A high-resolution optical satellite coastal zone change detection method driven by physical prior

Abstract

The application discloses a kind of high-resolution optical satellite coast zone change detection method driven by physical priori, and belongs to the technical field of remote sensing image processing and computer vision, it is related to the water and land change detection method of high-resolution optical satellite remote sensing image for coast zone.The purpose of the present application is to solve the problem of false change and unclear boundary in the existing high-resolution coast zone change detection task.The process is as follows: obtaining a trained deep learning model;obtaining high-resolution optical satellite remote sensing images of the same coast area before and after the time to be measured;obtaining the front time image of each window and the corresponding window of the back time image;inputting the front time image of each window and the corresponding window of the back time image into the trained deep learning model to output the semantic change detection result;processing the semantic change detection result to obtain a vector patch, and assigning the vector patch with the attributes of the front time, the back time, the image source, the confidence, the area, the perimeter and the change type.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and computer vision technology, and relates to a method for detecting coastal zone land and water changes based on high-resolution optical satellite remote sensing images. Background Technology

[0002] The coastal zone is a transitional area where land and sea meet, and where natural processes and human activities are highly coupled. It is also an area where marine activities are highly concentrated, such as port expansion, land reclamation, offshore photovoltaic construction, nearshore aquaculture, coastal city development, and shoreline ecological restoration. Continuous, detailed, and quantifiable monitoring of coastal zone changes is of great significance for coastal resource management, ecological environmental protection, territorial spatial supervision, and verification of marine activities. High-resolution optical remote sensing imagery, with its wide coverage, strong repeatability, and rich detail of ground features, can provide an effective data foundation for coastal zone change detection.

[0003] Existing remote sensing change detection methods are mostly geared towards urban buildings, roads, or general land cover scenarios, typically directly comparing the spectral or depth features of images from different time periods. However, coastal zones exhibit significant land-water interaction characteristics. Waves, tides, turbid water, solar flares, floating objects, ships, and short-term displacements of aquaculture facilities can all produce strong apparent differences. These environmental disturbances may not correspond to actual land-water transitions, but they are easily identified as changed areas by existing change detection models, thus creating false changes. Furthermore, the intertidal zone has a tortuous and fragmented boundary, and a dense distribution of aquaculture facilities and nearshore structures makes it prone to boundary adhesion, voids, and internal fragmentation.

[0004] High-resolution satellite imagery typically includes panchromatic and multispectral data. Different images may differ in imaging time, solar elevation angle, sensor response, geometric positioning accuracy, and spatial resolution. Directly comparing pixels from two temporal phases can lead to misclassification by the model as real changes due to orthorectification residuals, panchromatic / multispectral misalignment, and inconsistencies in the two-temporal pixel grid. Therefore, coastal zone change detection requires not only constructing a detection network adapted to the semantic features of land and water, but also establishing a standardized processing flow from raw high-resolution imagery to a unified pixel grid. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of false changes and unclear boundaries in existing high-resolution coastal zone change detection tasks, and to propose a physical prior-driven high-resolution optical satellite coastal zone change detection method.

[0006] The specific process of a high-resolution optical satellite coastal zone change detection method driven by physical priors is as follows:

[0007] Step 1: Construct the training set data;

[0008] Step 2: Build a deep learning model;

[0009] Step 3: Train the deep learning model based on the training set to obtain a trained deep learning model;

[0010] Step 4: Acquire high-resolution optical satellite remote sensing images of the same coastal area to be measured in the previous and subsequent phases; process the acquired high-resolution optical satellite remote sensing images of the same coastal area to be measured in the previous and subsequent phases to obtain the previous phase images of each window and the corresponding subsequent phase images of the window.

[0011] The preceding temporal image and the following temporal image of each window are input into the trained deep learning model, and the trained deep learning model outputs the semantic change detection results. ;

[0012] Step 5: Analyze the semantic change detection results output by the trained deep learning model. The vector patches are processed and then assigned attributes such as previous time phase, next time phase, image source, confidence level, area, perimeter, and change type.

[0013] The beneficial effects of this invention are as follows:

[0014] The purpose of this invention is to provide a physical prior-driven high-resolution optical satellite coastal zone change detection method to solve the problems of spurious changes and unclear boundaries in high-resolution coastal zone change detection tasks. By establishing a unified image standardization preprocessing workflow, the method reduces unrealistic changes caused by geometric misalignment and pixel grid inconsistency in dual-temporal images. At the same time, by utilizing physical priors and deep networks in synergy, the method suppresses environmental spurious changes such as tides, waves, and movable targets, thereby improving the accuracy of water-land change direction discrimination and fine change boundary extraction.

[0015] This invention provides a high-resolution optical satellite coastal zone change detection method driven by physical priors. Through a standardized preprocessing link, a dual-stream network architecture that is enhanced by synergistic enhancement of land-water physical prior constraints and time-frequency boundaries, it achieves fully automated detection of pixel-level coastal zone changes.

[0016] Compared with existing coastal zone remote sensing change detection technologies, this invention has significant technical advantages. Through the above-described complete technical solution, this invention can achieve high-precision, interference-resistant, and boundary-complete dynamic monitoring of the coastal zone, possessing strong engineering application value and industry promotion prospects. It has at least the following beneficial effects:

[0017] (1) By adopting a consistent digital elevation model, target projection and target pixel size for orthorectification of dual temporal images, and performing feature matching and fine registration in stable land areas after water interference, the geometric consistency and pixel matching accuracy of the preceding and following temporal images are significantly improved.

[0018] (2) By determining the spatial intersection of the effective data range of the dual-temporal orthophotos, and constructing the target cell grid based on the unified grid origin and target cell size, the preceding and following temporal images have the same spatial range, number of rows and columns, and affine geographic transformation parameters, thereby ensuring that the corresponding cells represent the same geographical location and reducing the pseudo-changes caused by inconsistent cell grids.

[0019] (3) By using panchromatic and multispectral subpixel registration and panchromatic multispectral fusion, high spatial resolution and multispectral information are taken into account, providing a data foundation for the identification of subtle shoreline changes, aquaculture facilities and nearshore structures.

[0020] (4) Using the normalized differential water index to generate a dual-temporal water mask, and constructing a binary change prior and a directional water-land semantic change prior, the coastal zone water-land change pattern is explicitly injected into the detection network, which helps to reduce water-land semantic confusion caused by turbid water, shadow and light differences.

[0021] (5) The frequency domain enhancement branch learns and modulates the spectral response of the difference features, which helps to highlight the spatially consistent land and water change features and reduce random texture interference such as waves, tides and reflections.

[0022] (6) The boundary-aware branch enhances the geometric representation of changing contours through a fixed-weight and differentiable Sobel operator and boundary-assisted supervision, which helps to alleviate boundary adhesion, holes and internal fragmentation of targets in densely aquaculture areas and slender shorelines.

[0023] (7) The binary-semantic dual-stream detection structure first determines whether the pixel has changed, and then uses the binary change probability to perform soft gating on the semantic discrimination of water body to land and land to water body, so that the semantic classification pays more attention to the high confidence change area and reduces the misclassification of unchanged background.

[0024] (8) The present invention forms a complete process from high-resolution image preprocessing, pixel-level alignment, deep learning model inference to raster and vector output, which can be used for monitoring scenarios such as reclamation, port expansion, nearshore aquaculture, offshore photovoltaic construction, shoreline management and coastal ecological restoration. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the overall process of the physical prior-driven high-resolution optical satellite coastal zone change detection method of the present invention.

[0026] Figure 2 This is a flowchart illustrating the high-resolution dual-temporal image radiometric calibration, atmospheric correction, orthorectification, panchromatic multispectral registration and fusion, dual-temporal fine registration, and unified grid construction of the intersection region in this invention.

[0027] Figure 3This is a schematic diagram of the overall structure of the high-resolution optical satellite coastal zone change detection model driven by physical priors in this invention.

[0028] Figure 4 This is a schematic diagram of the frequency domain enhancement branch of the present invention;

[0029] Figure 5 This is a schematic diagram of the boundary-aware branch structure of the present invention. Detailed Implementation

[0030] Specific Implementation Method 1: The specific process of this implementation method for detecting coastal zone changes using a high-resolution optical satellite driven by physical priors is as follows:

[0031] Step 1: Construct the training set data;

[0032] Step 2: Build a deep learning model;

[0033] Step 3: Train the deep learning model based on the training set to obtain a trained deep learning model;

[0034] Step 4: Acquire high-resolution optical satellite remote sensing images of the same coastal area to be measured at two different times (before and after); process the high-resolution optical satellite remote sensing images of the same coastal area to be measured at two different times (before and after) to obtain the before-time image of each window and the after-time image of the corresponding window.

[0035] The preceding temporal image and the following temporal image of each window are input into the trained deep learning model, and the trained deep learning model outputs the semantic change detection results. ;

[0036] The high resolution is defined as follows: when only multispectral data is available, its spatial resolution is better than (not less than) 5 meters; when both multispectral and panchromatic data are available, the resolution of the fused image is better than (not less than) 5 meters.

[0037] Step 5: Analyze the semantic change detection results output by the trained deep learning model. The vector patches are processed and then assigned attributes such as previous time phase, next time phase, image source, confidence level, area, perimeter, and change type.

[0038] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: Step 1, constructing training set data;

[0039] The specific process is as follows:

[0040] Step 11: Acquire high-resolution optical satellite remote sensing images of the same coastal area at two different times (before and after the initial phase);

[0041] Each time phase includes at least the blue, green, red, and near-infrared bands;

[0042] If the satellite has panchromatic imaging products, then panchromatic images of the corresponding time phase will be acquired simultaneously.

[0043] Synchronously acquire satellite sensor metadata and rational polynomial coefficient files;

[0044] Download the digital elevation model covering the imaging area;

[0045] Step 12: Perform radiometric calibration, atmospheric correction, orthorectification, and registration on the previous and subsequent temporal images in sequence to obtain the pixel-aligned previous temporal image. Post-temporal images aligned to pixel level ;

[0046] Use preset pixel size ( (pixel) Pixel-level aligned previous temporal images using a sliding window method Cut;

[0047] Use preset pixel size ( (pixel) Pixel-level aligned post-temporal images using a sliding window method Cut;

[0048] Record the images of each window before cropping. and The row and column offsets and affine geographic transformation parameters are used to stitch the prediction results together and restore them to the original geographic coordinates;

[0049] The affine geotransformation parameters include the horizontal coordinate of the top left corner of each window image, the horizontal pixel resolution of each window, the row rotation parameters of each window, the vertical coordinate of the top left corner of each window image, the column rotation parameters of each window, and the vertical pixel resolution of each window, which are used to establish the mapping relationship between image row and column indices and geographic coordinates.

[0050] Step 13: Based on the previous temporal images of each window and the corresponding post-temporal image in the same window Normalized differential water index was calculated using the green band and near-infrared band respectively. and ;

[0051]

[0052] in, Indicates phase The green band, Indicates phase The near-infrared band, To prevent extremely small positive numbers with a denominator of zero;

[0053] Step 14: Generate a water body mask for the preceding time phase using a preset fixed threshold. and post-phase water mask ;

[0054] Step 15: Based on the previous phase water body mask and post-phase water mask Generate prior knowledge of semantic changes between land and water ; indicates as:

[0055]

[0056] Step 16: Based on prior knowledge of semantic changes between land and water Generate binary transformation prior ; indicates as:

[0057]

[0058] in, To Take the absolute value;

[0059] According to the category definition of this invention, This indicates that the preceding time phase was a body of water and the following time phase was a body of land, i.e., a transition from water to land; This indicates that the preceding time phase was land and the following time phase was water, i.e., land turning into water. This indicates that the water and land attributes have not changed.

[0060] The other steps and parameters are the same as in Specific Implementation Method 1.

[0061] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: in step 12, radiometric calibration, atmospheric correction, orthorectification, and registration are performed sequentially on the preceding and following temporal images to obtain pixel-level aligned preceding temporal images. Post-temporal images aligned to pixel level ;

[0062] Use preset pixel size ( (pixel) Pixel-level aligned previous temporal images using a sliding window method Cut;

[0063] Use preset pixel size ( (pixel) Pixel-level aligned post-temporal images using a sliding window method Cut;

[0064] Record the row and column offsets of each window in the target cell grid and the affine geographic transformation parameters so that the prediction results can be stitched together and restored to the original geographic coordinates;

[0065] The affine geotransformation parameters include the horizontal coordinate of the top left corner of each window image, the horizontal pixel resolution of each window, the row rotation parameters of each window, the vertical coordinate of the top left corner of each window image, the column rotation parameters of each window, and the vertical pixel resolution of each window, which are used to establish the mapping relationship between image row and column indices and geographic coordinates.

[0066] The specific process is as follows:

[0067] Step 121: Perform radiometric calibration, atmospheric correction, and orthorectification based on rational polynomial coefficients and digital elevation model on the previous and subsequent time-phase images respectively.

[0068] The specific process is as follows:

[0069] Step 1211: For each band Original pixel value The apparent radiance was obtained by performing radiometric calibration. :

[0070]

[0071] in, and These are the calibration gain and bias recorded in the satellite metadata, respectively.

[0072] Step 1212: Subsequently, atmospheric correction is performed using the 6S radiative transfer model based on parameters such as solar zenith angle, Earth-Sun distance, and band solar irradiance to obtain surface reflectance. The same correction method is used for the preceding and following time phases to reduce spectral shift caused by differences in sensor response and atmospheric conditions.

[0073] Steps 1, 2, 13, as follows Figure 2 As shown, orthorectification is performed on the temporal images based on the rational polynomial coefficient model and the digital elevation model covering the study area provided with the satellite product. The digital elevation model preferably uses data with a spatial resolution of 30 meters or higher, and the target projection can use the UTM projection or other unified projection coordinate system corresponding to the study area.

[0074] The preceding and following time-phase images are orthorectified using the same digital elevation model, the same target projection, and the same target pixel size to avoid additional effects of different geometric processing parameters on the pixel correspondence between the two time phases.

[0075] After orthorectification, the projection coordinate system, pixel size, effective data range, and geometric positioning accuracy of the previous and subsequent temporal images are checked to provide a unified geometric basis for subsequent panchromatic multispectral registration and fusion and dual-temporal fine registration.

[0076] Step 122: If a panchromatic image exists, perform sub-pixel registration between the orthorectified previous-phase panchromatic image and the same-phase multispectral image. Then, fuse the registered previous-phase panchromatic image and the same-phase multispectral image using the Gram-Schmidt fusion algorithm to obtain the fused previous-phase image. Perform sub-pixel registration between the orthorectified subsequent-phase panchromatic image and the same-phase multispectral image. Then, fuse the registered subsequent-phase panchromatic image and the same-phase multispectral image using the Gram-Schmidt fusion algorithm to obtain the fused subsequent-phase image. Using the fused previous-phase image as a reference, perform sub-pixel registration on the fused subsequent-phase image to obtain the registered dual-phase image.

[0077] If a panchromatic image is not available, subpixel registration is performed on the orthorectified rear temporal image based on the orthorectified front temporal image to obtain a registered dual temporal image.

[0078] Subpixel registration refers to the fact that the error between two registrations is no more than 1 pixel, which can also be called "pixel-level registration".

[0079] For panchromatic and multispectral images of the same time phase, the multispectral image is registered to the panchromatic pixel grid, using the panchromatic image as a reference. Feature matching is performed between the intensity components of the multispectral image and the low-pass processed panchromatic image to estimate sub-pixel translation or local translation field, ensuring that the multispectral edges are consistent with the panchromatic edges. In a preferred embodiment, the residual misalignment after registration is less than 0.5 panchromatic pixels. A high spatial resolution fusion algorithm based on multispectral channel weighting is used to generate a four-band high-resolution fused image.

[0080] Step 123: Extract the spatial intersection of the effective data range of the registered dual-temporal images and construct a unified target pixel grid;

[0081] Step 124: Crop the registered dual-temporal images to the intersection range to obtain the front temporal intersection image and the back temporal intersection image;

[0082] Step 125: Resample the previous temporal intersection image to the target pixel grid (divide the previous temporal intersection image into a pixel grid according to the preset target pixel size and a unified grid origin) to obtain a pixel-aligned previous temporal image. (Pixel-aligned pre-temporal images) Includes a built-in projection coordinate system, spatial range, spatial resolution, and target pixel size (preset target pixel size). ), number of rows and columns, and no data value (pixel value is 0);

[0083] The later temporal intersection image is resampled to the target pixel grid (the later temporal intersection image is divided into pixel grids according to the preset target pixel size and a unified grid origin) to obtain a pixel-level aligned later temporal image. (Pixel-aligned post-temporal images) Includes a built-in projection coordinate system, spatial range, spatial resolution, and target pixel size (preset target pixel size). ), number of rows and columns, and no data value (pixel value is 0);

[0084] Pixel-aligned pre-temporal images Post-temporal images aligned to pixel level Having the same projected coordinate system, spatial extent, spatial resolution, number of rows and columns, and affine geographic transformation parameters, for the same row and column indices , and Representing the same geographical location;

[0085] After the unified mesh processing is completed, a consistency check is performed on the width, height, affine geographic transformation parameters, and invalid value mask of the dual-temporal image. If any parameter is inconsistent, the process returns to the target mesh construction step (steps 123, 124, and 125) for reprocessing to ensure that the model input strictly corresponds at the pixel level.

[0086] Step 126: Use the preset pixel size ( The pixels are cropped using a sliding window method to crop the pixel-level aligned previous temporal image, with the spacing between adjacent windows set. The overlap rate; the windows that do not change are screened and randomly sampled so that the number of unchanged windows retained is not less than 10% of the total number of windows, so as to ensure that the model has sufficient ability to represent the unchanged background, while avoiding class imbalance caused by too many background samples.

[0087] Use preset pixel size ( The pixel-level aligned post-temporal image is cropped using a sliding window method, with the spacing between adjacent windows set... The overlap rate; the windows that do not change are screened and randomly sampled so that the number of unchanged windows retained is not less than 10% of the total number of windows, so as to ensure that the model has sufficient ability to represent the unchanged background, while avoiding class imbalance caused by too many background samples.

[0088] Step 127: Record the row and column offsets and affine geographic transformation parameters of each window in the target cell grid so that the prediction results can be stitched together and restored to the original geographic coordinates;

[0089] The affine geotransformation parameters include the horizontal coordinate of the top left corner of each window image, the horizontal pixel resolution of each window, the row rotation parameters of each window, the vertical coordinate of the top left corner of each window image, the column rotation parameters of each window, and the vertical pixel resolution of each window, which are used to establish the mapping relationship between image row and column indices and geographic coordinates.

[0090] In one specific embodiment, the projection coordinate system of a large-scale remote sensing image to be processed is EPSG:32651, i.e., WGS84 / UTM zone 51N; its spatial range is: 303935.2815, 3586832.3950: 334402.4815, 3619431.5950. The spatial resolution of the large-scale remote sensing image is 0.8m, and the image size is 40749 rows and 38084 columns. A sliding window with a size of 512 rows × 512 columns is used to crop the large-scale remote sensing image. Taking the first window located in the upper left corner of the large-scale image as an example, the starting row offset and starting column offset of this window in the unified target cell grid are both 0, its projection coordinate system is still EPSG:32651, and its spatial resolution is still 0.8 m. The ground width and height corresponding to this window are both...

[0091] ;

[0092] Therefore, the spatial range of this window is: 303935.2815, 3619431.5950-409.6:303935.2815+409.6, 3619431.5950, or: 303935.2815, 3619021.9950:304344.8815, 3619431.5950. The first two values ​​represent the minimum x-coordinate and minimum y-coordinate of the window, respectively, and the last two values ​​represent the maximum x-coordinate and maximum y-coordinate of the window, respectively. After the window completes its prediction, the prediction results are backfilled into rows 0 to 511 and columns 0 to 511 of the entire predicted image, based on its starting row and column offsets. For other sliding windows, the spatial range of the corresponding window is calculated based on its starting row offset, starting column offset and spatial resolution of 0.8 m in the unified target cell grid, and the prediction results are backfilled to the corresponding positions in the unified target cell grid.

[0093] Other steps and parameters are the same as in specific implementation method one or two.

[0094] Specific Implementation Method Four: This implementation method differs from one of the specific implementation methods one to three in that: in step 123, the spatial intersection of the effective data range of the registered dual-temporal images is extracted to construct a unified target pixel grid.

[0095] The specific process is as follows:

[0096] Step 1231: Determine the effective data range of the previous temporal phase based on the non-nullable pixels of the registered previous temporal image. ;

[0097] The effective data range of the post-temporal image is determined based on the non-nullable pixels of the registered post-temporal image. ;

[0098] Step 1232: Expand the effective data range of the previous time phase. and the effective data range of the later phase Transform to the same target projection coordinate system and calculate in the same target projection coordinate system. and Spatial intersection ,Right now:

[0099]

[0100] The spatial intersection As a processing range that is effective for both preceding and subsequent temporal images;

[0101] Steps 1, 2, 3, and 3: Based on spatial intersection Preset target pixel size and unified grid origin (Intersection top left corner point) Construct the target cell mesh.

[0102] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0103] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that: in step 14, a pre-set fixed threshold is used to generate a pre-temporal water body mask. and post-phase water mask ; indicates as:

[0104]

[0105]

[0106] Among them, threshold A predetermined fixed threshold can be determined based on the sensor type, image reflectance range, and validation samples;

[0107] For indicator functions, The indicator function is set to 1. ; Indicator function takes 0, ;

[0108] For indicator functions, The indicator function is set to 1. ; Indicator function takes 0, ;

[0109] In the same batch of image processing, the corresponding temporal thresholds are kept constant; the water body masks of the previous temporal phases are respectively... and post-phase water mask Small-scale morphological processing was performed to remove isolated noise and preserve the main land-water boundaries;

[0110] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0111] Specific Implementation Method Six: This implementation method differs from one of Specific Implementation Methods One to Five in that: Step 2, constructing a deep learning model;

[0112] The specific process is as follows:

[0113] The deep learning model includes ResNet-18, Transformer encoder, Transformer decoder, frequency domain enhancement branch, boundary-aware branch, binary change detection head, and semantic change detection head.

[0114] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0115] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One to Six in that: Step 3, training the deep learning model based on the training set to obtain a trained deep learning model;

[0116] The specific process is as follows:

[0117] Step 31: Extract multi-scale high-dimensional features from the front-phase image of a certain window and the back-phase image of the corresponding window obtained in Step 12 using ResNet-18 with shared parameters, to obtain the high-dimensional features of the front-phase image of a certain window and the high-dimensional features of the back-phase image of the corresponding window.

[0118] Step 32: Generate the corresponding image of the previous temporal image from the high-dimensional features of a certain window using a semantic token (Semantic Tokenizer). Each of the following semantic tokens is a semantic token. dimensional vector;

[0119] Each semantic token has a shape of [4, 32], where 4 represents the number of semantic tokens (defined by the user) and 32 represents the feature dimension of each token (also defined by the user).

[0120] The high-dimensional features of the subsequent phase image of the corresponding window are used to generate the corresponding subsequent phase image using a semantic token (Semantic Tokenizer). Each of the following semantic tokens is a semantic token. dimensional vector;

[0121] Step 33: Match the previous time-phase images with... Each semantic tag and the corresponding post-temporal image The semantic tags are concatenated along the token quantity dimension (two [4, 32] are concatenated into [8, 32],) to obtain the concatenated semantic tags;

[0122] Step 34: Input the concatenated semantic tags into the Transformer encoder, and the Transformer encoder outputs the encoded token;

[0123] The encoded tokens are segmented according to the number of tokens. Each segmented token is then input into two Transformer decoders with shared weights to obtain pre-temporal features enhanced with global context. and post-phase characteristics ;

[0124] Step 35: Based on previous temporal features and post-phase characteristics Calculate the difference features ; indicates as:

[0125]

[0126] In other implementations, the convolutional neural network backbone with shared weights can also adopt other convolutional networks with multi-scale feature outputs, as long as they can cooperate with the spatiotemporal Transformer feature interaction network to obtain pre-temporal features, post-temporal features and difference features.

[0127] Step 36: Difference features Input the frequency domain enhancement branch and the boundary sensing branch respectively to obtain the frequency domain enhancement features. and boundary features ;

[0128] Step 37: Analyze the differential features Frequency domain enhancement features and boundary features The features are stitched together along the channel dimension and then fused through a convolutional layer to obtain spatiotemporal-frequency-boundary fused features. ;

[0129] Step 38: Define boundary features Through a Convolutional layers produce boundary prediction maps ;

[0130] Step 39: During model training, label the semantic changes between land and water. Perform the Sobel operation to obtain the Sobel result; land and water semantic change tags. The categories are 0, 1, and 2; 0, 1, and 2 represent no change, water body to land, and land body to water, respectively.

[0131] Binarize the Sobel operation result (specify a threshold, then change pixels smaller than the threshold to 0 and pixels greater than or equal to the threshold to 1) to obtain the boundary supervision label. Boundary monitoring labels The categories are 0 and 1; category 0 represents non-boundary, and category 1 represents boundary.

[0132] Step 310: Based on the boundary prediction map and boundary supervision labels Calculate boundary auxiliary loss ;

[0133] The boundary auxiliary loss To measure the difference between the prediction boundary and the supervision boundary, binary cross-entropy loss, Dice loss, or a combination of both can be used.

[0134] With boundary-assisted supervision, the model can focus on learning the gradient structure of the edge of reclamation, dikes, aquaculture facility boundaries and slender coastlines, reducing the adhesion and internal fragmentation of changing boundaries;

[0135] Step 311: Apply the binary transformation prior generated in step 16 Input label embedding layer to obtain binary transformation prior embedding. ;

[0136] The semantic change prior of land and water generated in step 15 Input the label embedding layer to obtain the prior embedding of water and land semantic changes. ;

[0137] Step 312: Fusion of spatiotemporal-frequency-boundary features With binary transformation prior embedding After concatenation, the data is input into a binary change detection head to obtain a binary change probability map. ; indicates as:

[0138]

[0139] in, This indicates a binary change detection head, which consists of a standard fully convolutional network (FCN) layer with 2 output channels;

[0140] express Activation function layer;

[0141] Step 313: Fusion of spatiotemporal-frequency-boundary features Prior embedding of semantic changes between water and land Concatenate the data and utilize the binary change probability map. Constructing binary change detection results ; indicates as:

[0142]

[0143] in, This indicates element-wise multiplication;

[0144] Step 314: Convert the binary change detection results Input the semantic change detection head and perform the operation along the category channel. The calculation yields pixel-level probability maps for each semantic change category. ; indicates as:

[0145]

[0146] in, The semantic change detection head consists of a standard fully convolutional network (FCN) layer with 3 output channels;

[0147] Step 315: Using the argmax function, select the category with the highest probability for each pixel to obtain the semantic change detection result. ; indicates as:

[0148]

[0149] in, Indicates semantic change category Pixel-level probability map; ;

[0150] Category 0, Category 1, and Category 2 represent no change, water body to land, and land body to water, respectively;

[0151] The soft gating does not use the binary change results to hard truncate the semantic change branch. Instead, it uses the binary change probability to spatially attention-weight the semantic change features, enhances the semantic response in regions with high change probability, and retains the basic response in regions with low change probability, thereby avoiding the complete blocking of the semantic change branch by missing the binary change branch.

[0152] Step 316: Based on the binary change probability diagram and binary change labels Calculate binary change loss ;

[0153] The binary change loss is adopted as binary cross-entropy loss, Dice loss, or a combination of the two.

[0154] Based on the semantic change probability diagram and semantic change tags Calculate semantic change loss ;

[0155] The semantic change loss employs multi-class cross-entropy loss, Dice loss, or a combination of both.

[0156] Step 317: The total loss during model training is expressed as:

[0157]

[0158] in, and The weights are non-negative; in one example, for , for ;

[0159] During training, random flipping, rotation, scaling transformation, and slight spectral perturbation are used, but strong geometric deformations that change the correspondence between bi-phase pixels are not used.

[0160] Step 318: Train the deep learning model based on all window data and the total loss from step 12 to obtain the trained deep learning model.

[0161] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0162] Specific Implementation Method Eight: This implementation method differs from one of Specific Implementation Methods One to Seven in that: Step 36: [The text abruptly ends here, likely due to an incomplete sentence or a missing section.] Input the frequency domain enhancement branch and the boundary sensing branch respectively to obtain the frequency domain enhancement features. and boundary features ;

[0163] The specific process is as follows:

[0164] Step 361: Difference features Input frequency domain enhancement branch, output frequency domain enhancement feature ;like Figure 4 As shown; the specific process is as follows:

[0165] Step 3611: To reduce the pseudo-changes caused by wave patterns, tide marks, and instantaneous ship edges, the differential features are... Perform a two-dimensional fast Fourier transform; represented as:

[0166]

[0167] in, Represents a two-dimensional Fast Fourier Transform; For complex spectrum;

[0168] Step 3612: Utilize shared convolutional filters respectively right The real and imaginary parts of the frequency response are modulated, and the modulated real and imaginary parts are then recombined to obtain the frequency response modulated feature. ; indicates as:

[0169]

[0170] in, express The real part, express The imaginary part, This indicates a shared convolutional filter. A shared convolutional filter refers to... convolution; Represents the imaginary unit. ;

[0171] Step 3613: Modulate the frequency response characteristics Performing an inverse fast Fourier transform yields the frequency domain enhancement features. ; indicates as:

[0172]

[0173] in, This represents the inverse fast Fourier transform;

[0174] Shared convolutional filters learn the spectral response of differential features through end-to-end training, which helps to enhance spatially consistent variation features such as large-scale land reclamation, shoreline advancement, or water expansion, and reduce random texture interference such as waves, reflections, and small floating objects.

[0175] Step 362: Difference features Input the boundary-aware branch to obtain boundary features. ;like Figure 5 As shown;

[0176] Boundary features Through a Convolutional layers produce boundary prediction maps ;

[0177] The specific process is as follows:

[0178] Step 3621: Difference features enter convolution, Convolution converts differential features Compression into single-channel features ;

[0179] Step 3622: Use a fixed-weight, differentiable Sobel kernel and Calculate the horizontal and vertical gradients separately; calculate the gradient magnitude based on the horizontal and vertical gradients; expressed as:

[0180]

[0181] in, Represents the gradient in the horizontal direction. Represents the gradient in the vertical direction. This represents the convolution operation. Indicates the gradient magnitude. Represents positive numbers. ;

[0182] Step 3623: Set the gradient magnitude Input boundary coding units for feature extraction to obtain boundary features. ;

[0183] Boundary coding units include, in sequence: Convolutional layers Convolutional layer;

[0184] Boundary features Through a Convolutional layers produce boundary prediction maps .

[0185] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0186] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One to Eight in that: in step 4, high-resolution optical satellite remote sensing images of the same coastal area to be measured at the preceding and following time phases (two different times before and after) are acquired; the high-resolution optical satellite remote sensing images of the same coastal area to be measured at the preceding and following time phases (two different times before and after) are processed to obtain the preceding time phase image of each window and the following time phase image of the corresponding window;

[0187] The preceding temporal image and the following temporal image of each window are input into the trained deep learning model, and the trained deep learning model outputs the semantic change detection results. ;

[0188] The specific process is as follows:

[0189] Step 41: Acquire high-resolution optical satellite remote sensing images of the same coastal area to be measured at two different times, before and after the previous time phase.

[0190] Each time phase includes at least the blue, green, red, and near-infrared bands;

[0191] If the satellite has panchromatic imaging products, then panchromatic images of the corresponding time phase will be acquired simultaneously.

[0192] Synchronously acquire satellite sensor metadata and rational polynomial coefficient files;

[0193] Download the digital elevation model covering the imaging area;

[0194] Step 42: Perform radiometric calibration, atmospheric correction, orthorectification, and registration on the previous and subsequent temporal images sequentially to obtain the pixel-aligned previous temporal image. Post-temporal images aligned to pixel level ;

[0195] Using a preset pixel size in a sliding window manner (using Pixel window and (Sliding the overlap ratio) for pixel-level aligned previous temporal images Cut;

[0196] Using a preset pixel size in a sliding window manner (using Pixel window and (Sliding the overlap ratio) for pixel-level aligned post-temporal images Cut;

[0197] Record the row and column offsets of each window in the target cell grid and the affine geographic transformation parameters so that the prediction results can be stitched together and restored to the original geographic coordinates;

[0198] The affine geotransformation parameters include the horizontal coordinate of the top left corner of each window image, the horizontal pixel resolution of each window, the row rotation parameters of each window, the vertical coordinate of the top left corner of each window image, the column rotation parameters of each window, and the vertical pixel resolution of each window, which are used to establish the mapping relationship between image row and column indices and geographic coordinates.

[0199] In step 12, windows that do not change are filtered and randomly sampled to ensure that the number of unchanged windows retained is not less than 10% of the total number of windows.

[0200] Change to:

[0201] Keep all windows;

[0202] Step 43: Input the previous temporal image of each window and the subsequent temporal image of the corresponding window into the trained deep learning model. The trained deep learning model outputs the semantic change detection results. ;

[0203] For the class probabilities of the same pixel obtained from multiple windows, a center-weighted average method is used for fusion (for the class probabilities of the same pixel obtained from 3 windows, a center-weighted average method is used for fusion). The weight of the central region of the window is greater than the weight of the edge region, so as to reduce the seams caused by convolution boundary padding and block splicing.

[0204] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0205] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that: step 5, the semantic change detection result output by the trained deep learning model... The vector patches are processed and then assigned attributes such as previous time phase, next time phase, image source, confidence level, area, perimeter, and change type.

[0206] The specific process is as follows:

[0207] Semantic change detection results based on the output of the trained deep learning model Connectivity analysis was performed on water-to-land and land-to-water categories respectively, and patches with an area smaller than the preset minimum mapping unit were deleted.

[0208] Morphological closure operations were performed on the preserved plaques to connect the narrow fractures, resulting in connected plaques.

[0209] Semantic change detection results Connected component denoising and morphological correction are performed sequentially to obtain the morphologically corrected image;

[0210] The morphologically corrected image is then rasterized to vector (by directly calling the GDAL library to convert the raster into a vector), resulting in a rasterized polygon vector file.

[0211] Assign attributes such as previous phase, next phase, image source, confidence level, area, perimeter, and variation type to each patch in the rasterized polygon vector file.

[0212] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0213] The beneficial effects of the present invention are verified using the following embodiments:

[0214] Example 1:

[0215] In one specific embodiment, coastal zone change detection data is constructed using Gaofen-1 dual-temporal imagery acquired from 2013 to 2025. A 2-meter panchromatic image and an 8-meter blue, green, red, and near-infrared multispectral image are fused to generate a 2-meter four-band fused image. The sample covers 11 coastal provinces in my country, including a total of 48 sets of dual-temporal imagery. After radiometric calibration, atmospheric correction, orthorectification, panchromatic multispectral registration and fusion, dual-temporal fine registration, and unified pixel grid processing, the normalized difference water index is used to assist in the manual annotation of water-to-land and land-to-water changes. Two rounds of verification are conducted to eliminate spurious changes caused by tidal fluctuations.

[0216] The 48 sets of dual-temporal images were divided into 32 sets of training images and 16 sets of validation images. Pixel window and The overlap ratio was used to prune the samples, resulting in 4810 training samples and 1955 validation samples; the training batch size was set to 8, and the number of iterations was set to 40000.

[0217] The method of this invention obtains [results] in the category of water-to-land conversion. IoU and F1, in the land-to-water category IoU and The F1 score was higher than that of the BIT baseline. Compared with the BIT baseline, the IoU for water-to-land conversion improved by 6.49%, and the IoU for land-to-water conversion improved by 3.55%, indicating that the synergistic effect of water-land physical priors, frequency domain enhancement, and boundary awareness can improve the accuracy of water-land conversion detection in complex coastal scenarios.

[0218] In the ablation experiment, when only the physical prior injection module was used, the IoU for water-to-land and land-to-water categories were respectively... and After adding the frequency domain enhancement branch, both were improved to [a higher level]. and After further adding the boundary-aware branch, both reached [the following values]. and The above results indicate that each technical module can produce positive technical effects.

[0219] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A high-resolution optical satellite method for detecting coastal zone changes driven by physical priors, characterized in that: The specific process of the method is as follows: Step 1: Construct the training set data; Step 2: Build a deep learning model; Step 3: Train the deep learning model based on the training set to obtain a trained deep learning model; Step 4: Acquire high-resolution optical satellite remote sensing images of the same coastal area to be measured in the previous and subsequent phases; process the acquired high-resolution optical satellite remote sensing images of the same coastal area to be measured in the previous and subsequent phases to obtain the previous phase images of each window and the corresponding subsequent phase images of the window. The preceding temporal image and the following temporal image of each window are input into the trained deep learning model, and the trained deep learning model outputs the semantic change detection results. ; Step 5: Analyze the semantic change detection results output by the trained deep learning model. The vector patches are processed and then assigned attributes such as previous time phase, next time phase, image source, confidence level, area, perimeter, and change type.

2. The high-resolution optical satellite coastal zone change detection method driven by physical priors according to claim 1, characterized in that: Step 1: Constructing the training set data; The specific process is as follows: Step 11: Acquire high-resolution optical satellite remote sensing images of the same coastal area in the preceding and subsequent time phases; Each time phase includes at least the blue, green, red, and near-infrared bands; If the satellite has panchromatic imaging products, then panchromatic images of the corresponding time phase will be acquired simultaneously. Synchronously acquire satellite sensor metadata and rational polynomial coefficient files; Download the digital elevation model covering the imaging area; Step 12: Perform radiometric calibration, atmospheric correction, orthorectification, and registration on the previous and subsequent temporal images in sequence to obtain the pixel-aligned previous temporal image. Post-temporal images aligned to pixel level ; The previous temporal image was aligned pixel-level using a sliding window method with a preset pixel size. Cut; Post-temporal images with pixel-level alignment using a sliding window method with preset pixel size. Cut; Record the images of each window before cropping. and The row and column offsets and affine geographic transformation parameters in the data; The affine geotransformation parameters include the horizontal coordinate of the top left corner of each window image, the horizontal pixel resolution of each window, the row rotation parameters of each window, the vertical coordinate of the top left corner of each window image, the column rotation parameters of each window, and the vertical pixel resolution of each window. Step 13: Based on the previous temporal images of each window and the corresponding post-temporal image in the same window Normalized differential water index was calculated using the green band and near-infrared band respectively. and ; Step 14: Generate a water body mask for the preceding time phase using a preset fixed threshold. and post-phase water mask ; Step 15: Based on the previous phase water body mask and post-phase water mask Generate prior knowledge of semantic changes between land and water ; indicates as: Step 16: Based on prior knowledge of semantic changes between land and water Generate binary transformation prior ; indicates as: in, To Take the absolute value; This indicates that the preceding time phase was a body of water and the following time phase was a body of land, i.e., a transition from water to land; This indicates that the preceding time phase was land and the following time phase was water, i.e., land turning into water. This indicates that the water and land attributes have not changed.

3. The high-resolution optical satellite coastal zone change detection method driven by physical priors according to claim 2, characterized in that: Step 12 involves sequentially performing radiometric calibration, atmospheric correction, orthorectification, and registration on the preceding and following temporal images to obtain pixel-level aligned preceding temporal images. Post-temporal images aligned to pixel level ; The previous temporal image was aligned at the pixel level using a sliding window method with a preset pixel size. Cut; Post-temporal images with pixel-level alignment using a sliding window method with preset pixel size. Cut; Record the row and column offsets of each window in the target cell grid and the affine geographic transformation parameters; The affine geotransformation parameters include the horizontal coordinate of the top left corner of each window image, the horizontal pixel resolution of each window, the row rotation parameters of each window, the vertical coordinate of the top left corner of each window image, the column rotation parameters of each window, and the vertical pixel resolution of each window. The specific process is as follows: Step 121: Perform radiometric calibration, atmospheric correction, and orthorectification based on rational polynomial coefficients and digital elevation model on the previous and subsequent time-phase images respectively. Step 122: If a panchromatic image exists, perform subpixel registration between the orthorectified previous-phase panchromatic image and the same-phase multispectral image. Then, fuse the registered previous-phase panchromatic image with the same-phase multispectral image to obtain a fused previous-phase image. Perform subpixel registration between the orthorectified subsequent-phase panchromatic image and the same-phase multispectral image. Then, fuse the registered subsequent-phase panchromatic image with the same-phase multispectral image to obtain a fused subsequent-phase image. Using the fused previous-phase image as a reference, perform subpixel registration on the fused subsequent-phase image to obtain a registered dual-phase image. If a panchromatic image is not available, subpixel registration is performed on the orthorectified rear temporal image based on the orthorectified front temporal image to obtain a registered dual temporal image. Step 123: Extract the spatial intersection of the effective data range of the registered dual-temporal images and construct a unified target pixel grid; Step 124: Crop the registered dual-temporal images to the intersection range to obtain the front temporal intersection image and the back temporal intersection image; Step 125: Resample the previous temporal intersection image to the target pixel grid to obtain a pixel-aligned previous temporal image. ; The later temporal intersection image is resampled to the target pixel grid to obtain a pixel-aligned later temporal image. ; Step 126: Crop the pixel-aligned previous temporal image using a sliding window method with a preset pixel size, and set the spacing between adjacent windows. The overlap rate; filtering and random sampling of windows that do not change, so that the number of unchanged windows retained is not less than 10% of the total number of windows; The pixel-level aligned post-temporal image is cropped using a sliding window method with a preset pixel size, and the spacing between adjacent windows is set. The overlap rate; filtering and random sampling of windows that do not change, so that the number of unchanged windows retained is not less than 10% of the total number of windows; Step 127: Record the row and column offsets and affine geographic transformation parameters of each window in the target cell grid; The affine geotransformation parameters include the horizontal coordinate of the top left corner of each window image, the horizontal pixel resolution of each window, the row rotation parameters of each window, the vertical coordinate of the top left corner of each window image, the column rotation parameters of each window, and the vertical pixel resolution of each window.

4. The high-resolution optical satellite coastal zone change detection method driven by physical priors according to claim 3, characterized in that: Step 123 extracts the spatial intersection of the effective data range of the registered dual-temporal images and constructs a unified target pixel grid. The specific process is as follows: Step 1231: Determine the effective data range of the previous temporal phase based on the non-nullable pixels of the registered previous temporal image. ; The effective data range of the post-temporal image is determined based on the non-nullable pixels of the registered post-temporal image. ; Step 1232: Expand the effective data range of the previous time phase. and the effective data range of the later phase Transform to the same target projection coordinate system and calculate in the same target projection coordinate system. and Spatial intersection ,Right now: Steps 1, 2, 3, and 3: Based on spatial intersection Preset target pixel size and unified grid origin Construct the target cell mesh.

5. The high-resolution optical satellite coastal zone change detection method driven by physical priors according to claim 4, characterized in that: Step 14 involves generating a pre-temporal water mask using a preset fixed threshold. and post-phase water mask ; indicates as: Among them, threshold A predetermined fixed threshold; For indicator functions, The indicator function is set to 1. ; Indicator function takes 0, ; For indicator functions, The indicator function is set to 1. ; Indicator function takes 0, .

6. The high-resolution optical satellite coastal zone change detection method driven by physical priors according to claim 5, characterized in that: Step 2: Construct a deep learning model; The specific process is as follows: The deep learning model includes ResNet-18, Transformer encoder, Transformer decoder, frequency domain enhancement branch, boundary-aware branch, binary change detection head, and semantic change detection head.

7. The high-resolution optical satellite coastal zone change detection method driven by physical priors according to claim 6, characterized in that: Step 3 involves training the deep learning model based on the training set to obtain a trained deep learning model. The specific process is as follows: Step 31: Extract features from the front temporal image of a certain window and the back temporal image of the corresponding window obtained in Step 12 using ResNet-18 with shared parameters to obtain the high-dimensional features of the front temporal image of a certain window and the high-dimensional features of the back temporal image of the corresponding window. Step 32: Generate the corresponding image of the previous time phase from the high-dimensional features of the previous time phase image of a certain window using semantic tokens. Each token is a semantic tag. dimensional vector; The high-dimensional features of the corresponding window's post-temporal image are used to generate the corresponding post-temporal image using semantic tokens. Each token is a semantic tag. dimensional vector; Step 33: Match the previous time-phase images with... Each semantic tag and the corresponding post-temporal image The semantic tags are concatenated along the dimension of the number of tokens to obtain the concatenated semantic tags; Step 34: Input the concatenated semantic tags into the Transformer encoder, and the Transformer encoder outputs the encoded token; The encoded tokens are segmented according to the number of tokens. Each segmented token is then input into two Transformer decoders with shared weights to obtain pre-temporal features enhanced with global context. and post-phase characteristics ; Step 35: Based on previous temporal features and post-phase characteristics Calculate the difference features ; indicates as: Step 36: Difference features Input the frequency domain enhancement branch and the boundary sensing branch respectively to obtain the frequency domain enhancement features. and boundary features ; Step 37: Analyze the differential features Frequency domain enhancement features and boundary features The features are stitched together along the channel dimension and then fused through a convolutional layer to obtain spatiotemporal-frequency-boundary fused features. ; Step 38: Define boundary features Through a Convolutional layers produce boundary prediction maps ; Step 39: Tag semantic changes between water and land Perform the Sobel operation to obtain the Sobel result; land and water semantic change tags. Categories 0, 1, and 2; Category 0, Category 1, and Category 2 represent no change, water body to land, and land body to water, respectively; Binarize the Sobel operation results to obtain the boundary supervision labels. Boundary monitoring labels The categories are 0 and 1; category 0 represents non-boundary, and category 1 represents boundary. Step 310: Based on the boundary prediction map and boundary supervision labels Calculate boundary auxiliary loss ; The boundary auxiliary loss Use binary cross-entropy loss; Step 311: Apply the binary transformation prior generated in step 16 Input label embedding layer to obtain binary transformation prior embedding. ; The semantic change prior of land and water generated in step 15 Input the label embedding layer to obtain the prior embedding of water and land semantic changes. ; Step 312: Fusion of spatiotemporal-frequency-boundary features With binary transformation prior embedding After concatenation, the data is input into a binary change detection head to obtain a binary change probability map. ; indicates as: in, This indicates a binary change detection head, which consists of a standard fully convolutional network (FCN) layer with 2 output channels; express Activation function layer; Step 313: Fusion of spatiotemporal-frequency-boundary features Prior embedding of semantic changes between water and land Concatenate the data and utilize the binary change probability map. Constructing binary change detection results ; indicates as: in, This indicates element-wise multiplication; Step 314: Convert the binary change detection results Input the semantic change detection head and perform the operation along the category channel. The calculation yields pixel-level probability maps for each semantic change category. ; indicates as: in, The semantic change detection head consists of a standard fully convolutional network layer with 3 output channels. Step 315: Using the argmax function, select the category with the highest probability for each pixel to obtain the semantic change detection result. ; indicates as: in, Indicates semantic change category Pixel-level probability map; ; Category 0, Category 1, and Category 2 represent no change, water body to land, and land body to water, respectively; Step 316: Based on the binary change probability diagram and binary change labels Calculate binary change loss ; The binary change loss is a binary cross-entropy loss. Based on the semantic change probability diagram and semantic change tags Calculate semantic change loss ; The semantic change loss employs multi-class cross-entropy loss. Step 317: The total loss during model training is expressed as: in, and Non-negative weights; Step 318: Train the deep learning model based on all window data and the total loss from step 12 to obtain the trained deep learning model.

8. The high-resolution optical satellite coastal zone change detection method driven by physical priors according to claim 7, characterized in that: Step 36: Difference features Input the frequency domain enhancement branch and the boundary sensing branch respectively to obtain the frequency domain enhancement features. and boundary features ; The specific process is as follows: Step 361: Difference features Input frequency domain enhancement branch, output frequency domain enhancement feature The specific process is as follows: Step 3611: Differential Features Perform a two-dimensional fast Fourier transform; represented as: in, Represents a two-dimensional Fast Fourier Transform; For complex spectrum; Step 3612: Utilize shared convolutional filters respectively right The real and imaginary parts of the frequency response are modulated, and the modulated real and imaginary parts are then recombined to obtain the frequency response modulated feature. ; indicates as: in, express The real part, express The imaginary part, This indicates a shared convolutional filter. A shared convolutional filter refers to... convolution; Represents the imaginary unit. ; Step 3613: Modulate the frequency response characteristics Performing an inverse fast Fourier transform yields the frequency domain enhancement features. ; indicates as: in, This represents the inverse fast Fourier transform; Step 362: Difference features Input the boundary-aware branch to obtain boundary features. ; Boundary features Through a Convolutional layers produce boundary prediction maps ; The specific process is as follows: Step 3621: Difference features enter convolution, Convolution converts differential features Compression into single-channel features ; Step 3622: Use a fixed-weight, differentiable Sobel kernel and Calculate the horizontal and vertical gradients separately; calculate the gradient magnitude based on the horizontal and vertical gradients; expressed as: in, Represents the gradient in the horizontal direction. Represents the gradient in the vertical direction. This represents the convolution operation. Indicates the gradient magnitude. Represents positive numbers. ; Step 3623: Set the gradient magnitude Input boundary coding units for feature extraction to obtain boundary features. ; Boundary coding units include, in sequence: Convolutional layers Convolutional layer; Boundary features Through a Convolutional layers produce boundary prediction maps .

9. The high-resolution optical satellite coastal zone change detection method driven by physical priors according to claim 8, characterized in that: Step 4: Acquire high-resolution optical satellite remote sensing images of the same coastal area to be measured in the previous and subsequent phases; process the acquired high-resolution optical satellite remote sensing images of the same coastal area to be measured in the previous and subsequent phases to obtain the previous phase images of each window and the subsequent phase images of the corresponding window. The preceding temporal image and the following temporal image of each window are input into the trained deep learning model, and the trained deep learning model outputs the semantic change detection results. ; The specific process is as follows: Step 41: Acquire high-resolution optical satellite remote sensing images of the same coastal area before and after the current time. Each time phase includes at least the blue, green, red, and near-infrared bands; If the satellite has panchromatic imaging products, then panchromatic images of the corresponding time phase will be acquired simultaneously. Synchronously acquire satellite sensor metadata and rational polynomial coefficient files; Download the digital elevation model covering the imaging area; Step 42: Perform radiometric calibration, atmospheric correction, orthorectification, and registration on the previous and subsequent temporal images sequentially to obtain the pixel-aligned previous temporal image. Post-temporal images aligned to pixel level ; The previous temporal image was aligned at the pixel level using a sliding window method with a preset pixel size. Cut; The pixel-level aligned post-temporal image is processed using a sliding window method with a preset pixel size. Cut; Record the row and column offsets of each window in the target cell grid and the affine geographic transformation parameters; The affine geotransformation parameters include the horizontal coordinate of the top left corner of each window image, the horizontal pixel resolution of each window, the row rotation parameters of each window, the vertical coordinate of the top left corner of each window image, the column rotation parameters of each window, and the vertical pixel resolution of each window. Windowings that do not change are filtered and randomly sampled to ensure that the number of unchanged windows retained is not less than 10% of the total number of windows; Change to: Keep all windows open; Step 43: Input the previous temporal image of each window and the subsequent temporal image of the corresponding window into the trained deep learning model. The trained deep learning model outputs the semantic change detection results. ; For the class probabilities of the same pixel obtained from multiple windows, a center-weighted average method is used for fusion, with the weight of the central region of the window being greater than that of the edge region.

10. A physical prior-driven high-resolution optical satellite method for detecting coastal zone changes according to claim 9, characterized in that: Step 5 involves analyzing the semantic change detection results output by the trained deep learning model. The vector patches are processed and then assigned attributes such as previous time phase, next time phase, image source, confidence level, area, perimeter, and change type. The specific process is as follows: Semantic change detection results based on the output of the trained deep learning model Connectivity analysis was performed on water-to-land and land-to-water categories respectively, and patches with an area smaller than the preset minimum mapping unit were deleted. Morphological closure operations were performed on the preserved plaques to connect the narrow fractures, resulting in connected plaques. Semantic change detection results Connected component denoising and morphological correction are performed sequentially to obtain the morphologically corrected image; The morphologically corrected image is then rasterized to obtain a rasterized polygon vector file. Assign attributes such as previous phase, next phase, image source, confidence level, area, perimeter, and variation type to each patch in the rasterized polygon vector file.

Images