Coastal zone ground feature coverage information extraction method and system based on remote sensing cloud platform

Abstract

The invention belongs to the technical field of image processing, and discloses a coastal zone ground feature coverage information extraction method and system based on a remote sensing cloud platform. Multi-source remote sensing data are integrated, and a standardized image is generated after preprocessing; an improved U-Net model is constructed, an encoder of the improved U-Net model adopts a residual convolution module to enhance feature extraction, a decoder recovers details through bilinear interpolation up-sampling and jump connection, the segmentation precision of an artificial shoreline and a natural shoreline is optimized in combination with a weighted cross entropy loss function, and the recognition robustness of a complex ground feature coverage area is improved; and superposing the coastline segmentation result and the ground feature classification data space, generating a vector result, and verifying the precision through a comprehensive evaluation system fusing multiple indexes. According to the method, high-precision extraction of coastal zone ground feature information is realized, and technical support is provided for ocean space planning and implementation and coastal zone resource management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a coastal zone ground cover information extraction method and system based on a remote sensing cloud platform. BACKGROUND

[0002] As a key area of sea-land interaction, the accurate extraction of ground cover information of the coastal zone is crucial for marine spatial planning and implementation and management of coastal zone resources. Traditional remote sensing interpretation methods mostly rely on manual visual interpretation or single algorithm model, and have limitations such as low efficiency, poor generalization and insufficient accuracy. The ground cover types of the coastal zone are complex and diverse, and the forms of artificial shorelines (such as ports and wharfs) and natural shorelines (such as beaches and bedrock) are significantly different. In addition, multi-source remote sensing data (such as Landsat, Sentinel and MODIS) have problems such as heterogeneous spatio-temporal resolution and redundant spectral features, which makes it difficult for traditional methods to balance the boundary segmentation accuracy and classification consistency. In the prior art, methods based on threshold segmentation or single machine learning models (such as decision trees and support vector machines) are easily disturbed by data noise and lack adaptability to complex ground cover scenes, and the evaluation system mostly relies on a single index (such as overall accuracy), which is difficult to fully reflect the model performance. In addition, standard deep learning models such as U-Net easily ignore the form difference between artificial and natural shorelines in the segmentation of the coastline, resulting in blurred segmentation in the weak boundary area and difficulty in meeting the high-precision application requirements. SUMMARY

[0003] To solve the above technical problems, the present application provides a coastal zone ground cover information extraction method and system based on a remote sensing cloud platform.

[0004] In a first aspect, the present application provides a coastal zone ground cover information extraction method based on a remote sensing cloud platform, comprising: Step 1: obtaining multi-source remote sensing image data of a target coastal zone region based on a GEE platform; Step 2: preprocessing the multi-source remote sensing image data to obtain a standardized image data set; Step 3: constructing an improved U-Net classification model, wherein the encoder of the U-Net classification model comprises a residual convolution module, the decoder adopts a bilinear interpolation upsampling and skip connection fusion mechanism, and the loss function is a weighted cross-entropy loss function; Step 4: based on the standardized image data set, using the improved U-Net classification model to extract a preliminary coastline and distinguish artificial shorelines from natural shorelines; Step 5: performing ground cover classification on the standardized image data set; Step 6: performing spatial overlay analysis on the preliminary coastline and the ground cover classification result to generate coastal zone ground cover vector data; Step 7: Based on measured shoreline data and field validation samples, calculate the confusion matrix and comprehensive evaluation index Z to verify the classification accuracy; the comprehensive evaluation index Z includes shoreline edge continuity. Step 8: Iteratively optimize the model parameters for misclassified areas and output the final coastal land cover information.

[0005] Optionally, the structure of the residual convolution module in step 3 is as follows: a batch normalization layer and a ReLU activation function are concatenated after the standard convolutional layer, and skip connections are added to form a residual structure. The expression for its output feature map is: ; Where σ represents the ReLU activation function, BN represents the batch normalization operation, Conv represents the convolution operation with a 3×3 convolution kernel, F(x) represents the output feature map of the residual structure, x represents the input feature map of the residual convolution module, δ represents the tidal influence factor, and G k denoted as the multi-scale gradient operator, and β as the boundary continuity decay coefficient.

[0006] Optionally, the expression for the weighted cross-entropy loss function in step 3 is: ; Among them, L unet N represents the weighted cross-entropy loss function value; batch The variable represents the total number of samples in the current training batch; i represents the sample index variable; y represents the total number of samples in the current training batch. i Let y represent the true label of the i-th sample. i ∈{0,1}, where 1 represents artificial coastline and 0 represents natural coastline; p i This represents the model's predicted probability that the i-th sample belongs to an artificial shoreline, with a value ranging from 0 to p. i ≤1; α is the category weight coefficient of artificial shoreline, ranging from 0.6 to 0.8, λ represents the fracture repair excitation coefficient, C i This represents the severity factor of local fracture.

[0007] Optionally, the base classifier training in step 5 includes: dividing the standardized image dataset into a training set and a validation set in a 7:3 ratio, and using the probability feature map output by the base classifier as the input data for the meta classifier.

[0008] Optionally, the method for segmenting the artificial shoreline in step 4 is: extracting the image edges and connecting the broken boundaries using morphological closing operations; The method for segmenting natural coastlines is as follows: the contrast, correlation, and texture entropy features of the image are calculated based on the gray-level co-occurrence matrix to segment the coastline region.

[0009] Optionally, the preprocessing method in step 2 includes fusing multi-source remote sensing image data, specifically including: performing band synthesis on Landsat series images to generate NDWI index, performing temporal median synthesis on Sentinel series images, and performing spatial downscaling interpolation on MODIS series images.

[0010] Optionally, in step 7, the comprehensive evaluation index Z is based on the entropy method, which integrates overall accuracy (OA), precision (Precision), recall (Recall), F1 score, Kappa coefficient (Kappa), false positive rate (FPR), area under the curve (AUC), Youden index, and shoreline edge continuity. The formula for calculating the comprehensive evaluation index Z is as follows: ; The weighting coefficients ω1 to ω9 are determined using the entropy method, satisfying Σω j =1, YoudenIndex is the Youden index, and Edge_Continuity is the shoreline edge continuity, which is used to quantify the boundary continuity of artificial shorelines.

[0011] Optionally, the iterative optimization of model parameters in step 8 includes: increasing the sample weights for false positive regions in the confusion matrix, adjusting the convolution kernel parameters of the U-Net classification model, and adjusting the splitting threshold of the base classifier during land cover classification.

[0012] Optionally, the spatial overlay analysis method in step 6 is as follows: topologically overlay the coastline vector with the land feature classification grid, use majority voting to determine the final land feature category for the overlapping area, and remove fragmented areas with an area of ​​less than 10 pixels.

[0013] Secondly, the present invention also provides a coastal zone cover information extraction system based on a remote sensing cloud platform, comprising: The acquisition module is used to acquire multi-source remote sensing image data of the target coastal zone based on the GEE platform; The preprocessing module is used to preprocess the multi-source remote sensing image data to obtain a standardized image dataset; The building module is used to build an improved U-Net classification model. The encoder of the U-Net classification model includes a residual convolution module, the decoder adopts a bilinear interpolation upsampling and skip connection fusion mechanism, and the loss function is a weighted cross-entropy loss function. The extraction module is used to extract preliminary coastlines based on the standardized image dataset using the improved U-Net classification model, and to distinguish between artificial coastlines and natural coastlines. The classification module performs land cover classification on the standardized image dataset; The generation module is used to perform spatial overlay analysis on the preliminary coastline and land cover classification results to generate coastal zone land cover vector data. The verification module is used to calculate the confusion matrix and comprehensive evaluation index Z based on measured shoreline data and field verification samples, and to verify the classification accuracy; the comprehensive evaluation index Z includes shoreline edge continuity. The output module is used to iteratively optimize the model parameters for misclassified regions and output the final coastal land cover information.

[0014] The present invention has the following technical effects: This invention significantly improves the accuracy and efficiency of coastal land cover classification by integrating multi-source remote sensing data and employing an improved U-Net model. In the improved U-Net model, the encoder's residual convolution module enhances the extraction capability of weak boundary features (such as intertidal zones) through batch normalization and skip connection mechanisms. The decoder's bilinear interpolation upsampling combined with skip connection fusion restores high-resolution details and avoids artifacts. The weighted cross-entropy loss function alleviates the class imbalance problem between artificial and natural coastlines through differentiated weight allocation, resulting in more complete edge segmentation of artificial coastlines and higher texture consistency of natural coastlines.

[0015] For artificial coastlines, morphological closing operations are employed to enhance edge segmentation, ensuring the boundary continuity of structures such as breakwaters. For natural coastlines, contrast, correlation, and texture entropy features are extracted based on the gray-level co-occurrence matrix before classification, improving accuracy in complex areas such as tidal flats. A comprehensive evaluation Z based on eight indicators using the entropy method is used to fully reflect the model's performance differences across different land cover regions. Combined with parameter optimization driven by the confusion matrix (such as adjusting the sample weights in false positive areas), a dynamic feedback mechanism is formed to continuously improve the spatial consistency of classification results. Based on the above scheme, efficient processing and long-term time-series analysis of large-scale coastal zone data are achieved, providing high-precision support for marine spatial planning and coastal resource management. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A schematic diagram of the method for extracting coastal land cover information based on a remote sensing cloud platform provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the coastal land cover information extraction system based on a remote sensing cloud platform provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] Figure 1 This is a schematic diagram of a method for extracting coastal cover information based on a remote sensing cloud platform, provided in an embodiment of the present invention. The method includes: Step 1: Acquire multi-source remote sensing image data of the target coastal area using the GEE platform; Step 2: Preprocess the multi-source remote sensing image data to obtain a standardized image dataset; Step 3: Construct an improved U-Net classification model. The encoder of the U-Net classification model includes a residual convolution module, and the decoder adopts a fusion mechanism of bilinear interpolation upsampling and skip connections. The loss function is the weighted cross-entropy loss function. Step 4: Based on the standardized image dataset, use the improved U-Net classification model to extract the preliminary coastline and distinguish between artificial coastlines and natural coastlines; Step 5: Perform land cover classification on the standardized image dataset; Step 6: Perform spatial overlay analysis on the preliminary coastline and land cover classification results to generate coastal zone land cover vector data; Step 7: Based on measured shoreline data and field validation samples, calculate the confusion matrix and comprehensive evaluation index Z to verify the classification accuracy; the comprehensive evaluation index Z includes shoreline edge continuity. Step 8: Iteratively optimize the model parameters for misclassified areas and output the final coastal land cover information.

[0020] A method for extracting coastal land cover information based on a remote sensing cloud platform first acquires multi-source remote sensing imagery data of the target area, including Landsat, Sentinel, and MODIS series images, through the Google Earth Engine (GEE) platform. This data covers different temporal and spatial resolutions, providing spectral, textural, and temporal variation information. Radiometric correction is applied to the acquired images to eliminate atmospheric interference, geometric correction ensures spatial coordinate consistency, and cloud-covered areas are filtered using cloud masking to generate a standardized image dataset. The preprocessed data preserves the spectral characteristics and spatial distribution patterns of coastal land cover, providing high-quality input for subsequent model training and classification.

[0021] When constructing the improved U-Net classification model, a residual convolutional structure is introduced into the encoder part. Batch normalization and skip connections enhance feature transfer efficiency and solve the gradient degradation problem of traditional convolutional networks in deep training. The decoder uses bilinear interpolation upsampling instead of traditional deconvolution operations, and combines skip connections to fuse shallow detail features with deep semantic information, improving the ability to recover the edges of artificial shorelines and the textures of natural shorelines. The loss function is designed as a weighted cross-entropy form, dynamically adjusting the weights according to the difference in sample distribution between artificial and natural shorelines to alleviate the impact of class imbalance on segmentation accuracy. After the model is trained, a standardized image dataset is input to extract preliminary coastline results for artificial and natural shorelines respectively. The artificial shoreline segmentation is enhanced by edge gradient to ensure the boundary continuity of structures such as breakwaters and wharves, while the natural shoreline segmentation is based on multi-scale texture features to preserve complex morphologies such as mudflats and bedrock.

[0022] Stacking ensemble learning algorithms can be used for land cover classification. Support vector machines, random forests, and decision trees are selected as base classifiers, extracting differential information from spectral reflectance, texture statistics, and temporal variation features, respectively. The output probability feature maps of the base classifiers serve as input to the logistic regression meta-classifier. Multi-model collaborative decision-making reduces the risk of misclassification by a single classifier in mixed land cover scenarios such as aquaculture areas and vegetated areas. The ensemble learning strategy effectively suppresses noise interference and overfitting, improving the classification consistency of land cover types such as woodland, farmland, and water bodies.

[0023] The preliminary coastline segmentation results and land cover classification results are spatially overlaid and analyzed. Raster-to-vector conversion generates coastal zone vector data including shoreline attributes and land cover type labels. A confusion matrix is ​​constructed based on field-measured shoreline data and validation samples, and a comprehensive evaluation index Z is calculated. This index integrates overall accuracy, precision, recall, F1 score, Kappa coefficient, false positive rate, area under the curve, and Youden index, dynamically allocating weights using the entropy method to comprehensively evaluate the model's performance differences across different land cover categories. For misclassified areas, the classification effect is iteratively improved by adjusting sample weights and optimizing model parameters, ultimately outputting high-precision coastal zone land cover information.

[0024] By combining multi-source data fusion, improved model structure, and ensemble learning strategies, this method demonstrates outstanding performance in terms of clarity of coastline boundary segmentation, accuracy of land feature classification, and adaptability to complex scenarios, providing a reliable data foundation for marine spatial planning and marine resource management.

[0025] In some implementations, the structure of the residual convolution module in step 3 is as follows: a batch normalization layer and a ReLU activation function are concatenated after the standard convolutional layer, and skip connections are added to form a residual structure. The expression for its output feature map is: ; Where σ represents the ReLU activation function, BN represents the batch normalization operation, Conv represents the convolution operation with a 3×3 convolution kernel, F(x) represents the output feature map of the residual structure, x represents the input feature map of the residual convolution module, γ represents the tidal influence factor, and G k denoted as the multi-scale gradient operator, and β as the boundary continuity decay coefficient.

[0026] Where γ = 0.2 × (daily tidal range / average tidal range). β = 1 - 0.5 × (predicted fracture length / total shoreline length). Diurnal tidal range represents the maximum daily tidal level change in the target area, mean tidal range represents the average annual tidal range of the target area, and γ is used to adjust the contribution of the gradient enhancement term to the convolution result. The scale parameter is σ k The Gaussian Laplacian operator, scale-adaptive, σ k Let σ be the Gaussian kernel standard deviation. k =2k, where k is the scale level. The larger k is, the more it represents the macroscopic coastline; the smaller k is, the more it represents the microscopic reefs. -k G is the scale attenuation coefficient. k It can solve the problem of large differences in the scale of coastal features. The predicted fault length represents the total length of discontinuous shoreline segments in the model output, and the total shoreline length represents the theoretical shoreline length of the target area. β can solve the problem of faults caused by shading and insufficient resolution of artificial shorelines.

[0027] In the improved U-Net classification model encoder, the construction process of the residual convolution module is as follows: A standard convolution operation with a 3×3 kernel is performed on the input feature map. The feature map output from the convolutional layer undergoes batch normalization to stabilize the training process and accelerate convergence. Subsequently, a ReLU activation function is used to introduce non-linear expressive power, enhancing the discriminative power of the features. During this process, the input feature map and the output feature map after convolution-normalization-activation are directly added through skip connections to form the residual structure. The mathematical expression for this process is shown above.

[0028] In this process, the input feature map x is processed by 3×3 convolution (Conv) to generate preliminary features, batch normalization (BN) is used to adjust the feature distribution, the ReLU activation function (σ) introduces a non-linear transformation, and the skip connection adds the original input x with the transformed features to ensure effective gradient propagation and alleviate the degradation problem in deep network training.

[0029] In the residual convolution module, the batch normalization layer standardizes the mean and variance of the features output by the convolution, stabilizing the feature distribution and avoiding training oscillations caused by differences in input distribution between layers. The ReLU activation function thresholds the normalized features, suppressing negative responses and preserving positive features, enhancing the model's sensitivity to the edges and textures of coastal features. Skip connections fuse the original input with the transformed features, preserving shallow details while transmitting deep semantic features. This ensures that the edge gradient features of artificial coastlines and the multi-scale texture features of natural coastlines are fully preserved during the encoding process, avoiding feature loss caused by excessive depth in traditional convolutional networks.

[0030] Through the above structure, the residual convolution module enhances the model's feature extraction capability and ensures the stability of the training process. It is especially suitable for the joint representation of weak boundaries (such as tidal channels and shoals) and complex textures (such as reefs and vegetation) in coastal images, and significantly improves the boundary continuity and internal consistency of the segmentation results.

[0031] In some implementations, the expression for the weighted cross-entropy loss function in step 3 is: ; Among them, L unet N represents the weighted cross-entropy loss function value; batch The variable represents the total number of samples in the current training batch; i represents the sample index variable; y represents the total number of samples in the current training batch. i Let y represent the true label of the i-th sample. i ∈{0,1}, where 1 represents artificial coastline and 0 represents natural coastline; p i This represents the model's predicted probability that the i-th sample belongs to an artificial shoreline, with a value ranging from 0 to p. i ≤1; α is the category weight coefficient of artificial shoreline, ranging from 0.6 to 0.8, δ represents the fracture repair excitation coefficient, C i This represents the severity factor of local fracture.

[0032] Experiments show that artificial coastlines typically constitute a low proportion of coastal features, exhibiting a highly unbalanced class distribution. Through grid search and cross-validation, it was found that when α ranges from 0.6 to 0.8, the model significantly improves the recall rate for artificial coastlines while maintaining a relatively stable F1 score for natural coastlines, achieving an optimal balance between class balance and overall accuracy. This range is dynamically adjusted by comparing the confusion matrix of the validation set under different α values ​​to ensure the integrity of artificial coastline boundary segmentation while avoiding over-segmentation of natural coastlines.

[0033] In the training process of the improved U-Net classification model, a weighted cross-entropy loss function is designed to mitigate the model bias problem caused by the imbalance in the number of artificial and natural shoreline samples. The specific form of the loss function is shown above. Where C...i = (Break length of shoreline segment containing sample i / Total length of shoreline segment), where the shoreline segment containing sample i can be a shoreline interval extending a certain distance to the left and right of sample i as the center, the break length can be the total length of discontinuous shoreline within this interval, and the total length of the shoreline segment can be the theoretical continuous length of the shoreline segment, C i =1 represents complete fracture, C i =0 represents complete continuity. δ controls the excitation intensity for fracture repair, δ=0.15+0.05×(regional historical maximum fracture rate / 0.3). This is achieved by introducing a local fracture severity factor C. i By quantifying shoreline morphological defects and embedding them into the fault repair excitation coefficient δ, high-fracture areas can be prioritized for optimization, spatial local perception capabilities can be improved, and real-time feedback can be achieved. C is updated in each iteration. i In response to the latest forecasts.

[0034] During model training, α is set to a value greater than 0.5. By increasing the loss weight of artificial shoreline samples, the model is forced to pay more attention to the classification error of this category. For example, when artificial shorelines account for a significantly smaller proportion of the training data than natural shorelines, increasing α can strengthen the model's learning of artificial shoreline features during backpropagation, avoiding missed detections due to differences in sample size. In the formula, This item applies a weighted penalty to the prediction error of artificial shoreline samples. The error of the natural shoreline is dynamically adjusted, and the two work together to balance the learning intensity of the model for different categories of features.

[0035] During training, the loss function optimizes model parameters using gradient descent. For artificial shoreline areas (such as breakwaters and wharves), the model, weighted by α, enhances its sensitivity to abrupt changes in edge gradients, reducing boundary breaks or blurring caused by insufficient samples. For natural shoreline areas (such as mudflats and bedrock), the model reduces the error weight of natural shorelines to avoid overfitting to noise interference from complex textured backgrounds. This loss function design enables the model to significantly improve the boundary integrity and internal connectivity of artificial shorelines while maintaining the consistency of natural shoreline segmentation.

[0036] By introducing class weight coefficients, this method effectively alleviates the class imbalance problem that is common in coastal land cover classification, enabling the model to achieve balanced classification performance in different scenarios. It is especially suitable for complex coastal environments where densely built-up areas and natural ecological areas coexist.

[0037] In some implementations, the base classifier training in step 5 of the classification process includes: dividing the standardized image dataset into a training set and a validation set in a 7:3 ratio, and using the probability feature map output by the base classifier as the input data for the meta classifier.

[0038] The 7:3 split ratio was determined based on the spatial distribution characteristics of the images and the generalization requirements of machine learning. A 70% training set adequately covers the spatial heterogeneity of coastal features (such as the intertidal zone, ports, and vegetated areas), while a 30% validation set ensures the independence and statistical significance of the meta-classifier's input data. Testing showed that this ratio resulted in lower Kappa coefficient fluctuations for the base classifier on the validation set, significantly outperforming other split ratios such as 6:4 or 8:2, indicating that it effectively balances model complexity and the risk of overfitting.

[0039] In the base classifier training phase of the Stacking ensemble learning algorithm, the standardized image dataset is divided into a training set and a validation set according to a preset ratio. The training set is used for independent model fitting of the base classifiers, and the validation set is used to generate the input features of the meta-classifier. After each base classifier (including support vector machines, random forests, and decision trees) completes parameter learning on the training set, it predicts the probability distribution of the validation set samples and their respective land cover categories. For a classification task with C land cover categories, the output probability of the m-th base classifier for the n-th sample is represented as a vector. ,in This represents the probability that the m-th base classifier predicts the n-th sample will be classified into class c. The probability outputs of all base classifiers are concatenated horizontally in sample order to construct the training feature matrix of the meta-classifier. Its dimensions are N×(M×C), where N is the total number of validation set samples and M is the number of base classifiers.

[0040] Logistic regression models, as meta-classifiers, receive... As input features, the probability prediction results of the base classifiers are fused by optimizing the weight coefficients. The loss function of the meta-classifier. Defined as: ; in, Let n be the total number of samples in the validation set, and n be the sample index in the validation set (n=1,2,…,N). val C represents the total number of land cover categories (such as artificial shoreline, natural shoreline, water area, woodland, etc.), c represents the index of land cover category (c=1,2,…,C), M represents the number of base classifiers (M=3, i.e., support vector machine, random forest, decision tree), and m represents the index of base classifier (m=1,2,…,M). Predict the probability that the nth sample belongs to class c for the mth base classifier. The true label (using one-hot encoding) for the nth sample belonging to the cth category. =1 indicates that sample n belongs to category c. =0 indicates that sample n does not belong to category c. Let be the combined weights of the m-th base classifier. The weight parameters are iteratively updated using the gradient descent algorithm to make the fused probability output approximate the true distribution. After training, the meta-classifier can adaptively adjust the contribution of each base classifier to specific land cover categories; for example, it can enhance the decision weights of the support vector machine in areas with artificial buildings, and emphasize the output of the random forest in areas with vegetation cover.

[0041] This method effectively integrates the complementary advantages of multiple classifiers through spatial fusion and dynamic weighting of probabilistic features, reduces misclassification of single models in spectrally confused regions (such as aquaculture ponds and natural water bodies), and suppresses overfitting caused by local bias in training data, thus significantly improving the spatial consistency and class discrimination of land cover classification results.

[0042] In some implementations, the method for segmenting the artificial shoreline in step 4 is: extracting the image edges and connecting the broken boundaries using morphological closing operations; The method for segmenting natural coastlines is as follows: the contrast, correlation, and texture entropy features of the image are calculated based on the gray-level co-occurrence matrix to segment the coastline region.

[0043] To address the segmentation requirements between artificial and natural coastlines, differentiated edge detection and texture analysis methods are employed. For artificial coastlines, Canny edge detection is first performed on the preprocessed image. Gaussian filtering is used to suppress noise interference, and the image gradient magnitude and direction are calculated. Combined with non-maximum suppression and double thresholding, initial edge contours are extracted. Since artificial structures (such as wharves and breakwaters) typically have regular geometric shapes and high-contrast boundaries, but may exhibit breaks due to image resolution or occlusion, morphological closing operations are further used to connect and fill the edges. A rectangular kernel is chosen as the structural element for the closing operation. A dilation-erosion operation is used to close small gaps, enhancing the boundary continuity of the artificial coastline and ensuring the complete representation of structures such as breakwaters and revetments.

[0044] For natural coastlines, multi-scale texture features are extracted based on the Gray-Level Co-occurrence Matrix (GLCM). Contrast, correlation, and texture entropy are calculated within a sliding window. Contrast reflects the intensity of local texture changes, correlation describes gray-level spatial dependencies, and texture entropy characterizes texture complexity. For example, tidal flats exhibit fine granular textures due to tidal action, resulting in lower contrast but higher entropy; bedrock coastlines, due to weathering and erosion forming irregular fissures, show significantly higher contrast and correlation than tidal flats. The extracted texture features are used as input to train a random forest classifier. Multiple decision trees vote on the texture features to suppress the sensitivity of a single classifier to local noise, thus improving the robustness of natural coastline segmentation.

[0045] The above methods effectively address the regularity of artificial coastline edges and the complexity of natural coastline textures. The combination of the Canny operator and morphological closing operations ensures clear and complete boundaries of artificial structures, while the collaborative analysis of gray-level co-occurrence matrices and random forests enhances the recognition capability of complex natural coastline textures. This method effectively distinguishes between two types of coastlines in mixed scenarios such as ports and bedrock coastlines, reducing missegmentation caused by feature confusion and improving the spatial accuracy and visual consistency of coastal feature classification.

[0046] In some implementations, the preprocessing method in step 2 includes fusing multi-source remote sensing image data, specifically including: performing band synthesis on Landsat series images to generate NDWI index, performing temporal median synthesis on Sentinel series images, and performing spatial downscaling interpolation on MODIS series images.

[0047] In the fusion processing of multi-source remote sensing image data, for Landsat series images, the Normalized Difference Water Index (NDWI) is generated by band calculation to enhance the spectral contrast between water bodies and land. Specifically, near-infrared and green band reflectance data are selected, and the NDWI value of each pixel is calculated according to the formula, highlighting the features of coastal waters, tidal channels, and wetlands, while suppressing interference from vegetation and bare land. For Sentinel series images, based on long-term series data, a median composite method is used to generate quarterly or annual composite images. By calculating the temporal median pixel by pixel, transient noise such as cloud cover and atmospheric disturbances is effectively eliminated, while preserving the stable spectral characteristics of ground features.

[0048] To address the low spatial resolution limitations of MODIS series imagery, a spatial downscaling interpolation method is employed to enhance resolution. The original imagery is resampled to a high-resolution grid using a bilinear interpolation algorithm and spatially aligned with Landsat or Sentinel imagery to ensure pixel-level matching of multi-source data within the same coordinate system. During interpolation, the spectral similarity of neighboring pixels is considered to maintain smooth transitions at ground feature boundaries and avoid jagged artifacts.

[0049] After fusing multi-source data, the enhanced water body information from the NDWI index, the stable spectral features from Sentinel temporal median synthesis, and the spatial consistency from MODIS downscaling interpolation complement each other, forming a standardized image dataset with high spatial resolution, temporal continuity, and spectral discriminability. This dataset can clearly express the regular geometric shape of artificial coastlines, the complex texture details of natural coastlines, and the dynamic changes of the intertidal zone, providing high-quality input for subsequent model training and significantly improving the accuracy and reliability of coastline segmentation and land cover classification.

[0050] In some implementations, the comprehensive evaluation index Z in step 7 is based on the entropy method, which integrates overall accuracy (OA), precision (Precision), recall (Recall), F1 score, Kappa coefficient, false positive rate (FPR), area under the curve (AUC), Youden index, and shoreline edge continuity. The formula for calculating the comprehensive evaluation index Z is as follows: ; The weighting coefficients ω1 to ω9 are determined using the entropy method, satisfying Σω j =1, YoudenIndex is the Youden index, and Edge_Continuity is the shoreline edge continuity, which is used to quantify the boundary continuity of artificial shorelines.

[0051] This scheme adds the concept of shoreline edge continuity to the traditional calculation method of comprehensive evaluation index Z. Through this improvement, the comprehensive evaluation index Z can be calculated more accurately.

[0052] Specifically, Edge_Continuity can be calculated as follows: Edge_Continuity = (Length of closed shoreline / Length of original shoreline) × (Number of artificial shoreline breaks / N) total ), where N total This represents the total number of original fracture boundaries of the artificial coastline. For example, if there are 5 fractures in the original artificial coastline, then Ntotal = 5; if 3 of the fractures are connected, then the numerator is 3.

[0053] Σω j The constraint of 1 is achieved through entropy normalization, ensuring the comparability of the weight coefficients of each evaluation indicator. Without normalization, the weights of high-entropy indicators (such as OA and Kappa) may be underestimated, leading to a bias in the overall evaluation towards local indicators (such as Precision and FPR). After normalization, the entropy method can dynamically reflect the information differences between indicators, making the Z-index more aligned with actual classification needs.

[0054] The comprehensive evaluation index Z is calculated based on the entropy method, which integrates multiple single precision indices to comprehensively evaluate the performance differences of the model in coastal land cover classification. First, for eight indices—overall accuracy, precision, recall, F1 score, Kappa coefficient, false positive rate, area under the curve, and Youden index—the results on the validation sample set are calculated to construct an initial evaluation matrix. For a validation set consisting of N samples, the value of the j-th index for the n-th sample is denoted as ρ. nj By eliminating dimensional differences through normalization, the standardized index value p is obtained. nj : ; Subsequently, the entropy value e of the j-th index is calculated.j : ; Entropy reflects the dispersion of indicator data; the smaller the entropy value, the higher the discrimination ability of the indicator in classification. The weight coefficients of each indicator are calculated based on the entropy value. : ;e k This represents the entropy value of the k-th index, and its calculation method (principle) is the same as e. j Same; I j Let I be the land cover importance coefficient of the j-th indicator. k Similarly, for example, an artificial coastline can be 1.5, while a natural coastline can be 1.0.

[0055] Ultimately, the comprehensive evaluation index Z final Represented as: .

[0056] Weights are dynamically allocated using the entropy method, Z final The metrics can adaptively highlight the role of key performance indicators. For example, in areas with dense artificial coastlines, the weights of the false positive rate and Youden index are significantly increased, suppressing errors of misidentifying breakwaters as natural coastlines; in areas with complex textures of natural coastlines, the weights of recall rate and F1 score are increased, reducing the under-scoring of tidal flats and bedrock. During training, Z... final Indicators serve as feedback signals to drive model parameter optimization, such as adjusting loss function weights or enhancing the base classifier's focus on specific land features. This method overcomes the limitations of traditional single-indicator evaluation, quantifying the comprehensive performance of the classification model from multiple dimensions. It provides a scientific and interpretable evaluation tool for coastal land feature classification, supporting model iteration and engineering applications. Here, λ is the nonlinear interaction coefficient, which can range from 0.15 to 0.25, with 0.2 being preferred. Recall is the model recall rate. Adding the λ×(Recall×Edge_Continuity) term enables a non-linear synergy between Recall (the ability to detect fractured shorelines) and Edge_Continuity (the effectiveness of morphological restoration), resulting in higher efficiency through joint optimization. It forces the model to simultaneously meet the Edge_Continuity threshold (Edge_Continuity ≥ 0.8) when the recall rate is high (Recall > 0.9), eliminating the negative correlation between high recall and a surge in fractures in traditional linear weighting. λ is set independently of the weighting system (∑ωj = 1), precisely controlling the synergy strength to avoid higher-order terms interfering with the balance of basic indicators, thus reducing model tuning complexity. The integrity-detection rate balance of artificial shoreline fracture boundary restoration is effectively improved, and the standard deviation of the weighting system's stability is lower.

[0057] In some implementations, the iterative optimization of model parameters in step 8 includes: increasing the sample weights for false positive regions in the confusion matrix, adjusting the convolution kernel parameters of the U-Net classification model, and adjusting the splitting threshold of the base classifier during land cover classification.

[0058] The iterative optimization phase of the model parameters focuses on targeted improvements to false positive regions in the confusion matrix and dynamic adjustments to the model parameters. False positive regions refer to the set of pixels where the model misclassifies non-target features as target categories, such as misclassifying bare land as artificial shorelines or vegetation-covered areas as natural shorelines. To address this type of error, during training, based on the statistical results of the confusion matrix, false positive samples are assigned higher loss weights, forcing the model to prioritize correcting such misclassifications during backpropagation. Weight adjustments are achieved by modifying the weight coefficients of the corresponding samples in the loss function, enabling the model to strengthen feature learning of easily confused regions in subsequent iterations and suppress misclassifications caused by similar spectral or textural backgrounds.

[0059] In the parameter optimization stage, the Adam optimizer adaptively adjusts the convolution kernel parameters of the U-Net classification model and the splitting threshold of the ensemble learning base classifier. The Adam optimizer combines a momentum term with an adaptive learning rate mechanism, dynamically adjusting the update step size based on the first and second moments of the parameter gradients. For the U-Net model, the optimizer focuses on adjusting the convolution kernel weights of the residual convolution module in the encoder to enhance the extraction capability of gradient features from artificial shoreline edges; simultaneously, it optimizes the parameters of the upsampling layer in the decoder to improve the reconstruction effect of natural shoreline texture details. For ensemble learning base classifiers (such as random forests), the optimizer adjusts the splitting threshold of the decision tree nodes to improve the discrimination of spectrally confused regions (such as water bodies and shadows).

[0060] Through the synergy of false positive sample weight adjustment and the Adam optimizer, the model converges rapidly to a stable state during training, and the classification boundaries gradually become clearer. Increasing the weight of false positive regions forces the model to widen the distance between easily confused categories in the feature space, for example, enhancing the spectral-texture differences between artificial shorelines (such as docks) and bare land, and between natural shorelines (such as mudflats) and vegetation. The adaptive learning rate mechanism of the Adam optimizer avoids the oscillation phenomenon during parameter updates in traditional gradient descent methods, ensuring stable convergence of the model in complex coastal scenarios.

[0061] This method demonstrates significant advantages in mixed landform areas such as ports, bedrock coasts, and intertidal zones. Dynamic adjustment of false positive weights reduces misclassification of artificial coastlines in shaded areas, while the Adam optimizer's fine-tuning of convolution kernel parameters enhances the model's sensitivity to weak boundary features, such as subtle textural variations in natural coastlines like tidal channels and reefs. Optimization of the splitting threshold in the ensemble learning base classifier further suppresses overfitting caused by local biases in the training data, ensuring the model's generalization ability under different lighting conditions and seasonal variations. Ultimately, the iteratively optimized model significantly improves the boundary integrity and classification confidence of artificial coastlines while maintaining consistency in natural coastline segmentation, providing highly reliable data support for coastal resource surveys and the formulation and implementation of marine spatial planning.

[0062] In some implementations, the spatial overlay analysis method in step 6 is as follows: topologically overlay the coastline vector with the land feature classification grid, use majority voting to determine the final land feature category for the overlapping area, and remove fragmented areas with an area of ​​less than 10 pixels.

[0063] Fragmented areas smaller than 10 pixels mainly originate from image noise or local model misclassification (such as cloud shadows or wave spray). Removing smaller patches can eliminate most isolated noise, and the loss of effective object information is negligible. This threshold was determined by analyzing the impact of fragments of different scales on the classification results, thereby improving the readability and engineering applicability of the mapping results while maintaining the spatial continuity of ground features.

[0064] In the spatial overlay analysis phase, preliminary coastline vector data is topologically overlaid with land cover classification raster data. First, a raster-to-vector conversion tool is used to convert the coastline vectors into raster data with the same spatial resolution as the classification raster, ensuring precise matching of pixel locations. During the overlay process, for each area covered by the coastline vector, the corresponding land cover classification raster pixel values ​​are extracted. A majority voting method is used to determine the percentage of pixels for each land cover category within that area, and the category with the highest percentage is used as the final land cover type label for that area. For example, in the area where port boundaries intersect with adjacent waters, the majority voting method can eliminate sporadic misclassified pixels caused by spectral confusion, ensuring that waters or facilities within the port structure are correctly classified as artificial shoreline appendages.

[0065] For isolated fragmented areas that are too small, morphological filtering is used to remove them. The classification result raster is traversed to identify fragments with connected regions smaller than a preset threshold, and their category labels are replaced with the dominant land cover type of the neighboring area. This operation effectively eliminates small erroneous patches caused by image noise or local modeling, such as isolated small water area pixels formed by tidal action in tidal flat areas. After fragmentation, these pixels are merged into neighboring tidal flat or land categories, avoiding unreasonable sporadic distributions in the classification map.

[0066] Through a collaborative process of majority voting and fragmented area removal, spatial overlay analysis significantly improves the spatial continuity and visual consistency of classification results while preserving the regular geometric features of artificial coastlines and the textural details of natural coastlines. The resulting coastal cover vector data has clear boundaries and pure categories.

[0067] Figure 2 A schematic diagram of the coastal land cover information extraction system based on a remote sensing cloud platform provided in this embodiment of the invention includes: Acquisition Module 1 is used to acquire multi-source remote sensing image data of the target coastal zone based on the GEE platform; Preprocessing module 2 is used to preprocess multi-source remote sensing image data to obtain a standardized image dataset; Module 3 is used to build an improved U-Net classification model. The encoder of the U-Net classification model includes a residual convolution module, and the decoder adopts a bilinear interpolation upsampling and skip connection fusion mechanism. The loss function is the weighted cross-entropy loss function. Extraction module 4 is used to extract preliminary coastlines based on a standardized image dataset using an improved U-Net classification model, and to distinguish between artificial coastlines and natural coastlines. Classification module 5 is used to classify ground cover in standardized image datasets; Module 6 is used to perform spatial overlay analysis on the preliminary coastline and land cover classification results to generate coastal zone land cover vector data. Validation module 7 is used to calculate the confusion matrix and comprehensive evaluation index Z based on measured shoreline data and field validation samples, and to verify the classification accuracy; the comprehensive evaluation index Z includes shoreline edge continuity. Output module 8 is used to iteratively optimize model parameters for misclassified regions and output the final coastal land cover information.

[0068] The system provided in this embodiment of the invention has the same technical features as the method described above, and therefore can solve the same technical problems and achieve the same technical effects, which will not be elaborated here.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting coastal cover information based on a remote sensing cloud platform, characterized in that, include: Step 1: Acquire multi-source remote sensing image data of the target coastal area using the GEE platform; Step 2: Preprocess the multi-source remote sensing image data to obtain a standardized image dataset; Step 3: Construct an improved U-Net classification model. The encoder of the U-Net classification model includes a residual convolution module, and the decoder adopts a bilinear interpolation upsampling and skip connection fusion mechanism. The loss function is the weighted cross-entropy loss function. Step 4: Based on the standardized image dataset, use the improved U-Net classification model to extract preliminary coastlines and distinguish between artificial coastlines and natural coastlines; Step 5: Perform land cover classification on the standardized image dataset; Step 6: Perform spatial overlay analysis on the preliminary coastline and land cover classification results to generate coastal zone land cover vector data; Step 7: Based on measured shoreline data and field validation samples, calculate the confusion matrix and comprehensive evaluation index Z to verify the classification accuracy; the comprehensive evaluation index Z includes shoreline edge continuity. Step 8: Iteratively optimize the model parameters for misclassified areas and output the final coastal land cover information.

2. The method according to claim 1, characterized in that, The structure of the residual convolution module in step 3 is as follows: a batch normalization layer and a ReLU activation function are concatenated after the standard convolutional layer, and skip connections are added to form a residual structure. Its output feature map expression is: ； Where σ represents the ReLU activation function, BN represents the batch normalization operation, Conv represents the convolution operation with a 3×3 convolution kernel, F(x) represents the output feature map of the residual structure, x represents the input feature map of the residual convolution module, γ represents the tidal influence factor, and G k denoted as the multi-scale gradient operator, and β as the boundary continuity decay coefficient.

3. The method according to claim 1, characterized in that, The expression for the weighted cross-entropy loss function in step 3 is: ； Among them, L unet N represents the weighted cross-entropy loss function value; batch The variable represents the total number of samples in the current training batch; i represents the sample index variable; y represents the total number of samples in the current training batch. i Let y represent the true label of the i-th sample. i ∈{0,1}, where 1 represents artificial coastline and 0 represents natural coastline; p i This represents the model's predicted probability that the i-th sample belongs to an artificial shoreline, with a value ranging from 0 to p. i ≤1; α is the category weight coefficient of artificial shoreline, ranging from 0.6 to 0.8, δ represents the fracture repair excitation coefficient, C i This represents the severity factor of local fracture.

4. The method according to claim 1, characterized in that, The base classifier training in step 5 of the classification process includes: dividing the standardized image dataset into a training set and a validation set in a 7:3 ratio, and using the probability feature map output by the base classifier as the input data for the meta classifier.

5. The method according to claim 1, characterized in that, The method for segmenting the artificial shoreline in step 4 is as follows: extract the image edges and connect the broken boundaries by combining morphological closing operations. The method for segmenting natural coastlines is as follows: the contrast, correlation, and texture entropy features of the image are calculated based on the gray-level co-occurrence matrix to segment the coastline region.

6. The method according to claim 1, characterized in that, The preprocessing method in step 2 includes fusing multi-source remote sensing image data, specifically including: performing band synthesis on Landsat series images to generate NDWI index, performing temporal median synthesis on Sentinel series images, and performing spatial downscaling interpolation on MODIS series images.

7. The method according to claim 1, characterized in that, In step 7, the comprehensive evaluation index Z is based on the entropy method, which integrates overall accuracy (OA), precision (Precision), recall (Recall), F1 score, Kappa coefficient, false positive rate (FPR), area under the curve (AUC), Youden index, and shoreline edge continuity. The formula for calculating the comprehensive evaluation index Z is as follows: ； The weighting coefficients ω1 to ω9 are determined using the entropy method, satisfying Σω j =1, YoudenIndex is the Youden index, and Edge_Continuity is the shoreline edge continuity, which is used to quantify the boundary continuity of artificial shorelines.

8. The method according to claim 1, characterized in that, The iterative optimization of model parameters in step 8 includes: increasing the sample weights for false positive regions in the confusion matrix, adjusting the convolution kernel parameters of the U-Net classification model, and adjusting the splitting threshold of the base classifier during land cover classification.

9. The method according to claim 1, characterized in that, The spatial overlay analysis method in step 6 is as follows: topologically overlay the coastline vector with the land feature classification grid, use the majority voting method to determine the final land feature category for the overlapping area, and remove fragmented areas with an area of ​​less than 10 pixels.

10. A coastal zone cover information extraction system based on a remote sensing cloud platform, characterized in that, include: The acquisition module is used to acquire multi-source remote sensing image data of the target coastal zone based on the GEE platform; The preprocessing module is used to preprocess the multi-source remote sensing image data to obtain a standardized image dataset; The building module is used to build an improved U-Net classification model. The encoder of the U-Net classification model includes a residual convolution module, the decoder adopts a bilinear interpolation upsampling and skip connection fusion mechanism, and the loss function is a weighted cross-entropy loss function. The extraction module is used to extract preliminary coastlines based on the standardized image dataset using the improved U-Net classification model, and to distinguish between artificial coastlines and natural coastlines. The classification module performs land cover classification on the standardized image dataset; The generation module is used to perform spatial overlay analysis on the preliminary coastline and land cover classification results to generate coastal zone land cover vector data. The verification module is used to calculate the confusion matrix and comprehensive evaluation index Z based on measured shoreline data and field verification samples, and to verify the classification accuracy; the comprehensive evaluation index Z includes shoreline edge continuity. The output module is used to iteratively optimize the model parameters for misclassified regions and output the final coastal land cover information.

Images