value of the LETC image is less than a second preset threshold and whose water frequency index is greater than a third preset threshold as the minimum range of seawater coverage at low tide, thereby obtaining a vector boundary of the minimum seawater coverage range;

[0039] The boundary of the seawater is refined by calculating the average digital elevation model (DEM) value within the vector boundary;

[0040] The intersection of the complementary set of the vector boundary of the minimum seawater coverage range and the vector boundary of the maximum seawater coverage range is the intertidal zone boundary.

[0041] Preferably, the difference in NDVI values ​​between summer and winter for each vegetation type is calculated based on the tidal flat differentiation image and the spectral index curve, thereby obtaining image data for distinguishing mangroves and salt marshes, namely, a coastal wetland classification image.

[0042] Preferably, when distinguishing image data of mangroves and salt marshes, the mangrove vegetation index (MVI) is used as an additional parameter to distinguish mangroves from inland vegetation to reduce classification errors.

[0043] The present invention also provides a device for automatically classifying coastal ecosystems based on multi-source remote sensing fusion, comprising:

[0044] An acquisition unit, used to acquire Landsat satellite image data, MODIS image data and multi-layer coastal wetland maps of the target area;

[0045] A filtering and screening unit, used to use Google Earth Engine to perform NSPI algorithm cloud layer filtering and cloud amount threshold screening on the Landsat satellite image data to obtain cloud-free and clear Landsat satellite image data;

[0046] A multi-source data fusion unit is used to fuse and reconstruct the cloud-free and clear Landsat satellite image data with the MODIS image data after being screened by a cloud threshold by adopting a regression-based spatiotemporal fusion model, so as to obtain fused image data with less cloud and uniform distribution over time;

[0047] The spectral index curve reconstruction unit is used to extract wetland vector data from the multi-layer coastal wetland map in layers according to wetland type; based on Google Earth Engine, a number of random points are randomly selected from the selected wetland type vector data, and the corresponding spectral index threshold is obtained by calculating the mean and standard deviation of the spectral index of the random points; and then the spectral index curve is reconstructed using a time linear interpolation method;

[0048] A tidal flat distinguishing unit, used to identify the tidal flat based on the fused image data and the spectral index curve, using the dynamic change of tide level and the feature of no vegetation, and obtain a tidal flat distinguishing image;

[0049] The classification result unit is used to distinguish the evergreen mangroves and the deciduous salt marshes according to the tidal flat distinction image and the spectral index curve, and to obtain the final classification result of the coastal wetland image by using phenological characteristics.

[0050] The present invention also provides an automatic classification device for coastal ecosystems based on multi-source remote sensing fusion, including a processor and a memory, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement an automatic classification method for coastal ecosystems based on multi-source remote sensing fusion as described above.

[0051] The present invention also provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, a method and device for automatic classification of coastal ecosystems based on multi-source remote sensing fusion as described above are implemented.

[0052] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0053] The present invention adopts a data fusion algorithm that can eliminate cloud interference and adapt to tidal dynamics to construct high-resolution remote sensing spectral image data of coastal wetlands; and uses a linear interpolation method to reconstruct the spectral index curve, and automatically extracts random points to construct an automated classification method for coastal wetlands based on classification thresholds. The present invention fuses and predicts Landsat and MODIS long-term images for multi-source data, effectively improving the accuracy of wetland classification, helping to accurately monitor the composition, structure and dynamic trends of the blue carbon ecosystem functions of coastal wetlands, and providing dynamic monitoring and protection warnings for the future sustainable development of the blue carbon ecosystem of coastal wetlands. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1 A schematic diagram of a method and device for automatic classification of coastal ecosystems based on multi-source remote sensing fusion provided in Example 1 of the present invention.

[0056] Figure 2 A schematic diagram of multi-source remote sensing data fusion provided in Example 1 of the present invention.

[0057] Figure 3 This is a schematic diagram of extracting classification thresholds and reconstructing NDVI curves provided in Example 1 of the present invention.

[0058] Figure 4 This is a schematic diagram of the time interpolation reconstruction principle provided in the first embodiment of the present invention.

[0059] Figure 5 The low tide image a and high tide image b of area A, as well as the corresponding minimum seawater coverage image c and maximum seawater coverage image d are provided in the first embodiment of the present invention.

[0060] Figure 6 A schematic diagram of the coastal wetland classification process provided in Example 1 of the present invention.

[0061] Figure 7 The original images a, b, c, and the fused images d, e, f of area B provided in the first embodiment of the present invention.

[0062] Figure 8 A schematic diagram of the NDVI curve of the high and medium tide reconstruction process of area A provided in Example 1 of the present invention, wherein the left figure represents the original curve of the high and medium tide vegetation index NDVI before interpolation; the middle curve graph represents the high and medium tide vegetation index NDVI curve after screening and removing the outliers in the curve; the right figure represents the high and medium tide vegetation index NDVI curve reconstructed after interpolation and re-filling the blank part after removing the outliers.

[0063] Fig. 9 Graph a showing the difference in NDVI between mangroves and salt marshes in summer and winter in region A provided in Example 1 of the present invention; and graph b showing the minimum NDVI value of mangroves in winter and the maximum and minimum NDVI values ​​of salt marshes in winter in region A.

[0064] Fig.10 This is a classification result diagram of coastal wetlands in Region A provided in Example 1 of the present invention.

[0065] Fig.11 A schematic diagram of an automatic classification device for coastal ecosystems based on multi-source remote sensing fusion provided in Example 2 of the present invention.

[0066] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0068] Embodiment 1

[0069] As the key interface of the coordinated development of land and sea in the coastal zone, coastal wetlands play a vital role in addressing climate change. Large-scale and precise monitoring of the structure and function of blue carbon ecosystems in coastal wetlands requires remote sensing data with high temporal and spatial resolution. However, due to the influence of high-frequency changes in cloud interference and tidal inundation, the quantity and quality of effective images of coastal wetlands are limited. Multi-source remote sensing fusion technology, as a technical means of efficient and continuous monitoring, can reconstruct data based on missing original images, but it is still in the initial development stage in the dynamic monitoring of coastal wetland ecosystems. Based on the basic theory of traditional coastal wetland classification and mapping, the present invention constructs a data fusion algorithm that can eliminate cloud interference and adapt to tidal dynamics, reconstruct remote sensing spectral data of coastal wetlands with high temporal resolution, and realizes automatic image classification based on the significant spectral differences of coastal wetlands at different tide levels and phenological stages.

[0070] Embodiment 1 of the present invention provides a method for automatic classification of coastal ecosystems based on multi-source remote sensing fusion, which can be implemented by an automatic classification device for coastal ecosystems based on multi-source remote sensing fusion (hereinafter referred to as the automatic classification device for coastal ecosystems), and in particular, is executed by one or more processors in the automatic classification device for coastal ecosystems.

[0071] In this embodiment, the coastal ecosystem automatic classification device may be an electronic device equipped with a processor, which carries a computer program of the coastal ecosystem automatic classification method based on multi-source remote sensing fusion and the computer program can be executed, such as a computer, a smart phone, a smart tablet, a workstation, etc., which is not limited here.

[0072] like Figure 1 As shown, a method and device for automatic classification of coastal ecosystems based on multi-source remote sensing fusion includes steps S1 to S6.

[0073] S1, obtain Landsat satellite image data, MODIS image data and multi-layer coastal wetland maps of the target area.

[0074] The Landsat satellite image data includes Landsat-5, Landsat-7 and Landsat-8 satellite image data.

[0075] The MODIS image data is a globally free remote sensing data source with a wide spectral range, a short revisit period, and can be downloaded through NASA's official website.

[0076] The multi-layer coastal wetland map used in this embodiment is the 2020 East Asia Multi-layer Tidal Wetland Map, which is based on the MTWM-TP classification framework and generated using the Sentinel-2 image time series.

[0077] S2, using Google Earth Engine, after performing NSPI algorithm cloud layer filtering and cloud amount threshold screening on the Landsat satellite image data, cloud-free and clear Landsat satellite image data is obtained.

[0078] Specifically, Figure 2 As shown in the figure, before image fusion, the image of the target area needs to be filtered by NSPI algorithm cloud threshold to obtain high-quality cloud-free images.

[0079] In this embodiment, the NSPI (Non-Stationary Standardized Precipitation Index) algorithm is a method for removing cloud contamination from satellite images based on an improved neighborhood similar pixel interpolation method, which uses a weighted linear model to predict the spectral spatial and spectral temporal information of a target pixel from similar pixels adjacent to the target pixel (i.e., an unscanned pixel).

[0080] After selecting the target area, use the area calculation function to obtain the area of ​​the selected target area, and filter and preprocess the images of the target research area before downloading to reduce invalid data. Taking Landsat 8 data as an example, the steps for cloud filtering and cloud threshold screening using the NSPI algorithm are as follows:

[0081] Using the powerful cloud computing capability of Google Earth Engine, bit-by-bit operations are performed on the data recorded in the QA_PIXEL band of the Landsat satellite image data in turn to perform cloud threshold filtering processing;

[0082] During the cloud threshold filtering process, the first 0-4 bits of the QA_PIXEL band are compared to zero to identify pixel fill values, puff clouds, high confidence cirrus clouds, clouds, and cloud shadows.

[0083] Select pixels in the QA_PIXEL band whose first 0-4 bits are all zero to create a mask to mask invalid pixels, such as occluding clouds and cloud shadows.

[0084] Then, pixels with zero values ​​in the QA_RADSAT band were selected to indicate that the image was unsaturated, and a saturation filter mask was generated to exclude pixels with saturated spectral information in the image. This mask was used to exclude pixels with saturated spectral information in subsequent processing to ensure the accuracy and reliability of the data.

[0085] Apply the created mask and saturation filter mask to the target area image data, select the red band and convert it into a vector area, and you can get an image without the above interference, that is, the corresponding Landsat cloud-free image.

[0086] According to the area of ​​Landsat cloud-free images, the cloud-free area and total area are calculated to obtain the cloud-free area ratio and cloud pollution area ratio. According to the cloud pollution area ratio, the cloud pollution value is added to the image to adjust the filter threshold of Landsat cloud-free images.

[0087] According to the adjusted Landsat cloud-free image, similar pixels are identified using an adaptive moving window and a preset threshold, and the spectral values ​​of cloud-covered pixels are predicted based on a weighted linear model using similar pixels adjacent in space and time, so as to effectively remove the cloud amount in the Landsat cloud-free image, thereby obtaining cloud-free and clear Landsat satellite image data.

[0088] S3, using a regression-based spatiotemporal fusion model to fuse and reconstruct the cloud-free and clear Landsat satellite image data with the MODIS image data after cloud threshold screening, thereby obtaining fused image data with less cloud cover and uniform distribution over time.

[0089] Specifically, since the temporal resolution of Landsat images is low and there are limited images suitable for constructing a complete time series after excluding high cloud images, the spatiotemporal fusion model based on the multi-source data fusion algorithm provides the potential to reconstruct a complete series and improve the temporal resolution. High temporal resolution images are combined with low spatial resolution images, and mathematical methods such as weighted functions are used to construct images of a specific period.

[0090] The present invention uses a spatiotemporal fusion model combined with spectral autocorrelation to reconstruct time series data. Compared with other spatiotemporal fusion models, this model has a higher efficiency in data utilization. On the one hand, the present invention utilizes the autocorrelation characteristics of the image band to improve the accuracy of the reconstructed data; on the other hand, it requires less input data, only a pair of basic images and a reference image are needed to complete the image reconstruction, which is more suitable for situations where the original image is limited.

[0091] In this embodiment, the spatiotemporal fusion model based on multivariate linear regression is a novel multi-spectral spatiotemporal fusion method.

[0092] After cloud cover screening, the obtained cloud-free and clear Landsat satellite image data are combined with MODIS data.

[0093] Specifically, an adaptive moving window and threshold are used to identify similar pixels, and similar pixels adjacent in space and time are used to predict the spectral values ​​of cloud-covered pixels based on a weighted linear model. For example, for an image with clouds at time t1, an auxiliary image at the nearest date t2 is obtained through a regression-based spatiotemporal fusion model prediction (these auxiliary images are cloudless in the cloudy part of the t1 image). For pixels at the edge of the cloud, when predicting based on spectral spatial information, they are given greater weights to better maintain spatial continuity and reduce edge effects. For pixels in the center of the cloud, more weights are given when predicting based on spectral-temporal information. This spectral-temporal prediction can effectively remove cloud cover from Landsat satellite images. Subsequently, the cloud-free and clear Landsat satellite image data processed by the NSPI algorithm is merged with the MODIS data using a regression-based spatiotemporal fusion model.

[0094] The regression-based spatiotemporal fusion model integrates partial least squares regression (PLSR) into the regression-based spatiotemporal fusion method, specifically:

[0095] Based on PLSR, the regression-based spatiotemporal fusion model is established and trained using MODIS image data filtered by cloud threshold, and the spatiotemporal fusion model corresponding to the regression coefficient under the optimal PLS component is selected to obtain the trained regression-based spatiotemporal fusion model.

[0096] The cloud-free and clear Landsat satellite image data are sequentially input into the trained regression-based spatiotemporal fusion model to predict the predicted Landsat image at the next moment.

[0097] Specifically, using PLSR, a regression model is established between the coarse-resolution MODIS image C1 at time t1 and the MODIS image C2 at time t2, and the regression model under the optimal PLS component is selected, which can improve the accuracy of the regression coefficient, reduce the computational complexity, and optimize the prediction results.

[0098] Assuming that the regression model remains constant at the local scale, the relationship established at coarse resolution can be applied to fine spatial resolution. Therefore, the regression coefficients trained on MODIS images can be transferred and directly applied to Landsat image F1 at time t1 to produce the predicted Landsat image F2 at time t2.

[0099] Since the MODIS image (in this example, the resolution is 100 meters 100 meters) and Landsat images (in this example, the resolution is 30 meters 30 meters) has a large difference in spatial resolution. Using the spatial interpolation residual method from coarse resolution to fine resolution may introduce errors and smooth spatial textures. Therefore, an adaptive residual compensation method is used to optimize the errors in the fusion prediction results. Compared with the prediction results without residual compensation, a more robust fusion result can be produced, thereby obtaining the PLSR predicted Landsat image.

[0100] The PLSR predicted Landsat image is merged with the cloud-free and clear Landsat satellite image data to obtain fused image data with less cloud cover and uniform distribution over time.

[0101] In this embodiment, PLSR is a dimensionality reduction technique for supervised learning, which is particularly suitable for situations with many factors and high collinearity. In the regression-based spatiotemporal fusion model, PLSR is used to solve the problem of multivariate multiple regression, that is, using multiple independent variables to simultaneously predict multiple dependent variables (many-to-many framework). This method can make full use of the spectral information of multiple bands and alleviate the cascade effect, thereby optimizing the prediction results.

[0102] The regression-based spatiotemporal fusion model makes full use of the spectral autocorrelation of multiple bands. The spectral values ​​of multiple bands usually change simultaneously and with certain rules, rather than being independent of each other. Therefore, using multiple bands instead of a single band in regression can obtain more information to build a more accurate model to capture complex temporal variation patterns and more accurate classification accuracy.

[0103] The prediction results can be further optimized by selecting the optimal number of PLS ​​components. This helps to reduce the computational complexity while maintaining the accuracy of the model.

[0104] The adaptive residual compensation ARC process is used to further optimize the prediction results. By adaptively adjusting the residual, the prediction accuracy of the model can be further improved.

[0105] S4, extracting wetland vector data from the multi-layer coastal wetland map in layers according to wetland type; based on Google Earth Engine, randomly extracting a number of random points from the selected wetland type vector data, and obtaining the corresponding spectral index threshold by calculating the mean and standard deviation of the spectral index of the random points; and then reconstructing the spectral index curve using a time linear interpolation method.

[0106] Specifically, the spectral index includes: Normalized Difference Vegetation Index NDVI, Normalized Difference Water Index NDWI, Mangrove Vegetation Index MVI and Enhanced Vegetation Index EVI, which are used to identify and distinguish surface water bodies and wetland vegetation.

[0107] like Figure 3The flowchart for extracting classification thresholds and reconstructing NDVI curves is shown. By automatically extracting random points from the existing land cover type dataset to generate corresponding thresholds for different blue carbon ecosystems, the classification efficiency is greatly improved while effectively ensuring the classification accuracy.

[0108] Manually extracting sample points from existing data sets to calculate thresholds is not only time-consuming, but also has poor stability and randomness, resulting in low accuracy of threshold data. Therefore, the present invention uses random point coding to randomly extract a number of points (in this embodiment, set to 100) from the image data of different phenological and tidal characteristics of surface cover type samples. Based on the acquired random sample points, the time linear interpolation of spectral index curves such as NDVI, NDWI and MVI curves is reconstructed to improve the accuracy of the classification results of coastal blue carbon ecosystems.

[0109] The specific steps are as follows:

[0110] First, in this embodiment, a multi-layer coastal wetland map of the target area in 2020 is imported, and wetland vector data is extracted layer by layer according to wetland types, wherein the wetland types include mangroves, salt marshes, and tidal flats;

[0111] Then, based on the ee.FeatureCollection.RandomPoints encoding of the GEE (Google Earth Engine) platform, 150 random points were randomly selected from the wetland type vector data;

[0112] Finally, the mean and standard deviation of the index data of the random points were calculated to obtain the classification threshold of the corresponding type. To avoid the influence of extreme values ​​on the data, the data within the range of ±2 standard deviations of the mean were retained. The classification threshold of the land cover type obtained can more accurately classify the composition of wetlands.

[0113] Compared with the manual method, this method of automatically extracting random sample points is more convenient.

[0114] Then reconstruct the time linear interpolation of NDVI, NDWI and MVI curves. The Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Water Index (NDWI) are closely related to the tide level. The NDVI value is negatively correlated with the tidal height, and the NDWI value is positively correlated with the tidal height. The minimum NDVI value and the minimum NDWI value theoretically represent the highest tide and the lowest tide, respectively (because the NDVI value on the water body is lower and the NDWI value is higher). Since mangroves and salt marshes are affected by tidal fluctuations, and Landsat images are often affected by clouds, this embodiment establishes an annual vegetation index curve and a water index curve.

[0115] According to the classification threshold of each wetland type, outliers in random points are removed to reduce the interference of the exponential curve, and the blanks in the removed values ​​are filled using the time linear interpolation method to reconstruct the spectral index curve; wherein the spectral index curve includes NDVI curve, NDWI curve and MVI curve.

[0116] The above process aims to obtain a more accurate exponential curve by improving the extraction of random points and threshold calculation, thereby improving the accuracy of subsequent wetland classification. Other existing methods can also be used to extract data points and calculate thresholds, which will not be repeated here.

[0117] The interpolated spectral index and time interpolation reconstruction calculation formula are shown in Table 1.

[0118] Table 1 Spectral indices used for interpolation

[0119]

[0120] Among them, ρ nir Represents the reflectivity of the near-infrared band in remote sensing images; ρ red Represents the reflectivity of the infrared band in the remote sensing image; ρ blue Represents the reflectivity of the blue band in the remote sensing image; ρ green Represents the reflectivity of the green band in the remote sensing image; ρ swir1 Represents the reflectivity of the shortwave infrared band in the remote sensing image.

[0121] In another preferred embodiment, assuming that the missing value to be interpolated from the NDVI curve is x, to obtain the time of three points a, b, and x and the values ​​of two points a and b, the calculation steps are as follows:

[0122] 1) Add the complete DOY (day of year) band and the masked DOY band to two image queues Landsat imageCollection respectively, and get two image queues and name them L8_Series_2020 and L8_Series_2020_MaskDate respectively.

[0123] 2) For L8_Series_2020_MaskDate, use ee.reduce.lastNONnull() and ee.reduce.firstnonnull() to obtain the values ​​and dates of points a and b;

[0124] 3) For L8_Series_2020, get the date of point x;

[0125] 4) Find the value of x according to the following formula

[0126] ;

[0127] Among them, NDVI() represents the NDVI curve function; Indicates the time difference between points a and x in the NDVI curve; Indicates the time difference between points a and b in the NDVI curve.

[0128] According to the above method, the curve is interpolated and reconstructed to obtain the reconstructed curve. Figure 4 As shown, HNDVI (blue) is the reconstructed curve of the vegetation index at high tide level, and MNDVI1 (orange) is the reconstructed curve of the vegetation index at mid tide level.

[0129] S5, according to the fused image data and the spectral index curve, the tidal flat is identified by utilizing the tidal level characteristics and the vegetation-free characteristics to obtain a tidal flat distinguishing image.

[0130] The absence of vegetation on tidal flats sets them apart from other wetland ecosystems such as mangroves and salt marshes. This distinction is particularly evident when observing satellite imagery and analyzing vegetation indices. The Normalized Difference Vegetation Index (NDVI) is widely used to quantify the greenness and density of vegetation. On tidal flats, NDVI values ​​remain low at low tide due to the lack of photosynthetically active vegetation. Mangroves and salt marshes have higher NDVI values ​​due to their dense and diverse vegetation structure.

[0131] In addition to NDVI, the Normalized Difference Water Index (NDWI) is key to understanding the hydrological dynamics of tidal flats. NDWI is used to measure and monitor soil moisture content and flooding status, which is key to distinguishing water-saturated surfaces from dry ground. At high tide, NDWI values ​​are minimum for tidal flats, reflecting flooding conditions, while at low tide, NDWI values ​​in these areas are maximum, reflecting the exposure and relative dryness of the tidal flats compared to flooded conditions. The significant changes in tides make the Normalized Difference Water Index (NDWI) values ​​of tidal flats, NDWI values ​​at high tide (minimum), and NDWI values ​​at low tide (maximum), important indicators for distinguishing tidal flats from other wetlands and various land cover types inland.

[0132] Specifically, according to the fused image data, a high tide image with NDWI band is generated by using the Google Earth Engine qualitymosaic method, so as to obtain the highest tide level image HETC.

[0133] According to the fused image data, the Landsat low tide image is synthesized by the ninety percentile method, so as to obtain the lowest tide image LETC.

[0134] The highest tidal water pixel of the highest tidal level image and the lowest tidal water pixel of the lowest tidal level image are identified as tidal flats to determine the intertidal zone boundary and obtain a tidal flat distinguishing image for distinguishing tidal flats.

[0135] Specifically, when determining the intertidal zone boundaries:

[0136] Since the NDWI value is positively correlated with the tidal height, the vector boundary whose normalized water index NDWI value of the HETC image is greater than the first preset threshold (in this embodiment, set to 0.2) is selected as the maximum range of seawater at high tide, thereby obtaining the vector boundary of the maximum seawater range.

[0137] Due to the potential interference of inland water pixels and the large elevation differences between inland water bodies, the high tide boundary is refined by calculating the average digital elevation model (DEM) value within the vector boundary. In order to remove inland errors, the area with an average DEM value less than 1 is retained, and the seawater coverage is the largest.

[0138] Since the NDVI value is negatively correlated with the tidal height, the vector boundary whose normalized water index NDWI value of the LETC image is less than the second preset threshold (in this embodiment, set to 0) and whose water frequency index is greater than the third preset threshold (in this embodiment, set to 97.5) is selected as the minimum range of seawater at low tide, thereby obtaining the vector boundary of the minimum seawater range. Considering the possible interference of inland pixels, the area with an average DEM value less than 1 is retained to refine the minimum seawater range to eliminate inland errors.

[0139] In this embodiment, the water body frequency index is an indicator used to measure the frequency of water bodies in a specific area. It is calculated based on remote sensing data and can reflect the distribution characteristics of water bodies in time and space. It can be calculated by obtaining the ready-made JRC data set from the GEE platform.

[0140] In the synthesized Landsat image, the intersection of the complementary set of the vector boundary of the minimum seawater range and the vector boundary of the maximum seawater range is the intertidal zone boundary.

[0141] like Figure 5 As shown, in another preferred embodiment, taking a coastal wetland in a certain area (referred to as area A) as an example, high and low tide images and regional maps of maximum and minimum seawater difference vectors are extracted. Figure a is a low tide image of area A, Figure b is a high tide image of area A, Figure c is the corresponding minimum sea width map of area A, and Figure d is the corresponding maximum sea width map of area A.

[0142] S6, according to the tidal flat distinction image and the spectral index curve, using phenological characteristics to distinguish evergreen mangroves and deciduous salt marshes, so as to obtain a final classification result of the coastal wetland image.

[0143] Since both mangroves and salt marshes are affected by tidal fluctuations, tidal characteristics alone are not sufficient to effectively distinguish between these two types of coastal wetlands. Therefore, this classification method incorporates phenological characteristics, as salt marshes have a strong seasonal green feature, while mangroves are evergreen.

[0144] Specifically, Figure 6 As shown, based on the tidal flat differentiation image and the spectral index curve, the difference in NDVI values ​​in summer and winter for each vegetation type is calculated, thereby obtaining image data for distinguishing mangroves and salt marshes with significant seasonal changes, that is, the classification result of coastal wetland images.

[0145] In the present embodiment, the images of January, February, March, October, November and December are synthesized by calculating the median of the winter images. Similarly, the images from April to September are synthesized by calculating the median of the summer images. Since the normalized difference vegetation index (NDVI) is positively correlated with green vegetation, and the seasonality of evergreen mangroves is weak, while the deciduous salt marsh has obvious seasonal characteristics of long summer and weak winter, the winter NDVI value of the salt marsh should be lower than that of the mangrove, and the NDVI difference between the summer and winter of the salt marsh should be greater than that of the mangrove. By calculating the maximum NDVI in winter of the salt marsh plot and the minimum NDVI difference between summer and winter, the salt marsh plot can be distinguished from the mangrove. Considering that the tidal flat lacks vegetation coverage, its winter NDVI should be lower than that of the salt marsh. Therefore, by calculating the minimum value of the winter NDVI of the salt marsh sample, the tidal flat can be distinguished from the salt marsh. The winter NDVI value of the mangrove is significantly higher than that of the deciduous salt marsh, and it shows greater stability during seasonal changes. Therefore, by calculating the minimum NDVI value of mangrove samples in winter and the maximum difference between summer and winter NDVI, mangroves can be distinguished from deciduous salt marshes.

[0146] In addition, when distinguishing the image data of mangroves and salt marshes, the mangrove vegetation index (MVI) was used as an additional parameter to distinguish mangroves from inland vegetation to reduce classification errors.

[0147] In another preferred embodiment, it is observed that many inland pixels exhibit seasonal characteristics similar to salt marshes, resulting in potential errors. To address this issue, a coastal buffer is created by synthesizing the 10th percentile of the Landsat image to produce a mid-tide image. The portion with NDVI less than 0 and average DEM value less than 1 is retained as the vector boundary of the seawater area. A 400-meter buffer is established along the edge of the seawater vector boundary, and pixels outside the buffer are removed from the salt marsh image, effectively eliminating the errors caused by inland objects.

[0148] In another preferred embodiment, a highly dynamically changing coastal area (referred to as Region B) is selected as a case, and Landsat-8 images are selected as the basic data set for 2020. After the cloud filter is implemented, the frequency of available images is significantly reduced. A detailed visual inspection is performed on each image to exclude images with unqualified sensor performance. This method uses a regression-based spatiotemporal fusion model to perform spatiotemporal data fusion prediction, which helps to estimate the images of missing months in each region, thereby reconstructing an image sequence with less cloud cover and more uniform distribution over time. After cloud cover threshold screening, a total of 5 original images with minimal or no cloud cover were collected. These images are then fused with MODIS images using a regression-based spatiotemporal fusion model to obtain 19 forecast images for Region B in 2020. The temporal resolution of the reconstructed time series for this area is increased to two images per month. As Figure 7 As shown, it is the data fusion result of area B. The first three images a, b, and c are original images, and the last three images d, e, and f are the prediction results after image data fusion.

[0149] It can be seen that the reconstruction quality of region B is significantly higher. Although the temporal resolution of the original images is low, their individual and overall quality is quite good. The extremely low cloud cover of each original image ensures the maximum validity of the data. Overall, the distribution of the original images is quite uniform, and the prediction time span required by the regression-based spatiotemporal fusion model is short, while taking advantage of the strengths of Landsat-8 and MODIS images. The reconstructed image produced by this integration shows enhanced spatial and temporal resolution.

[0150] like Figure 8 As shown in the figure, in order to demonstrate the effectiveness of different steps, area A is selected as a case in this step to show the NDVI reconstruction curve of the high and medium tide vegetation index. Among them, HNDVI (blue) is the high tide vegetation index reconstruction curve, and MNDVI1 (orange) is the mid-tide vegetation index reconstruction curve. The three figures from left to right respectively represent: the original NDVI curve of the high and medium tide vegetation index before interpolation, the NDVI curve of the high and medium tide vegetation index after screening and removing the outliers in the curve (indicated by the red dotted line), and the new NDVI curve of the high and medium tide vegetation index obtained by interpolating and refilling the blank part after removing the outliers. The trend of the reconstructed spectral index curve is more consistent with the actual vegetation index change trend, which is conducive to the extraction of wetland classification thresholds.

[0151] In the 2020 East Asian multi-level coastal wetland map, sampling points were randomly selected to calculate the thresholds for different wetland types. In this study, 90 thresholds generated by random points were selected as the basis for classification, which can ensure better classification results. The thresholds for tidal flat wetlands in Area A were determined as follows: for high-tide images, NDWI > 0.484; for low-tide images, NDWI < 0.496, and -0.197 < NDVI < 0.2. For salt marsh wetlands, the thresholds for sNDVI_winter in winter images were defined as: 0.122 < sNDVI_winter < 0.548, and for sNDVI_summer in summer images as: 0.252 < sNDVI_summer < 0.480, and the difference in sNDVI between summer and winter (sNDVIdiff) was greater than 0.201. The classification threshold criteria for mangrove wetlands were: the difference in mNDVIdiff in winter images > 0.264, the difference in mNDVIdiff between summer and winter < 0.013, and the modified vegetation index (MVI) > 4.

[0152] As Fig. 9 shown, the graph of the key vegetation indices used to distinguish mangroves, salt marshes, and tidal flats in the wetlands of Area A is presented.

[0153] As Fig.10 shown, by using the tidal and phenological feature classification method to obtain thresholds, the three wetland types were classified to generate a wetland change map of Area A from 2011 to 2020.

[0154] Generally speaking, from 2011 to 2020, mangroves and salt marshes in Area A faced certain pressures and challenges, but through effective protection measures and restoration projects, these precious coastal wetland blue carbon ecosystems have been protected and restored to a certain extent. Continuous protection work and scientific research are crucial for the future sustainable development of these areas.

[0155] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0156] The present invention adopts a data fusion algorithm that can eliminate cloud interference and adapt to tidal dynamics to construct high-resolution remote sensing spectral image data of coastal wetlands; and uses the linear interpolation method to reconstruct the spectral index curve, and constructs an automated classification method for coastal wetlands based on classification thresholds by automatically extracting random points. The present invention performs multi-source data fusion and prediction on Landsat and MODIS long-time series images, effectively improving the accuracy of wetland classification and contributing to the precise monitoring of the composition, structure, and dynamic trends of the functions of coastal wetland blue carbon ecosystems.

[0157] Example Two

[0158] As Fig.11As shown, the second embodiment of the present invention further provides an automatic classification device for coastal ecosystems based on multi-source remote sensing fusion, comprising:

[0159] An acquisition unit, used to acquire Landsat satellite image data, MODIS image data and multi-layer coastal wetland maps of the target area;

[0160] A filtering and screening unit, used to use Google Earth Engine to perform NSPI algorithm cloud layer filtering and cloud amount threshold screening on the Landsat satellite image data to obtain cloud-free and clear Landsat satellite image data;

[0161] A multi-source data fusion unit is used to fuse and reconstruct the cloud-free and clear Landsat satellite image data with the MODIS image data after being screened by a cloud threshold by adopting a regression-based spatiotemporal fusion model, so as to obtain fused image data with less cloud and uniform distribution over time;

[0162] The spectral index curve reconstruction unit is used to extract wetland vector data from the multi-layer coastal wetland map in layers according to wetland type; based on Google Earth Engine, a number of random points are randomly selected from the selected wetland type vector data, and the corresponding spectral index threshold is obtained by calculating the mean and standard deviation of the spectral index of the random points; and then the spectral index curve is reconstructed using a time linear interpolation method;

[0163] A tidal flat distinguishing unit, used to identify the tidal flat based on the fused image data and the spectral index curve, using the dynamic change of tide level and the feature of no vegetation, and obtain a tidal flat distinguishing image;

[0164] The classification result unit is used to distinguish the evergreen mangroves and the deciduous salt marshes according to the tidal flat distinction image and the spectral index curve, and to obtain the final classification result of the coastal wetland image by using phenological characteristics.

[0165] Embodiment 3

[0166] The third embodiment of the present invention also provides an automatic classification device for coastal ecosystems based on multi-source remote sensing fusion, which includes a memory and a processor, wherein the memory stores a computer program, and the computer program can be executed by the processor to implement the automatic classification method for coastal ecosystems based on multi-source remote sensing fusion as described above.

[0167] Embodiment 4

[0168] The fourth embodiment of the present invention also provides a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor of a device where the computer-readable storage medium is located, the automatic classification method of coastal ecosystems based on multi-source remote sensing fusion as described above is implemented.

[0169] In several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed apparatus and method can also be implemented in other ways. The apparatus and method embodiments described above are merely schematic. For example, the flowcharts in the accompanying drawings show the possible architecture, functions and operations of the apparatus, method and computer program product according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0170] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0171] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, an electronic device, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code. It should be noted that in this article, the term "include", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such process, method, article or device. Without more constraints, an element defined by the phrase "comprising a..." does not exclude the existence of other identical elements in the process, method, article or apparatus comprising the element.

[0172] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0173] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0174] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0175] The "first\second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order for the objects. It is understandable that the "first\second" can be interchanged with the specific order or sequence where permitted. It should be understood that the objects distinguished by "first\second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0176] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An automatic classification method for coastal ecosystems based on multi-source remote sensing fusion, characterized in that: include: S1, obtain Landsat satellite image data, MODIS image data and multi-layer coastal wetland maps of the target area; S2, using Google Earth Engine, performing NSPI algorithm cloud layer filtering and cloud threshold screening on the Landsat satellite image data to obtain cloud-free and clear Landsat satellite image data; wherein the NSPI algorithm is a method for removing cloud pollution from satellite images based on an improved neighborhood similar pixel interpolation method; The steps of using NSPI algorithm for cloud filtering and cloud threshold screening are as follows: Using Google Earth Engine, bit-by-bit operations are performed on the data recorded in the QA_PIXEL band of the Landsat satellite image data in turn to perform cloud threshold filtering processing; During the cloud threshold filtering process, the first four bits of the QA_PIXEL band are compared with zero to identify pixel fill values, expansion clouds, high-confidence cirrus clouds, clouds, and cloud shadows; Select pixels whose first four bits in the QA_PIXEL band are all zero to create a mask to block invalid pixels; Select pixels with a QA_RADSAT band value of zero in the Landsat satellite image data to generate a saturation filter mask to exclude pixels with saturated spectral information in the image; Applying the created mask and saturation filter mask to the Landsat satellite image data to obtain a corresponding Landsat cloud-free image; According to the area of ​​Landsat cloud-free images, the cloud-free area ratio and cloud pollution area ratio are obtained, and the filtering threshold of Landsat cloud-free images is adjusted according to the cloud pollution area ratio; According to the adjusted Landsat cloud-free image, similar pixels are identified using an adaptive moving window and a preset threshold, and the spectral values ​​of cloud-covered pixels are predicted based on a weighted linear model using similar pixels adjacent in space and time, so as to effectively remove the cloud amount in the Landsat cloud-free image, thereby obtaining cloud-free and clear Landsat satellite image data; S3, adopting a regression-based spatiotemporal fusion model to fuse and reconstruct the cloud-free and clear Landsat satellite image data with the MODIS image data after cloud threshold screening, so as to obtain fused image data with less cloud and uniform distribution over time; wherein, the regression-based spatiotemporal fusion model integrates partial least squares regression PLSR into the regression-based spatiotemporal fusion method, and S3 is specifically: Based on PLSR, the regression-based spatiotemporal fusion model is established and trained using MODIS image data filtered by cloud threshold, and the spatiotemporal fusion model corresponding to the regression coefficient under the optimal PLSR component is selected to obtain the trained spatiotemporal fusion model. Inputting the cloud-free and clear Landsat satellite image data into the trained regression-based spatiotemporal fusion model in sequence to predict the Landsat image at the next moment; Since the spatial resolution of MODIS images and Landsat images is quite different, adaptive residual compensation is used to optimize the error in the fusion process to obtain PLSR predicted Landsat images; The PLSR predicted Landsat image is merged with the cloud-free and clear Landsat satellite image data to obtain fused image data with less cloud cover, uniform distribution over time, and full regional coverage; S4, extracting wetland vector data from the multi-layer coastal wetland map in layers according to wetland type; randomly extracting a number of random points from the selected wetland type vector data based on Google Earth Engine, and obtaining the corresponding spectral index threshold by calculating the mean and standard deviation of the spectral index of the random points; and then reconstructing the spectral index curve using a time linear interpolation method; S5, according to the fused image data and the spectral index curve, the tidal flat is identified by using the dynamic change of tide level and the feature of no vegetation, and the tidal flat distinction image is obtained; S6, according to the tidal flat distinction image and the spectral index curve, using phenological characteristics to distinguish evergreen mangroves and deciduous salt marshes, so as to obtain a final classification result of the coastal wetland image.

2. The automatic classification method of coastal ecosystem based on multi-source remote sensing fusion according to claim 1 is characterized in that ,The spectral index includes: Normalized Difference Vegetation Index NDVI, Normalized Difference Water Index NDWI, Mangrove Vegetation Index MVI and Enhanced Vegetation Index EVI, which are used to identify and distinguish surface water and wetland vegetation.

3. The automatic classification method of coastal ecosystem based on multi-source remote sensing fusion according to claim 2 is characterized in that , the S4 is specifically: For the multi-layer coastal wetland map, wetland vector data is extracted layer by layer according to wetland type; wherein, The wetland types described include mangroves, salt marshes, and tidal flats; Based on Google Earth Engine, a number of random points are randomly extracted from the wetland type vector data, and the classification threshold of the corresponding wetland type is obtained by calculating the mean and standard deviation of the random point index data; According to the classification threshold of each wetland type, outliers in random points are removed, and the spectral index curve is reconstructed using the time linear interpolation method; wherein the spectral index curve includes the normalized vegetation index NDVI curve, the enhanced vegetation index EVI curve, the mangrove vegetation index MVI curve and the normalized water body index NDWI curve.

4. The automatic classification method of coastal ecosystem based on multi-source remote sensing fusion according to claim 3 is characterized in that , the S5 is specifically: Based on the fused image data, a high tide image with an NDWI band is generated by using a maximum synthesis method of Google Earth Engine, thereby obtaining a highest tide level image; Based on the fused image data, synthesizing the Landsat low tide image by the ninety percentile method, thereby obtaining the lowest tide image to ensure the spatial range of subsequent processing steps; The highest tidal water pixel of the highest tidal level image and the lowest tidal water pixel of the lowest tidal level image are identified as tidal flats to determine the intertidal zone boundary and obtain a tidal flat distinction image.

5. The automatic classification method of coastal ecosystem based on multi-source remote sensing fusion according to claim 4 is characterized in that: The high tide image with NDWI band is defined as HETC image, and the low tide image of Landsat is defined as LETC image. When determining the boundary of the intertidal zone: Selecting a vector boundary whose normalized difference water index NDWI value of the HETC image is greater than a first preset threshold as the maximum range of seawater at high tide, thereby obtaining a vector boundary of the maximum seawater range; Selecting a vector boundary whose normalized water index NDWI value of the LETC image is less than a second preset threshold and whose water frequency index is greater than a third preset threshold as the minimum range of seawater at low tide, thereby obtaining a vector boundary of the minimum seawater range; The boundary of the seawater is refined by calculating the average digital elevation model (DEM) value within the vector boundary; The intersection of the complementary set of the vector boundary of the minimum seawater range and the vector boundary of the maximum seawater range is the intertidal zone boundary.

6. The method for automatic classification of coastal ecosystems based on multi-source remote sensing fusion according to claim 5 is characterized in that ,According to the tidal flat distinction image and the spectral index curve, ,the difference of NDVI values ​​in summer and winter for each vegetation type is ,calculated, thereby obtaining image data for distinguishing mangroves and ,salt marshes, that is, the classification result of coastal wetland images.

7. The method for automatic classification of coastal ecosystems based on multi-source remote sensing fusion according to claim 6 is characterized in that ,When distinguishing the image data of mangroves and salt marshes, the mangrove vegetation index MVI is used as an ,additional parameter to distinguish mangroves from inland vegetation to reduce the ,classification error.

8. An automatic classification device for coastal ecosystems based on multi-source remote sensing fusion, characterized in that: include: An acquisition unit, used to acquire Landsat satellite image data, MODIS image data and multi-layer coastal wetland maps of the target area; A filtering and screening unit is used to use Google Earth Engine to perform NSPI algorithm cloud layer filtering and cloud threshold screening on the Landsat satellite image data to obtain cloud-free and clear Landsat satellite image data; wherein the NSPI algorithm is a method for removing cloud pollution from satellite images based on an improved neighborhood similar pixel interpolation method; The steps of using NSPI algorithm for cloud filtering and cloud threshold screening are as follows: Using Google Earth Engine, bit-by-bit operations are performed on the data recorded in the QA_PIXEL band of the Landsat satellite image data in turn to perform cloud threshold filtering processing; During the cloud threshold filtering process, the first four bits of the QA_PIXEL band are compared with zero to identify pixel fill values, expansion clouds, high-confidence cirrus clouds, clouds, and cloud shadows; Select pixels whose first four bits in the QA_PIXEL band are all zero to create a mask to block invalid pixels; Select pixels with a QA_RADSAT band value of zero in the Landsat satellite image data to generate a saturation filter mask to exclude pixels with saturated spectral information in the image; Applying the created mask and saturation filter mask to the Landsat satellite image data to obtain a corresponding Landsat cloud-free image; According to the area of ​​Landsat cloud-free images, the cloud-free area ratio and cloud pollution area ratio are obtained, and the filtering threshold of Landsat cloud-free images is adjusted according to the cloud pollution area ratio; According to the adjusted Landsat cloud-free image, similar pixels are identified using an adaptive moving window and a preset threshold, and the spectral values ​​of cloud-covered pixels are predicted based on a weighted linear model using similar pixels adjacent in space and time, so as to effectively remove the cloud amount in the Landsat cloud-free image, thereby obtaining cloud-free and clear Landsat satellite image data; The multi-source data fusion unit is used to fuse and reconstruct the cloud-free and clear Landsat satellite image data with the MODIS image data after cloud threshold screening by adopting a regression-based spatiotemporal fusion model, so as to obtain fused image data with less cloud and uniform distribution over time; wherein the regression-based spatiotemporal fusion model integrates partial least squares regression PLSR into the regression-based spatiotemporal fusion method, and the multi-source data fusion unit is specifically: Based on PLSR, the regression-based spatiotemporal fusion model is established and trained using MODIS image data filtered by cloud threshold, and the spatiotemporal fusion model corresponding to the regression coefficient under the optimal PLSR component is selected to obtain the trained spatiotemporal fusion model. Inputting the cloud-free and clear Landsat satellite image data into the trained regression-based spatiotemporal fusion model in sequence to predict the Landsat image at the next moment; Since the spatial resolution of MODIS images and Landsat images is quite different, adaptive residual compensation is used to optimize the error in the fusion process to obtain PLSR predicted Landsat images; The PLSR predicted Landsat image is merged with the cloud-free and clear Landsat satellite image data to obtain fused image data with less cloud cover, uniform distribution over time, and full regional coverage; The spectral index curve reconstruction unit is used to extract wetland vector data from the multi-layer coastal wetland map in layers according to wetland type; based on Google Earth Engine, a number of random points are randomly selected from the selected wetland type vector data, and the corresponding spectral index threshold is obtained by calculating the mean and standard deviation of the spectral index of the random points; and then the spectral index curve is reconstructed using a time linear interpolation method; A tidal flat distinguishing unit, used to identify the tidal flat based on the fused image data and the spectral index curve, using the dynamic change of tide level and the feature of no vegetation, and obtain a tidal flat distinguishing image; The classification result unit is used to distinguish the evergreen mangroves and the deciduous salt marshes according to the tidal flat distinction image and the spectral index curve, and to obtain the final classification result of the coastal wetland image by using phenological characteristics.

Images