A multi-source remote sensing image analysis system and method for dynamic monitoring of wetlands

Abstract

The application provides a multi-source remote sensing image analysis system and method for dynamic monitoring of wetlands. The system determines the seasonal timing of the wetland vegetation and water features from a multi-source timing image set, extracts spatial texture and geometric features from the multi-spectral image, and performs double-path change detection on the timing features, spatial texture and geometric features to obtain evolution patterns and heterogeneous structures. The system determines the change intensity map from the evolution patterns and heterogeneous structures, extracts multiple change map patches from the change intensity map according to a conversion rule library, filters and reclassifies all change map patches to obtain the spatial distribution of the wetlands and multiple types of conversion events, and integrates all conversion events and the spatial distribution of the wetlands in space and time to generate a hotspot area of wetland degradation. The application can strip the complex spectral features in multi-source remote sensing data and accurately distinguish real land class conversion, thereby achieving high-precision identification and positioning of key changes in wetlands.

Description

Technical Field

[0001] This application relates to the field of dynamic monitoring technology, and more specifically, to a multi-source remote sensing image analysis system and method for dynamic monitoring of wetlands. Background Technology

[0002] Wetlands are important national ecological resources and the "kidneys of the earth." Their dynamic changes directly affect regional ecological security, water resource balance, and biodiversity. Utilizing remote sensing technology to dynamically monitor wetlands is a key means to achieve wetland resource protection and scientific management.

[0003] Existing wetland remote sensing monitoring methods mainly rely on single-type (such as optical) remote sensing images or simple combinations of multi-source data. Changes are identified by analyzing changes in vegetation and water indices or by comparing images after classification. However, optical images are easily affected by clouds and fog, making it difficult to obtain continuous time-series data and penetrate the vegetation canopy to directly reflect surface moisture conditions. On the other hand, wetland surface cover is complex, and the spectral characteristics of vegetation and water are easily confused. In addition, strong seasonal and interannual fluctuations make it difficult for traditional methods to accurately distinguish true land cover transitions. The change detection results have a high false alarm rate and blurred boundaries. Furthermore, they fail to deeply integrate the unique eco-hydrological processes and land cover transition logic of wetlands, resulting in inaccurate identification of key processes such as wetland degradation. Therefore, how to extract the complex spectral features from multi-source remote sensing data and accurately distinguish true land cover transitions to achieve high-precision identification and location of key wetland changes has become a challenge for the industry. Summary of the Invention

[0004] This application provides a multi-source remote sensing image analysis system and method for dynamic monitoring of wetlands, which can strip away complex spectral features from multi-source remote sensing data and accurately distinguish real land cover transformations, thereby achieving high-precision identification and location of key changes in wetlands.

[0005] In a first aspect, this application provides a multi-source remote sensing image analysis method for dynamic monitoring of wetlands, comprising the following steps:

[0006] Synthetic aperture radar and multispectral images of the target wetland were acquired simultaneously to obtain a multi-source time-series image set of the wetland in a spatiotemporal coordination.

[0007] Based on the multi-source temporal image set, the seasonal phenological characteristics of wetland vegetation and water are determined, and the spatial texture and geometric features inside the wetland are extracted from the multispectral images.

[0008] By performing dual-path change detection on the temporal phenological features and the spatial texture and geometric features, the seasonal evolution pattern and boundary heterogeneous structure of wetlands are obtained, and then the change intensity map of wetlands is determined through the evolution pattern and the heterogeneous structure.

[0009] A wetland type conversion rule base is obtained, and multiple change patches of the change intensity map are extracted according to the conversion rule base. All change patches are judged, filtered and reclassified to obtain the spatial distribution of wetlands and multiple types of conversion events.

[0010] All transformation events and the spatial distribution of wetlands are spatiotemporally integrated and visualized to generate hotspots of wetland degradation.

[0011] In some embodiments, the simultaneous acquisition of synthetic aperture radar imagery and multispectral imagery of the target wetland to obtain a spatiotemporally coordinated multi-source time-series image set of the wetland specifically includes:

[0012] Synthetic aperture radar and multispectral images covering the target wetland are acquired simultaneously within a preset monitoring period.

[0013] The synthetic aperture radar image is subjected to noise suppression and moisture enhancement.

[0014] Atmospheric correction and cloud masking were performed on the multispectral images;

[0015] The processed synthetic aperture radar image and multispectral image were spatiotemporally registered to obtain a multi-source time-series image set of wetland spatiotemporally coordinated.

[0016] In some embodiments, determining the seasonal phenological characteristics of wetland vegetation and water based on the multi-source temporal image set specifically includes:

[0017] The spectral index is calculated pixel-by-pixel from the multispectral image sequence to obtain the spectral index time series.

[0018] The backscattering coefficients were extracted from the synthetic aperture radar images to obtain the time series of changes in surface moisture and roughness.

[0019] The start and end dates, peak values, amplitudes, and seasonal integral features of the growing season within the hydrological year are extracted from the spectral index time series. The seasonal mean and variation range features are extracted from the time series to obtain the time series phenological features.

[0020] In some embodiments, extracting spatial texture and geometric features of the wetland interior from the multispectral image specifically includes:

[0021] The red, green, blue, and near-infrared bands with the highest spatial resolution in the multispectral image were selected.

[0022] Determine the spatial texture features within the wetland based on all wavebands;

[0023] The multispectral image is divided into multiple homogeneous regions;

[0024] The area, perimeter, aspect ratio, and shape index of each homogeneous region are determined as the geometric features of the wetland interior.

[0025] In some embodiments, performing dual-path change detection on the temporal phenological features and the spatial texture and geometric features to obtain the seasonal evolution pattern of wetlands and the heterogeneous structure of their boundaries specifically includes:

[0026] Determine the temporal feature differences of the temporal phenological features across all pixels between the two temporal phases;

[0027] The pseudo-image sequence is determined based on all temporal feature differences;

[0028] The evolution pattern of wetlands within seasonal intervals in the pseudo-image sequence is captured by the gated loop unit of the first detection path, thereby obtaining the seasonal evolution pattern of wetlands.

[0029] Determine the differences in texture features of all pixels within the dual temporal phase and neighborhood window for the spatial texture and geometric features;

[0030] The convolutional neural network of the second detection path extracts the structural changes of wetland boundary contour and internal texture in all texture feature differences, thus obtaining the heterogeneous structure of the wetland boundary.

[0031] In some embodiments, determining the wetland change intensity map through the evolution pattern and the heterogeneous structure specifically includes:

[0032] Based on the evolution pattern and the heterogeneous structure, multiple joint feature vectors of the multi-source remote sensing image are determined;

[0033] All joint feature vectors are mapped to change confidence levels to obtain a wetland change intensity map.

[0034] In some embodiments, extracting multiple change patches from the change intensity map based on the transformation rule base specifically includes:

[0035] Identify multiple potential change pixels in the change intensity map;

[0036] Based on all the potential change cells, multiple initial change patches are formed that are connected;

[0037] Based on the transformation rule base, the transformed features are extracted from all the initial transformed features.

[0038] In some embodiments, the determination, filtering, and reclassification of all changed patches to obtain the spatial distribution of wetlands and multiple types of transformation events specifically include:

[0039] Identify multiple land cover types for all changed patches over biphasic time;

[0040] The land use type of all changed land parcels is determined and filtered using the aforementioned conversion rule base, and the corresponding changed land parcels are marked.

[0041] All marked change patches are classified to obtain the spatial distribution of wetlands and multiple types of transformation events.

[0042] In some embodiments, spatiotemporal integration and visualization of all transformation events and the spatial distribution of wetlands are performed to generate wetland degradation hotspots, specifically including:

[0043] The transformation events from multiple monitoring periods are overlaid with the spatial distribution to calculate the total frequency of degradation-type transformations within each spatial unit.

[0044] Smoothing and clustering analysis are performed on the spatial units of total frequency to identify spatial clustering regions.

[0045] By overlaying the identified clusters with the boundaries of nature reserves and the intensity of human activities, high-risk wetland degradation hotspots are delineated.

[0046] Secondly, this application provides a multi-source remote sensing image analysis system for dynamic monitoring of wetlands, comprising:

[0047] The acquisition module is used to simultaneously acquire synthetic aperture radar images and multispectral images of the target wetland to obtain a multi-source time-series image set of the wetland in a spatiotemporal coordination.

[0048] The processing module is used to determine the seasonal phenological characteristics of wetland vegetation and water based on the multi-source temporal image set, and to extract the spatial texture and geometric features of the wetland from the multispectral images.

[0049] The processing module is also used to perform dual-path change detection on the temporal phenological features and the spatial texture and geometric features to obtain the seasonal evolution pattern and the heterogeneous structure of the boundary of the wetland, and then determine the change intensity map of the wetland through the evolution pattern and the heterogeneous structure.

[0050] The processing module is also used to obtain a wetland type conversion rule base, extract multiple change patches from the change intensity map according to the conversion rule base, and perform judgment filtering and reclassification on all change patches to obtain the wetland spatial distribution and multiple types of conversion events;

[0051] The execution module is used to perform spatiotemporal integration and visualization of all transformation events and the spatial distribution of wetlands, and to generate hotspot areas of wetland degradation.

[0052] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0053] This application provides a multi-source remote sensing image analysis system and method for dynamic monitoring of wetlands. First, synthetic aperture radar (SAR) images and multispectral images of the target wetland are simultaneously acquired to obtain a spatiotemporally coordinated multi-source time-series image set. Based on the multi-source time-series image set, the seasonal phenological characteristics of wetland vegetation and water content are determined. Spatial texture and geometric features within the wetland are extracted from the multispectral images. Dual-path change detection is performed on the seasonal phenological characteristics and the spatial texture and geometric features to obtain the seasonal evolution pattern and heterogeneous structure of the wetland boundaries. Then, the change intensity map of the wetland is determined through the evolution pattern and the heterogeneous structure. A wetland type conversion rule base is obtained. Multiple change patches are extracted from the change intensity map based on the conversion rule base. All change patches are judged, filtered, and reclassified to obtain the spatial distribution of the wetland and multiple types of conversion events. All conversion events and the spatial distribution of the wetland are spatiotemporally integrated and visualized to generate hotspot areas of wetland degradation.

[0054] Therefore, in the process of a multi-source remote sensing image analysis method for wetland dynamic monitoring, this application firstly determines the seasonal phenological characteristics of wetland vegetation and water based on the multi-source time-series image set, and extracts the spatial texture and geometric features of the wetland interior from the multispectral images; performs dual-path change detection on the seasonal phenological characteristics and the spatial texture and geometric features to obtain the seasonal evolution pattern and the heterogeneous structure of the wetland boundary, and then determines the change intensity map of the wetland through the evolution pattern and the heterogeneous structure, wherein the change intensity map is a probability distribution map that initially identifies the change area and reflects the probability of change from the first time phase to the second time phase in pixels; secondly, obtains a wetland type conversion rule base, extracts multiple change patches from the change intensity map based on the conversion rule base, and performs judgment filtering and reclassification on all change patches to obtain the wetland spatial distribution and multiple types of conversion events; and performs spatiotemporal integration and visualization on all conversion events and the wetland spatial distribution to generate wetland degradation hotspot areas. The above scheme can strip away the complex spectral features in multi-source remote sensing data and accurately distinguish the real land cover transformation, thereby achieving high-precision identification and location of key changes in wetlands. Attached Figure Description

[0055] Figure 1 This is an exemplary flowchart of a multi-source remote sensing image analysis method for dynamic monitoring of wetlands, as shown in some embodiments of this application.

[0056] Figure 2 This is an exemplary flowchart illustrating the determination of temporal phenological characteristics according to some embodiments of this application;

[0057] Figure 3 This is a schematic diagram illustrating the effect of wetland change and degradation hotspot identification according to some embodiments of this application;

[0058] Figure 4 This is a schematic diagram of the structure of a multi-source remote sensing image analysis system for dynamic monitoring of wetlands, as shown in some embodiments of this application.

[0059] Figure 5 This is a schematic diagram of the structure of a computer device for implementing a multi-source remote sensing image analysis method for dynamic monitoring of wetlands, according to some embodiments of this application. Detailed Implementation

[0060] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0061] refer to Figure 1 The figure is an exemplary flowchart of a multi-source remote sensing image analysis method for dynamic monitoring of wetlands, according to some embodiments of this application. This method mainly includes the following steps:

[0062] In step 101, synthetic aperture radar images and multispectral images of the target wetland are acquired simultaneously to obtain a multi-source time-series image set of the wetland in a spatiotemporal coordination.

[0063] In some embodiments, the simultaneous acquisition of synthetic aperture radar imagery and multispectral imagery of the target wetland to obtain a spatiotemporally coordinated multi-source time-series image set of the wetland can be achieved through the following steps:

[0064] Synthetic aperture radar and multispectral images covering the target wetland are acquired simultaneously within a preset monitoring period.

[0065] The synthetic aperture radar image is subjected to noise suppression and moisture enhancement.

[0066] Atmospheric correction and cloud masking were performed on the multispectral images;

[0067] The processed synthetic aperture radar image and multispectral image were spatiotemporally registered to obtain a multi-source time-series image set of wetland spatiotemporally coordinated.

[0068] In practice, the synchronous acquisition of synthetic aperture radar (SAR) images and multispectral images covering the target wetland within a preset monitoring period can be achieved in the following ways: A fixed monitoring frequency (e.g., once every six months) is determined based on the wetland hydrological and vegetation phenological cycles; within the same hydrological year, SAR and multispectral images covering the target wetland and with imaging dates as close as possible are queried and downloaded through a satellite data platform; each pair of images acquired is used as a basic observation unit; noise suppression and moisture information enhancement of the SAR images can be achieved in the following ways: First, the original SAR images are filtered using a Refined Lee filter to suppress inherent speckle noise; then, terrain correction is performed using a digital elevation model to eliminate the influence of terrain undulations on the signal, thereby obtaining radar images with enhanced surface moisture and roughness information.

[0069] In specific implementation, atmospheric correction and cloud masking of the multispectral image can be achieved in the following way: First, atmospheric correction is performed on the multispectral image using an atmospheric correction processor such as Sen2Cor, converting apparent radiance into surface reflectance; then, using the cloud detection probability band inherent in the image, a threshold is set (e.g., cloud probability greater than 20%), and all pixels above this threshold are marked as clouds or cloud shadows. These marked pixels are then masked out of the analysis to obtain a high-quality surface reflectance image with no or few clouds; the processed synthetic aperture radar image and multispectral image are then compared with the cloud masking. Spatiotemporal registration of spectral images to obtain a spatiotemporally coordinated multi-source time-series image set of wetlands can be achieved in the following way: using a high-resolution multispectral image that has undergone geometrical fine correction in a certain period as a spatial reference benchmark, a registration method based on feature points (such as SIFT features) is used to match all other temporal multispectral images and synthetic aperture radar images with the spatial reference benchmark to complete spatial geometric registration and ensure that the same geographic coordinates are aligned on all images; then, all images are arranged in chronological order to construct a multi-source time-series image set that has been coordinated in both time and space.

[0070] It should be noted that the multi-source time-series image set in this application is an image acquisition constraint that ensures the temporal comparability of data from different sensors and constitutes an effective observation sequence. It is a standardized data basis that has undergone radiometric and geometric normalization processing for subsequent unified feature extraction and change detection.

[0071] In step 102, the seasonal phenological characteristics of wetland vegetation and water are determined based on the multi-source temporal image set, and the spatial texture and geometric features inside the wetland are extracted from the multispectral images.

[0072] In some embodiments, reference Figure 2As shown, this figure is an exemplary flowchart for determining temporal phenological characteristics in some embodiments of this application. In this embodiment, determining the seasonal phenological characteristics of wetland vegetation and water based on the multi-source temporal image set can be achieved by the following steps:

[0073] First, in step 1021, the index is calculated pixel by pixel in the multispectral image sequence to obtain the spectral index time series;

[0074] Secondly, in step 1022, the backscattering coefficient is extracted from the synthetic aperture radar image to obtain the time series of surface moisture and roughness changes;

[0075] Finally, in step 1023, the start and end dates, peak values, amplitudes, and seasonal integral features of the growing season within the hydrological year are extracted from the spectral index time series, and the seasonal mean and variation amplitude features are extracted from the time series, thereby obtaining the time series phenological features.

[0076] In specific implementation, the pixel-by-pixel calculation of the index to obtain the spectral index time series can be achieved in the following way: For each pixel in the multispectral image, the normalized differential vegetation index is calculated using its reflectance values ​​in the near-infrared and red bands; simultaneously, the normalized differential water index is calculated using its reflectance values ​​in the green and near-infrared bands; secondly, the differential vegetation index and differential water index calculated for each pixel in all temporal images are arranged in chronological order to form the time series of the differential water index and the time series of the differential vegetation index for that pixel; the set of the above two time series for all pixels is taken as the spectral index time series.

[0077] In specific implementation, the backscattering coefficients extracted from the synthetic aperture radar image to obtain the time series of surface moisture and roughness changes can be achieved in the following way: read the backscattering coefficient of each pixel in the synthetic aperture radar image; for each pixel, arrange its backscattering coefficient values ​​in all temporal images in chronological order to form the backscattering coefficient time series of that pixel; and use the set of backscattering coefficient time series of all pixels as the time series of surface moisture and roughness changes.

[0078] In specific implementation, the following methods can be used to extract the start and end dates, peak values, amplitudes, and seasonal integral features of the growing season within a hydrological year from the spectral index time series, and extract the seasonal mean and variation range features from the time series to obtain the time-series phenological features: First, the spectral index time series of each pixel is smoothed using a Savitzky-Golay filter. Then, the start and end dates of the vegetation growing season within each hydrological year are identified using a dynamic thresholding method. Next, the extreme values ​​and amplitudes of the differential vegetation index within the growing season are calculated, as well as the seasonal integral value obtained by integrating the time series of the spectral index within the growing season. Simultaneously, the average value (seasonal mean) and standard deviation (variation range) of the backscattering coefficient time series of the pixel within a complete hydrological year are calculated. Finally, the vector composed of the extracted start and end dates of the growing season, the peak value of the time series, the amplitude of the time series, the seasonal integral of the differential vegetation index, the seasonal mean of the backscattering coefficient, and the variation range of the backscattering coefficient is used as the phenological vector of the pixel, and the set of all phenological vectors is used as the time-series phenological features.

[0079] It should be noted that the spectral index time series in this application refers to the parameter values ​​of the spectral indicators that quantify the growth status of wetland vegetation and the surface water content over time; the variation amplitude characteristics are key quantitative indicators that condense the original time series curve into typical phenological stages and seasonal dynamics of wetlands.

[0080] In some embodiments, extracting spatial texture and geometric features of the wetland interior from the multispectral image can be achieved using the following steps:

[0081] The red, green, blue, and near-infrared bands with the highest spatial resolution in the multispectral image were selected.

[0082] Determine the spatial texture features within the wetland based on all wavebands;

[0083] The multispectral image is divided into multiple homogeneous regions;

[0084] The area, perimeter, aspect ratio, and shape index of each homogeneous region are determined as the geometric features of the wetland interior.

[0085] In specific implementation, selecting the red, green, blue, and near-infrared bands with the highest spatial resolution in the multispectral image can be achieved in the following way: read the combination of bands with the highest spatial resolution among all bands in the multispectral image; among them, the red, green, blue, and near-infrared bands have the highest spatial resolution of 10 meters; thereby extracting the reflectance data of the above four bands in the multispectral image; determining the spatial texture features inside the wetland based on all bands can be achieved in the following way: for each pixel in the multispectral image, define a neighborhood window of a fixed size (e.g., a 7x7 pixel) centered on it; within this window, for the near-infrared band, calculate the gray-level co-occurrence matrix, and further calculate the three texture metrics: contrast, homogeneity, and entropy; where contrast reflects the clarity of the texture, homogeneity reflects local uniformity, and entropy reflects the complexity; finally, the set of contrast values, homogeneity values, and entropy values ​​calculated for all pixels is taken as the spatial texture features inside the wetland.

[0086] In specific implementation, dividing the multispectral image into multiple homogeneous regions can be achieved in the following way: A multi-resolution segmentation algorithm is used to process the multispectral image. This algorithm, based on spectral heterogeneity and shape heterogeneity criteria, merges adjacent pixels from bottom to top. The size and regularity of the generated objects are controlled by setting scale parameters, shape weights, and compactness weights. For example, a larger scale parameter is set to generate objects that can cover homogeneous vegetation patches or water bodies in wetlands. Finally, the output divides the entire image into several internally homogeneous polygonal regions with clear boundaries, and these polygonal regions are all considered homogeneous regions. Each homogeneous region is then determined... The area, perimeter, aspect ratio, and shape index of a region can be used as geometric features of the wetland interior in the following way: For each homogeneous polygon, under a given map projection coordinate system (such as the universal transverse Mercator projection or the Gauss-Kruger projection coordinate system), calculate its actual surface area, boundary perimeter, and the ratio of the long side to the short side of the smallest bounding rectangle to obtain the aspect ratio; in addition, calculate the ratio of its perimeter to the circumference of a circle of the same area as the shape index to measure the degree of irregularity of the boundary; finally, the set of area, perimeter, aspect ratio, and shape index calculated for each homogeneous region is taken as the geometric features of the wetland interior.

[0087] It should be noted that spatial texture features are a set of metrics used to help distinguish different land features within wetlands and describe the heterogeneity and regularity of spectral spatial distribution within the neighborhood of a pixel; homogeneous regions are image objects with ecological significance that aggregate adjacent pixels with similar spectral and spatial features in an image to calculate geometric features; and geometric features within wetlands are a set of spatial morphological attributes that quantify the shape, size, and boundary complexity of wetland land feature patches.

[0088] In step 103, dual-path change detection is performed on the temporal phenological features and the spatial texture and geometric features to obtain the seasonal evolution pattern and the heterogeneous structure of the wetland boundary. Then, the change intensity map of the wetland is determined by the evolution pattern and the heterogeneous structure.

[0089] In some embodiments, the following steps can be used to perform dual-path change detection on the temporal phenological features and the spatial texture and geometric features to obtain the seasonal evolution pattern and heterogeneous structure of wetland boundaries:

[0090] Determine the temporal feature differences of the temporal phenological features across all pixels between the two temporal phases;

[0091] The pseudo-image sequence is determined based on all temporal feature differences;

[0092] The evolution pattern of wetlands within seasonal intervals in the pseudo-image sequence is captured by the gated loop unit of the first detection path, thereby obtaining the seasonal evolution pattern of wetlands.

[0093] Determine the differences in texture features of all pixels within the dual temporal phase and neighborhood window for the spatial texture and geometric features;

[0094] The convolutional neural network of the second detection path extracts the structural changes of wetland boundary contour and internal texture in all texture feature differences, thus obtaining the heterogeneous structure of the wetland boundary.

[0095] In specific implementation, determining the temporal characteristic differences of all pixels between two time phases of the temporal phenological characteristics can be achieved in the following way: From a multi-source temporal image set, select two images at different times as comparison benchmarks according to the monitoring target, namely, time phase T1 and time phase T2. For each pixel, obtain its temporal phenological characteristic vector in time phase T1 and its temporal phenological characteristic vector in time phase T2, respectively; calculate the absolute difference of the corresponding dimension values ​​of these two characteristic vectors to generate the difference feature vector of each pixel, that is, the temporal characteristic difference of the pixel; each dimension value in the difference feature vector represents the change of the pixel in specific phenological indicators such as vegetation growth peak, growing season length, and mean backscattering; according to All temporal feature differences can be used to determine the pseudo-image sequence in the following way: Arrange the temporal feature differences of each pixel according to the original geographical location of the pixel; fill all pixel values ​​of the first dimension of the temporal feature differences (e.g., NDVI peak difference) into an empty matrix with the same spatial range as the original image to form the first layer of difference image; process each dimension of the temporal feature differences in turn to generate multi-layer difference images; stack the multi-layer difference images in the channel dimension to form a multi-channel pseudo-image sequence; wherein, the pseudo-image sequence maintains the geographical pattern of the wetland in space, and each channel represents the difference of a phenological index, which can be used as input for subsequent sequence model analysis.

[0096] In specific implementation, the evolution pattern of wetlands within seasonal intervals in the pseudo-image sequence is captured by the gated recurrent unit of the first detection path, thereby obtaining the seasonal evolution pattern of wetlands. This can be achieved in the following way: Construct the first detection path, the core of which is a gated recurrent unit network; First, input the data of each channel of the pseudo-image sequence into the gated recurrent unit network along the time dimension. This network, through its internal gating mechanism, memorizes and integrates the dependencies between different difference indicators, simulating the transition process of the phenological state of wetlands in each pixel from phase T1 to T2; Finally, the network outputs a fixed-length feature vector for each pixel. This vector encodes the overall change pattern of the phenological characteristics of each pixel (such as macro trends such as "slow aridification" and "rapid flooding"). The set of feature vectors of the phenological characteristics of each pixel is taken as the seasonal evolution pattern of wetlands.

[0097] It should be noted that the gated loop cell network in this application consists of at least one stacked gated loop cell layer, and each cell contains an update gate and a reset gate mechanism. The network learns and memorizes the dependencies and transition logics between different phenological difference indicators through its internal gating mechanism, thereby simulating the possible evolution path of the ecological environment status (such as water conditions and vegetation vitality) of each wetland pixel within the seasonal interval. The network takes a pseudo-image sequence as input, where each channel of the pseudo-image sequence represents the difference value of a phenological indicator (such as peak vegetation growth, growing season length, etc.) between two consecutive time phases. The entire pseudo-image constitutes an input data volume that is spatially fully covered and contains multiple indicator differences in the channel dimension. Finally, the network extracts the data and transforms and compresses it through a fully connected layer, ultimately outputting a fixed-length feature vector for each pixel. The training process uses a cross-entropy loss function and an adaptive moment estimation optimizer, with an initial learning rate of 0.001. By training on wetland pixel data and associating the network's output feature vector with known change types (such as "droughtification," "flooding," and "vegetation recovery" as interpreted by experts), the network learns to map complex original differences into abstract patterns with clear phenological semantics.

[0098] In specific implementation, the determination of the differences in texture features of all pixels in the spatial texture and geometric features between the two temporal phases and within the neighborhood window can be achieved in the following way: For time phases T1 and T2, the spatial texture features (contrast, homogeneity, entropy) of each pixel and the geometric features (area, shape index, etc.) of its homogeneous region are obtained respectively; secondly, the absolute difference of the texture features of each pixel across the temporal phases is calculated, and the mean of the texture features of all pixels in its neighborhood window (e.g., 5x5 pixels) in time phase T2 is calculated with the texture feature value of the central pixel in time phase T1, thereby capturing the structural changes of the local context; the difference between the difference between the pixel's own cross-temporal phase difference and the local context difference is combined to form the texture feature difference of the pixel in spatial texture and geometric features.

[0099] In specific implementation, the structural changes of wetland boundary contours and internal textures are extracted from all texture feature differences through the convolutional neural network of the second detection path. The heterogeneous structure of the wetland boundary can be obtained in the following way: Construct a second detection path, the core of which is a convolutional neural network; rearrange the texture feature differences of each pixel into a multi-channel difference image corresponding to the image space, and input it into the convolutional neural network; the network extracts features from the local neighborhood step by step through multi-layer convolution and pooling operations, thereby capturing the changes in the edge contours and internal texture heterogeneity of the wetland caused by boundary erosion, vegetation fragmentation, or newly formed water bodies; the convolutional neural network finally outputs a fixed-length feature vector for each pixel. This vector represents the subtle structural changes in the local area where the pixel is located. The set of feature vectors output by each pixel is taken as the heterogeneous structure of the wetland boundary.

[0100] It should be noted that the convolutional neural network in this application adopts an encoder structure, which includes four feature extraction stages consisting of convolutional layers, batch normalization layers, modified linear unit activation functions, and max pooling layers connected sequentially. Each stage's convolutional layer uses a 3x3 convolutional kernel, and the pooling layer uses a 2x2 window for downsampling, thereby gradually expanding the receptive field and abstracting multi-scale spatial features. The network receives the aforementioned multi-channel texture difference image as input, and through layer-by-layer transmission and transformation, finally outputs a fixed-length feature vector at each pixel location. This output vector encodes the structural change information contained in all texture and geometric differences within the local neighborhood centered on that pixel in the input image. During the training process, the network maps specific change patterns such as "boundary erosion" and "vegetation fragmentation" onto unique vector expressions through the optimization of a large number of labeled wetland change samples and loss functions (such as mean squared error loss). Therefore, by collecting these feature vectors output from all pixels according to their spatial locations, a heterogeneous structure of the wetland boundary that can finely characterize the contour and internal texture heterogeneity of the wetland boundary is obtained.

[0101] It should be noted that the pseudo-image sequence is a multi-channel difference image data structure with geographic reference, which is constructed by spatially rearranging phenological difference indicators as input to the temporal change detection model; the evolution pattern is a set of abstract feature vectors that comprehensively quantify the intensity of change and characterize the macroscopic change trend of wetland pixel phenological state within the seasonal interval; and the heterogeneous structure is a set of abstract feature vectors that comprehensively quantify the intensity of change and characterize the subtle changes in the local spatial structure and texture of wetland pixels in the preceding and following time phases.

[0102] In some embodiments, determining the intensity map of wetland change based on the evolution pattern and the heterogeneous structure can be achieved through the following steps:

[0103] Based on the evolution pattern and the heterogeneous structure, multiple joint feature vectors of the multi-source remote sensing image are determined;

[0104] All joint feature vectors are mapped to change confidence levels to obtain a wetland change intensity map.

[0105] In specific implementation, determining multiple joint feature vectors of the multi-source remote sensing image based on the evolution pattern and the heterogeneous structure can be achieved in the following way: For each pixel, its feature vector in the evolution pattern and the feature vector in the heterogeneous structure are concatenated end-to-end in the dimensional direction of the vectors. For example, if the length of the feature vector in the evolution pattern is 64 and the length of the feature vector in the heterogeneous structure is 64, then a new feature vector with a length of 128 is generated after concatenation, which is the joint feature vector of the pixel in the multi-source remote sensing image. The joint feature vector combines the overall change pattern of the pixel in the phenological time series and the structural variation information in the local space, comprehensively describing the changes that have occurred at the location.

[0106] In practical implementation, mapping all joint feature vectors to change confidence levels to obtain the wetland change intensity map can be achieved as follows: A multilayer perceptron model is constructed, where the number of neurons in the input layer matches the dimension of the joint feature vectors; the model contains at least one hidden layer, consisting of a fully connected layer and a nonlinear activation function (such as a rectified linear unit), used to learn and express complex nonlinear relationships between features; the output layer is a single neuron, using the sigmoid function as its activation function to ensure the final output value is constrained to between 0 and... Within the range of 1, the joint feature vector of each pixel is input into the multilayer perceptron. The multilayer perceptron learns the complex mapping relationship from joint features to change confidence through its fully connected layers and nonlinear activation functions. The activation value of the neuron output (constrained between 0 and 1 by the Sigmoid function) is the change confidence of that pixel. The closer the change confidence is to 1, the greater the probability that the pixel will change from phase T1 to phase T2. The change confidence of all pixels is arranged into a single-band grayscale image according to their geographical location, which is the change intensity map of the wetland.

[0107] It should be noted that the joint feature vector is a comprehensive high-dimensional feature expression that integrates macroscopic evolution patterns and microscopic heterogeneous structural information to determine the certainty of change; the change intensity map is a probability distribution map that initially identifies the change area and reflects the probability of change from the first time phase to the second time phase in pixels; among them, wetland change is manifested as both slow phenological seasonal evolution and sudden spatial structural anomalies (such as drastic boundary changes caused by dam breaches and vegetation patch fragmentation); traditional single methods are difficult to accurately capture both types of signals at the same time; the first path uses gated recurrent units to model the pseudo-image sequence composed of temporal phenological feature differences, specifically learning the continuous evolution law of wetland pixel state at the seasonal scale, thereby effectively separating normal seasonal fluctuations; the second path uses convolutional neural networks to perceive local differences in spatial texture and geometric features, focusing on capturing subtle structural changes in the boundary contours and internal textures within the pixel neighborhood, thereby extracting fine heterogeneous structures, enabling the model to understand change from both macroscopic laws and microscopic evidence levels, significantly improving the ability to distinguish between real changes and natural fluctuations in complex wetland scenarios.

[0108] In step 104, a wetland type conversion rule base is obtained, and multiple change patches of the change intensity map are extracted according to the conversion rule base. All change patches are judged, filtered and reclassified to obtain the wetland spatial distribution and multiple types of conversion events.

[0109] It should be noted that the wetland type conversion rule base is a set of logical criteria built on wetland ecology, hydrology, and vegetation succession laws; its fundamental function is to distinguish the authenticity and classify the nature of wetland change events; the content of this rule base mainly includes two categories of rules: unreasonable conversion rules: used to filter out high-probability pseudo-changes; reasonable conversion rules: used to qualitatively characterize real changes, such as degradation, restoration, and stabilization.

[0110] In some embodiments, extracting multiple change patches from the change intensity map based on the transformation rule base can be achieved using the following steps:

[0111] Identify multiple potential change pixels in the change intensity map;

[0112] Based on all the potential change cells, multiple initial change patches are formed that are connected;

[0113] Based on the transformation rule base, the transformed features are extracted from all the initial transformed features.

[0114] In specific implementation, determining multiple potential changed pixels in the intensity map can be achieved in the following way: the intensity map is automatically segmented using the Otsu thresholding method. This method treats the change confidence values ​​of all pixels in the intensity map as a dataset, and automatically finds an optimal segmentation threshold by calculating and maximizing the inter-class variance between potentially changed and unchanged pixels in the data. The distinction criteria are: all pixels with change confidence values ​​higher than this automatically calculated threshold are considered as potentially changed pixels and their change confidence is marked as 1 in the intensity map; all pixels with change confidence values ​​lower than or equal to this threshold are considered as unchanged pixels and marked as 0. Thus, a grayscale intensity map is transformed into a binary image containing only 0 and 1 values, where all pixels marked as 1 constitute multiple potential changed pixels.

[0115] In practice, the formation of multiple initial change patches based on all potential change pixels can be achieved in the following way: First, perform morphological closing operation on the binary image. The closing operation first performs dilation to connect neighboring discrete potential change pixels and fill small gaps. Then, perform erosion to restore the approximate object boundary and remove small, isolated noise points. Finally, perform connected component analysis on the processed binary image to mark all spatially connected pixels with a value of 1 as the same independent region, thus obtaining an initial change patch.

[0116] In specific implementation, extracting change patches from all initial change patches based on the aforementioned transformation rule base can be achieved in the following way: A wetland type transformation rule base is predefined, containing a series of unreasonable transformation rules, such as "direct conversion from permanent water bodies to bare land (non-seasonal drying)" being considered highly suspicious; each initial change patch is overlaid onto a dual-temporal wetland type data layer to obtain the dominant pre-change and post-change types within the patch; this data is generated quickly from extracted temporal phenological, spatial texture, and geometric multi-dimensional features, and shares the same feature base and spatial registration relationship as the change intensity map, through a lightweight classifier (such as random forest) or a classification branch in a dual-path network; this type transformation pair is compared with entries in the rule base; if any unreasonable transformation rule is matched, the initial change patch is determined to be a pseudo-change and filtered out; all initial change patches that do not match unreasonable rules or conform to reasonable ecological succession rules are retained as change patches.

[0117] It should be noted that potential change pixels are a set of pixels that may change, initially extracted from the change intensity map; initial change patches are preliminary change areas formed by the aggregation of spatially adjacent potential change pixels for rule filtering; and change patches are the change areas to be classified that are retained after preliminary filtering by ecological rationality rules for final determination.

[0118] In some embodiments, the following steps can be used to filter and reclassify all changed patches to obtain the spatial distribution of wetlands and multiple types of transformation events:

[0119] Identify multiple land cover types for all changed patches over biphasic time;

[0120] The land use type of all changed land parcels is determined and filtered using the aforementioned conversion rule base, and the corresponding changed land parcels are marked.

[0121] All marked change patches are classified to obtain the spatial distribution of wetlands and multiple types of transformation events.

[0122] In practice, determining the multiple land use types of all changed patches in two phases of time can be achieved as follows: For each changed patch to be identified, its polygon boundaries are superimposed onto high-precision wetland type data layers of phases T1 and T2, respectively. Then, the wetland type with the highest frequency of occurrence within the polygon range of the changed patch in the phase T1 wetland type data layer is identified as the pre-change land use type of the changed patch. Similarly, the wetland type with the highest frequency of occurrence in the phase T2 wetland type data layer is identified as the post-change land use type of the changed patch. For example, a patch might be classified as... in phase T1. "Herbaceous marshland" was classified as "exposed mudflats" in phase T2. The land use type of all changed patches was determined and filtered using the aforementioned transformation rule base, and the corresponding changed patches were marked. This was achieved by: comprehensively comparing the <pre-change land use type, post-change land use type> combination of each changed patch with the rule base; if a "degradation transformation" rule was matched, the changed patch was marked as "confirmed degradation transformation"; if a "restoration transformation" rule was matched, it was marked as "confirmed restoration transformation"; if an "unreasonable transformation" rule was matched, it was marked as "pseudo-change" and removed; if no predefined rule was matched, it was marked as "change pending verification".

[0123] In practice, classifying all marked changed patches to obtain the wetland spatial distribution and multiple types of transformation events can be achieved in the following way: Based on the determination nature of the changed patches, i.e., "confirmed degradation change" or "confirmed restoration change," assign a corresponding type code to each changed patch. Then, vectorize all changed patches to generate a vector polygon layer. In the attribute table of this layer, record the unique identifier, the dominant land use type before the change, the dominant land use type after the change, the change nature (degradation / restoration), and the actual changed area calculated from the polygon area for each polygon (i.e., each changed patch). Use these records as the transformation events for the corresponding changed patch type. Merge all changed patches marked "confirmed degradation change" into one set, and merge all changed patches marked "confirmed restoration change" into another set. Perform spatial fusion processing on these two sets to eliminate redundant boundaries between adjacent patches and reconstruct two independent vector polygon layers. One layer represents the spatial distribution of degraded wetlands, and the other layer represents the spatial distribution of restored wetlands, thus obtaining the wetland spatial distribution.

[0124] In some embodiments, reference Figure 3As shown, this figure is a schematic diagram of the effect of wetland change and degradation hotspot identification in some embodiments of this application, used to demonstrate the complete processing flow of multi-source remote sensing wetland dynamic monitoring; the upper left sub-figure is the original remote sensing image of the wetland at time T1, and the lower left sub-figure is the original remote sensing image of the wetland at time T2; the black solid lines in the images represent the outlines of stable land features such as rivers, and the black dashed lines represent the wetland boundaries, serving as the original input data for time series comparison. The middle section shows the algorithm processing flow, which sequentially performs preprocessing, feature extraction, change detection, and patch selection. After patch selection, a dual-path branch is formed: a temporal path and a spatial path. The temporal path extracts the wetland temporal phenological evolution pattern based on GRU, while the spatial path extracts the spatial heterogeneous structure of the wetland. The results of the two analyses are then fused. The right side shows the algorithm output image. In the result image, the black solid lines represent the boundaries of land features such as rivers, and the black dashed lines represent the wetland boundaries. The diagonally filled areas are the wetland real change patches identified by the dual-path fusion. Solid black dots represent identified wetland degradation areas. Hollow circles are pseudo-change pixels removed in the patch selection stage. Pseudo-changes are caused by interference such as clouds, fog, and short-term surface moisture fluctuations. This image fully presents the principle of inputting two periods of original images, processing them through a dual-path algorithm, distinguishing between real and pseudo-changes, and finally outputting a wetland change and degradation hotspot distribution map.

[0125] It should be noted that the land use types before and after the change are determined based on the transformation rule base, identifying the dominant wetland category to which the changed patches belong in the preceding and following time periods; the wetland spatial distribution and multiple types of transformation events are a vector dataset of wetland degradation and restoration change results, containing spatial geometric information and a complete attribute table, used for spatiotemporal integrated analysis.

[0126] In step 105, all transformation events and the spatial distribution of the wetlands are spatiotemporally integrated and visualized to generate hotspot areas of wetland degradation.

[0127] In some embodiments, spatiotemporal integration and visualization of all transformation events and the spatial distribution of wetlands to generate wetland degradation hotspots can be achieved through the following steps:

[0128] The transformation events from multiple monitoring periods are overlaid with the spatial distribution to calculate the total frequency of degradation-type transformations within each spatial unit.

[0129] Smoothing and clustering analysis are performed on the spatial units of total frequency to identify spatial clustering regions.

[0130] By overlaying the identified clusters with the boundaries of nature reserves and the intensity of human activities, high-risk wetland degradation hotspots are delineated.

[0131] In practice, the transformation events from multiple monitoring periods are overlaid with spatial distributions. The total frequency of degradation-type transformations within each spatial unit can be calculated as follows: A spatial distribution map layer of degraded wetlands generated from multiple consecutive monitoring periods (e.g., one period per year over the past 5 years) is overlaid and covered with a regular grid (e.g., 1 km x 1 km). (1 km), for each grid cell, the total frequency of degraded patches falling within that cell for different monitoring periods is statistically analyzed; the spatial frequency of the spatial cells is smoothed and clustered to identify spatial clusters. This can be achieved by using a kernel density estimation algorithm to perform spatial smoothing and clustering analysis on the degradation frequency values ​​of all grid cells, thereby generating a continuous kernel density surface map; continuous spatial ranges with kernel density values ​​significantly higher than the surrounding background area are identified as spatial clusters of wetland degradation; the identified clusters are overlaid with the boundaries of nature reserves and the intensity of human activities to delineate high-risk wetland degradation hotspots. This can be achieved by using the following method: the boundaries of nature reserves refer to the spatial vector boundaries of nature reserves at all levels. The vector layer of nature reserve boundaries can be directly obtained from the official ecological protection database. The layer is then subjected to coordinate projection transformation to ensure consistency with the coordinate system of the wetland analysis image, and standardized spatial data of nature reserve boundaries are extracted; human activities... Human activity intensity is a quantitative grid index characterizing the degree of human disturbance in a region. Raw spatial data from road network density maps and land use change maps are acquired and unified to the same coordinate system as the wetland analysis imagery. Using grids as statistical units, the road network density and the degree of construction land encroachment on wetlands within each statistical unit are weighted to calculate the corresponding human activity intensity value for each grid, forming a raster dataset of human activity intensity. The identified degraded cluster vector polygons are then spatially overlaid with auxiliary data such as nature reserve boundary vector layers, road network density maps, and land use change maps. For example, a high-density degraded cluster located within the core area of ​​a nature reserve and close to a newly constructed road will be assessed as having an "extremely high" ecological risk level. Based on the overlay analysis results, different degraded clusters are labeled "high risk," "medium risk," and "low risk." Finally, the spatial extent of all degraded clusters assessed as "high risk" is merged and designated as priority wetland degradation hotspots.

[0132] Furthermore, in another aspect of this application, in some embodiments, this application provides a multi-source remote sensing image analysis system for dynamic monitoring of wetlands, referring to... Figure 4 The figure is a schematic diagram of the structure of a multi-source remote sensing image analysis system for dynamic monitoring of wetlands according to some embodiments of this application. The multi-source remote sensing image analysis system for dynamic monitoring of wetlands includes: an acquisition module 401, a processing module 402, and an execution module 403, which are described below:

[0133] Acquisition module 401, in this application, is mainly used to simultaneously acquire synthetic aperture radar images and multispectral images of the target wetland to obtain a multi-source time-series image set of the wetland in a spatiotemporal coordination.

[0134] Processing module 402, in this application, is used to determine the seasonal phenological characteristics of wetland vegetation and water based on the multi-source temporal image set, and to extract the spatial texture and geometric features of the wetland from the multispectral images;

[0135] It should be noted that the processing module 402 in this application is also used to perform dual-path change detection on the temporal phenological features and the spatial texture and geometric features to obtain the seasonal evolution pattern and the heterogeneous structure of the boundary of the wetland, and then determine the change intensity map of the wetland through the evolution pattern and the heterogeneous structure.

[0136] Additionally, it should be noted that the processing module 402 in this application is also used to obtain a wetland type conversion rule base, extract multiple change patches of the change intensity map according to the conversion rule base, and perform judgment filtering and reclassification on all change patches to obtain the wetland spatial distribution and multiple types of conversion events;

[0137] The execution module 403 in this application is mainly used to perform spatiotemporal integration and visualization of all transformation events and the spatial distribution of wetlands, and to generate hotspot areas of wetland degradation.

[0138] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described multi-source remote sensing image analysis method for dynamic monitoring of wetlands.

[0139] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device for implementing a multi-source remote sensing image analysis method for dynamic wetland monitoring, according to some embodiments of this application. The multi-source remote sensing image analysis method for dynamic wetland monitoring in the above embodiments can be implemented through... Figure 5 The computer device shown is used to implement this, and the computer device includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0140] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0141] The communication bus 502 can be used to transmit information between the aforementioned components.

[0142] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0143] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. The method used in the above embodiments can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0144] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0145] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0146] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0147] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described multi-source remote sensing image analysis method for dynamic monitoring of wetlands.

[0148] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0149] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for analyzing multi-source remote sensing images for dynamic monitoring of wetlands, characterized in that, Includes the following steps: Synthetic aperture radar and multispectral images of the target wetland were acquired simultaneously to obtain a multi-source time-series image set of the wetland in a spatiotemporal coordination. Based on the multi-source temporal image set, the seasonal phenological characteristics of wetland vegetation and water are determined, and the spatial texture and geometric features inside the wetland are extracted from the multispectral images. By performing dual-path change detection on the temporal phenological features and the spatial texture and geometric features, the seasonal evolution pattern and boundary heterogeneous structure of wetlands are obtained, and then the change intensity map of wetlands is determined through the evolution pattern and the heterogeneous structure. A wetland type conversion rule base is obtained, and multiple change patches of the change intensity map are extracted according to the conversion rule base. All change patches are judged, filtered and reclassified to obtain the spatial distribution of wetlands and multiple types of conversion events. All transformation events and the spatial distribution of wetlands are spatiotemporally integrated and visualized to generate hotspots of wetland degradation.

2. The method as described in claim 1, characterized in that, Synthetic aperture radar (SAR) images and multispectral images of the target wetland were acquired simultaneously to obtain a spatiotemporally coordinated multi-source time-series image set of the wetland, specifically including: Synthetic aperture radar and multispectral images covering the target wetland are acquired simultaneously within a preset monitoring period. The synthetic aperture radar image is subjected to noise suppression and moisture enhancement. Atmospheric correction and cloud masking were performed on the multispectral images; The processed synthetic aperture radar image and multispectral image were spatiotemporally registered to obtain a multi-source time-series image set of wetland spatiotemporally coordinated.

3. The method as described in claim 1, characterized in that, Determining the seasonal phenological characteristics of wetland vegetation and water based on the multi-source temporal image set specifically includes: The spectral index is calculated pixel-by-pixel from the multispectral image sequence to obtain the spectral index time series. The backscattering coefficients were extracted from the synthetic aperture radar images to obtain the time series of changes in surface moisture and roughness. The start and end dates, peak values, amplitudes, and seasonal integral features of the growing season within the hydrological year are extracted from the spectral index time series. The seasonal mean and variation range features are extracted from the time series to obtain the time series phenological features.

4. The method as described in claim 1, characterized in that, Extracting spatial texture and geometric features of the wetland interior from the multispectral imagery specifically includes: The red, green, blue, and near-infrared bands with the highest spatial resolution in the multispectral image were selected. Determine the spatial texture features within the wetland based on all wavebands; The multispectral image is divided into multiple homogeneous regions; The area, perimeter, aspect ratio, and shape index of each homogeneous region are determined as the geometric features of the wetland interior.

5. The method as described in claim 1, characterized in that, By performing dual-path change detection on the temporal phenological features and the spatial texture and geometric features, the evolution pattern of wetland seasons and the heterogeneous structure of its boundaries are obtained, specifically including: Determine the temporal feature differences of the temporal phenological features across all pixels between the two temporal phases; The pseudo-image sequence is determined based on all temporal feature differences; The evolution pattern of wetlands within seasonal intervals in the pseudo-image sequence is captured by the gated loop unit of the first detection path, thereby obtaining the seasonal evolution pattern of wetlands. Determine the differences in texture features of all pixels within the dual temporal phase and neighborhood window for the spatial texture and geometric features; The convolutional neural network of the second detection path extracts the structural changes of wetland boundary contour and internal texture in all texture feature differences, thus obtaining the heterogeneous structure of the wetland boundary.

6. The method as described in claim 1, characterized in that, Determining the intensity map of wetland change through the evolutionary pattern and the heterogeneous structure specifically includes: Based on the evolution pattern and the heterogeneous structure, multiple joint feature vectors of the multi-source remote sensing image are determined; All joint feature vectors are mapped to change confidence levels to obtain a wetland change intensity map.

7. The method as described in claim 1, characterized in that, Extracting multiple change patches from the change intensity map based on the transformation rule base specifically includes: Identify multiple potential change pixels in the change intensity map; Based on all the potential change cells, multiple initial change patches are formed that are connected; Based on the transformation rule base, the transformed features are extracted from all the initial transformed features.

8. The method as described in claim 1, characterized in that, All changed patches are filtered and reclassified to obtain the spatial distribution of wetlands and multiple types of transformation events, including: Identify multiple land cover types for all changed patches over biphasic time; The land use type of all changed land parcels is determined and filtered using the aforementioned conversion rule base, and the corresponding changed land parcels are marked. All marked change patches are classified to obtain the spatial distribution of wetlands and multiple types of transformation events.

9. The method as described in claim 1, characterized in that, The spatiotemporal integration and visualization of all transformation events and the aforementioned wetland spatial distribution generate wetland degradation hotspots, specifically including: The transformation events from multiple monitoring periods are overlaid with the spatial distribution to calculate the total frequency of degradation-type transformations within each spatial unit. Smoothing and clustering analysis are performed on the spatial units of total frequency to identify spatial clustering regions. By overlaying the identified clusters with the boundaries of nature reserves and the intensity of human activities, high-risk wetland degradation hotspots are delineated.

10. A multi-source remote sensing image analysis system for dynamic monitoring of wetlands, characterized in that, include: The acquisition module is used to simultaneously acquire synthetic aperture radar images and multispectral images of the target wetland to obtain a multi-source time-series image set of the wetland in a spatiotemporal coordination. The processing module is used to determine the seasonal phenological characteristics of wetland vegetation and water based on the multi-source temporal image set, and to extract the spatial texture and geometric features of the wetland from the multispectral images. The processing module is also used to perform dual-path change detection on the temporal phenological features and the spatial texture and geometric features to obtain the seasonal evolution pattern and the heterogeneous structure of the boundary of the wetland, and then determine the change intensity map of the wetland through the evolution pattern and the heterogeneous structure. The processing module is also used to obtain a wetland type conversion rule base, extract multiple change patches from the change intensity map according to the conversion rule base, and perform judgment filtering and reclassification on all change patches to obtain the wetland spatial distribution and multiple types of conversion events; The execution module is used to perform spatiotemporal integration and visualization of all transformation events and the spatial distribution of wetlands, and to generate hotspot areas of wetland degradation.

Images