High-precision survey monitoring method for land subsidence of reclamation area

Abstract

The invention relates to the field of intelligent analysis, and provides a reclamation area land subsidence high-precision survey monitoring method, which comprises the following steps: carrying out ROI generation and visual multi-scale feature extraction on a first reclamation area image and a second reclamation area image by using an image recognition and analysis algorithm based on machine vision; local neighborhood perception enhancement is carried out on the extracted first reclamation area ROI visual feature description feature and the extracted second reclamation area ROI visual feature description feature; and intelligently obtaining a settlement monitoring result of the settlement height based on visual difference feature representation between the enhanced first enhanced reclamation area ROI visual feature description feature and the enhanced second enhanced reclamation area ROI visual feature description feature. Therefore, the accuracy of reclamation area land subsidence height calculation can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent analysis, and more specifically, to a high-precision investigation and monitoring method for ground subsidence in reclamation areas. Background Art

[0002] Reclamation refers to land that was originally unsuitable for construction and has been transformed into usable land through land reclamation, river diversion or wetland drainage. However, these reclaimed lands are usually located in saturated soft soil areas, such as coastal tidal flats or estuary deltas, and are therefore prone to ground subsidence. Ground subsidence not only affects the safety of buildings and infrastructure, but can also lead to a series of safety issues such as damage to underground pipelines, restrictions on underground space development and even flooding.

[0003] Traditional leveling and GPS measurement methods have obvious deficiencies in monitoring range. Specifically, leveling is a classic ground elevation measurement method that relies on manual operation to set measurement points along a specific route at fixed time intervals. It is not only inefficient and costly, but also difficult to achieve high-frequency or real-time monitoring. Although GPS measurement is relatively flexible, the signal is easily affected in complex terrain or signal-blocked areas, such as mountainous areas in reclamation areas and densely built-up areas, and the detection accuracy of small settlements is limited, resulting in some areas being unable to be effectively monitored.

[0004] Therefore, a high-precision survey and monitoring method for land subsidence in reclamation areas is desired. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present application provides a high-precision survey and monitoring method for ground subsidence in reclamation areas.

[0006] According to one aspect of the present application, a high-precision survey and monitoring method for ground subsidence in a reclamation area is provided, which includes: acquiring a first reclamation area image and a second reclamation area image at a first time point and a second time point acquired by a camera; performing dynamic ROI generation on the first reclamation area image and the second reclamation area image to obtain a first reclamation area ROI image and a second reclamation area ROI image; performing visual multi-scale feature extraction on the first reclamation area ROI image and the second reclamation area ROI image respectively to obtain a first reclamation area ROI visual feature description feature map and a second reclamation area ROI visual feature description feature map; performing local neighborhood-aware feature enhancement on the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map respectively to obtain a first enhanced reclamation area ROI visual feature description feature map and a second enhanced reclamation area ROI visual feature description feature map; calculating a reclamation area ROI visual difference feature description feature map between the first enhanced reclamation area ROI visual feature description feature map and the second enhanced reclamation area ROI visual feature description feature map; and obtaining a settlement monitoring result based on the reclamation area ROI visual difference feature description feature map.

[0007] The present application has significant technical effects due to the adoption of the above technical solutions: the high-precision survey and monitoring method for land subsidence in reclamation areas provided by the present application uses an image recognition and analysis algorithm based on machine vision to generate ROIs and extract visual multi-scale features of the first reclamation area image and the second reclamation area image, and then performs local neighborhood perception enhancement on the extracted first reclamation area ROI visual feature description features and the second reclamation area ROI visual feature description features, thereby intelligently obtaining the settlement monitoring results of the settlement height based on the visual difference feature representation between the enhanced first enhanced reclamation area ROI visual feature description features and the second enhanced reclamation area ROI visual feature description features. In this way, the accuracy of calculating the land subsidence height in the reclamation area can be effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 It is a flow chart of a high-precision investigation and monitoring method for ground subsidence in a reclamation area according to an embodiment of the present application.

[0010] Figure 2 Schematic diagram of data flow of a high-precision investigation and monitoring method for ground subsidence in reclamation areas according to an embodiment of the present application.

[0011] Figure 3 This is a flowchart of step S140 in the method for high-precision investigation and monitoring of ground subsidence in reclamation areas according to an embodiment of the present application. DETAILED DESCRIPTION

[0012] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0013] In view of the above technical problems, the technical concept of the present application is to obtain the first reclamation area image and the second reclamation area image at the first time point and the second time point collected by the camera, and use the image recognition and analysis algorithm based on machine vision to generate ROI and extract visual multi-scale features of the first reclamation area image and the second reclamation area image, and then perform local neighborhood perception enhancement on the extracted first reclamation area ROI visual feature description features and the second reclamation area ROI visual feature description features, so as to intelligently obtain the settlement monitoring results of the settlement height based on the visual difference feature representation between the enhanced first enhanced reclamation area ROI visual feature description features and the second enhanced reclamation area ROI visual feature description features. Compared with the leveling measurement that is limited to setting measurement points along a specific route, the camera can be flexibly installed at any position in the reclamation area to achieve comprehensive coverage of large-area reclamation areas. In addition, the present application can obtain image data at any time as needed, capture surface deformation information at different scales, so that even in a complex background, subtle surface changes can be clearly distinguished and small settlement amounts can be detected, thereby improving the accuracy of settlement height calculation.

[0014] Figure 1 It is a flow chart of a high-precision investigation and monitoring method for ground subsidence in a reclamation area according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the high-precision survey and monitoring method for land subsidence in reclamation areas according to an embodiment of the present application. Figure 1 and Figure 2As shown, according to the embodiment of the present application, the high-precision survey and monitoring method for land subsidence in a reclamation area includes: S110, acquiring a first reclamation area image and a second reclamation area image at a first time point and a second time point acquired by a camera; S120, performing dynamic ROI generation on the first reclamation area image and the second reclamation area image to obtain a first reclamation area ROI image and a second reclamation area ROI image; S130, performing visual multi-scale feature extraction on the first reclamation area ROI image and the second reclamation area ROI image respectively to obtain a first reclamation area ROI visual feature description feature map and a second reclamation area ROI visual feature description feature map. S140, performing local neighborhood-aware feature enhancement on the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map respectively to obtain a first enhanced reclamation area ROI visual feature description feature map and a second enhanced reclamation area ROI visual feature description feature map; S150, calculating a reclamation area ROI visual difference feature description feature map between the first enhanced reclamation area ROI visual feature description feature map and the second enhanced reclamation area ROI visual feature description feature map; S160, obtaining a settlement monitoring result based on the reclamation area ROI visual difference feature description feature map.

[0015] In step S110, the first reclamation area image and the second reclamation area image at the first time point and the second time point collected by the camera are obtained. It should be understood that the first reclamation area image mainly contains the surface texture information (such as soil cracks, vegetation coverage, building form, etc.) of the monitoring area at the first time point, the surface marker information (such as roads, bridges, boundary signs, etc.), and the visual performance of the surface height (such as shadows, height differences caused by changes in illumination) information, and the second reclamation area image mainly contains the surface texture information, surface marker information, and visual performance information of the surface height of the monitoring area at the second time point. The image at a single time point can only provide static information of the reclamation area at a certain moment, and here, by obtaining the reclamation area images at two time points and comparing and analyzing the relevant contents in the two reclamation area images, it can be known whether the ground has subsided and how it has changed. In general, by comprehensively considering the changing characteristics of the objects in the reclamation area at different time points, the changes in the topography and landforms, etc., a high-precision analysis of the ground subsidence height of the reclamation area can be achieved.

[0016] In step S120, dynamic ROI generation is performed on the first reclamation area image and the second reclamation area image to obtain the first reclamation area ROI image and the second reclamation area ROI image. Specifically, in an embodiment of the present application, the step S120 includes: inputting the first reclamation area image and the second reclamation area image into a dynamic ROI generator based on a U-Net network to obtain the first reclamation area ROI image and the second reclamation area ROI image. It should be understood that the reclamation area image usually contains a large amount of background information, such as large areas without settlement changes. That is, the ground settlement in the reclamation area is not evenly distributed, and some specific areas may be more susceptible to or more sensitive to settlement. Therefore, in order to focus the analysis on key areas where ground settlement may occur, reduce the processing of irrelevant information, and thus achieve higher resolution settlement monitoring, the present application obtains the first reclamation area ROI image and the second reclamation area ROI image by inputting the first reclamation area image and the second reclamation area image into a dynamic ROI generator based on a U-Net network. Those skilled in the art should know that the U-Net network is a convolutional neural network that performs well in the field of image segmentation. It has a unique encoder-decoder structure, where the encoder is responsible for downsampling to extract image features, and the decoder restores the low-resolution feature map to the original image size by upsampling, while retaining rich contextual information through jump connections during the encoding and decoding process. This structure enables the U-Net network to achieve good results in image segmentation tasks with complex details and irregular shapes. For reclamation area images, there are also complex topographic details, and the U-Net network can effectively extract relevant features, providing a basis for accurate ROI division. And "dynamic" means that the generator can adaptively determine the region of interest based on the image content. The ground subsidence in the reclamation area may have different manifestations in different areas, and the subsidence area at different time points may also change. The use of a dynamic ROI generator can accurately delineate the key areas related to ground subsidence based on the actual situation of each reclamation area image. This flexibility greatly improves the pertinence and accuracy of subsequent analysis.

[0017] The following is a detailed description of a specific implementation process of "inputting the first reclamation area image and the second reclamation area image into the dynamic ROI generator based on the U-Net network to obtain the first reclamation area ROI image and the second reclamation area ROI image": Preliminary preparation is the basis of the entire implementation process. In the data collection and collation stage, it is necessary to widely collect image data of the reclamation area at different times. The high-definition camera carried by the drone can overlook from the air and obtain the overall image information of the reclamation area. The collected images must cover various features and scenes of the reclamation area, such as different topography, building distribution, vegetation coverage, etc., so that the subsequent model can learn comprehensive and rich information. Data preprocessing is to make the image data more suitable for model input and processing. Image standardization is an important step. By unifying the pixel values ​​to a specific range, such as [0,1] or [-1,1], the model training speed can be accelerated to avoid the problem of training instability caused by large differences in pixel value ranges. Resizing is also essential. Resizing all images to the same size can meet the requirements of the U-Net network input layer and ensure that the network can process these images normally. In addition, data augmentation operations can increase the diversity of data. Through methods such as rotation, flipping, and brightness adjustment, the model can be exposed to more images of different forms during training, thereby improving the generalization ability of the model and enabling it to perform better when facing various complex situations in practical applications.

[0018] Building a dynamic ROI generator based on the U-Net network is the core work of the implementation process. As an advanced convolutional neural network, the unique encoder-decoder structure of the U-Net network plays a key role. The encoder downsamples the input image through convolution and pooling operations. In this process, the spatial size of the image gradually decreases, but the abstract level of the features continues to increase, and high-level semantic features can be extracted, such as the overall layout characteristics of the building group in the reclamation area and the texture characteristics of a large area of ​​land. The decoder uses deconvolution and jump connections to upsample, merge the high-level semantic features extracted by the encoder with the low-level detail features, restore the spatial information of the image, and enable the final output image to retain more details of the original image. A dynamic ROI generation mechanism is introduced based on the U-Net network. Specifically, a binary classification convolution layer is set at the output layer of the decoder. This convolution layer can predict pixel by pixel whether the pixel belongs to the region of interest (ROI) based on the characteristics of the input image. This dynamic generation mechanism enables the network to flexibly and intelligently determine the ROI according to the actual content of different reclamation area images, greatly improving the accuracy and adaptability of ROI determination.

[0019] After the model is built, it needs to be trained. First, a labeled data set should be created. Some representative images are selected from the collected reclamation area images. Professionals manually mark the ROIs in them and generate corresponding labeled images. In the labeled images, the pixels in the ROI part are marked as 1 and the non-ROI part is marked as 0. These labeled images become the standard reference for model training and provide a clear goal for model learning. Determining the appropriate loss function and optimizer is crucial to the model training effect. The binary cross entropy loss function is usually used, which can accurately measure the difference between the ROI mask predicted by the network and the true annotation. By continuously calculating this difference, the model is guided to learn how to identify the ROI more accurately. The choice of optimizer will also have a significant impact on model training. Common optimizers such as stochastic gradient descent (SGD) and Adam are used to continuously adjust the parameters of the network to continuously reduce the value of the loss function, thereby improving the performance of the model. During the training process, the preprocessed image and the corresponding labeled image are input into the U-Net network, and the model will perform multiple iterative calculations. Generally, multiple training cycles are required. As the training continues, the model's ability to recognize ROI will gradually improve until the model's performance reaches a stable state. At this time, the model can achieve high accuracy and stability in identifying ROI.

[0020] When the model training is completed, it enters the stage of generating ROI images. The preprocessed first reclamation area image and the second reclamation area image are respectively input into the trained dynamic ROI generator, and the network will perform forward propagation calculations. In this process, the network outputs the corresponding binary mask based on the characteristics of the input image. This binary mask has the same size as the original image, where pixels with a value of 1 represent ROI, and pixels with a value of 0 represent non-ROI. Then, by multiplying the original image by the generated binary mask pixel by pixel, the first reclamation area ROI image and the second reclamation area ROI image can be obtained. These two ROI images only retain the areas in the reclamation area related to ground subsidence monitoring, removing a large amount of irrelevant background information, providing more targeted and streamlined data for subsequent feature extraction and analysis, which helps to improve the efficiency and accuracy of subsequent monitoring and analysis work.

[0021] Model evaluation and optimization are important links to ensure the continuous improvement of the performance of the entire model. Use a dedicated test data set to evaluate the trained model. Common evaluation indicators include accuracy, recall, F1 value, etc. Accuracy reflects the proportion of correct predictions by the model, recall measures the proportion of ROIs that the model can correctly identify, and F1 value comprehensively considers accuracy and recall to more comprehensively evaluate the performance of the model. Through these indicators, we can understand in detail the advantages and disadvantages of the model in identifying ROIs. If the evaluation results of the model are not ideal, the model needs to be optimized. Optimization measures can be taken from multiple aspects, such as adjusting the network structure, trying to increase or decrease the number of network layers, changing the number of neurons, etc., to explore a network architecture that is more suitable for the current task; you can also increase training data and collect more reclamation area images of different types and scenes to allow the model to learn more diverse features; in addition, optimize hyperparameters, such as learning rate, batch size, etc., and these subtle adjustments to hyperparameters may have a significant impact on the training speed and convergence effect of the model. Through continuous model evaluation and optimization, the performance and generalization ability of the model will continue to improve, which will help to more accurately generate ROI images of the reclamation area, and thus provide more reliable and accurate data support for ground subsidence monitoring in the reclamation area.

[0022] In step S130, visual multi-scale feature extraction is performed on the first reclamation area ROI image and the second reclamation area ROI image respectively to obtain a first reclamation area ROI visual feature description feature map and a second reclamation area ROI visual feature description feature map. Specifically, in an embodiment of the present application, the step S130 includes: inputting the first reclamation area ROI image and the second reclamation area ROI image into a visual multi-scale feature extractor based on a dynamic convolutional labeling model respectively to obtain the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map. Accordingly, it is considered that ground subsidence may present different characteristics at different scales. For example, it may appear as an overall terrain undulation in the region on a large scale, and may be reflected in subtle structures such as local ground cracks on a small scale. In order to consider these different scales of information at the same time and provide a more comprehensive understanding, in the technical solution of the present application, a visual multi-scale feature extractor based on a dynamic convolutional labeling model is used to process the first reclamation area ROI image and the second reclamation area ROI image respectively, so as to comprehensively capture various features in the reclamation area ROI image from macro to micro, and obtain the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map. In particular, it should be understood that the dynamic convolutional labeling model has unique advantages in describing the characteristics of the reclamation area. It can comprehensively capture different scale features from local details to overall structure. Specifically, the dynamic convolutional labeling model divides the input image into multiple local blocks, and uses a convolutional neural network (CNN) to process each local block, thereby extracting fine local features, such as soil texture changes in a specific area, small-scale edge deformation, etc., which is the strength of CNN. Subsequently, the feature maps of these local blocks are spliced ​​together and globally processed through the Transformer model. With its powerful self-attention mechanism, Transformer can capture the long-distance dependencies between local blocks, thereby well capturing the long-distance dependencies between the parts of the local image of the reclamation area and learning richer global semantic information. Unlike traditional static models, the "dynamic" characteristics of the dynamic convolutional labeling model enable it to automatically adjust internal operations according to the characteristics of the reclamation area image data, such as changing the attention weight distribution or choosing to analyze certain key local blocks more carefully, improving the adaptability and processing efficiency of reclamation scenes of different complexity.

[0023] The following is a detailed description of a specific implementation process of "inputting the first reclamation area ROI image and the second reclamation area ROI image into the visual multi-scale feature extractor based on the dynamic convolution labeling model to obtain the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map": Since the collected reclamation area ROI image may have problems such as uneven illumination, noise interference and inconsistent resolution, if it is not processed, it will seriously affect the accuracy and efficiency of subsequent feature extraction. Therefore, before the specific feature extraction, the input ROI image needs to be strictly preprocessed. First, the reclamation area ROI image is corrected for illumination, so as to use histogram equalization or adaptive histogram equalization technology to adjust the brightness and contrast of the image to ensure that the details of each area in the image can be clearly presented and avoid certain features being covered up due to illumination differences. As for the noise in the reclamation area ROI image, algorithms such as median filtering and Gaussian filtering can be used to remove it to make the image smoother and provide a clean data basis for subsequent feature extraction. At the same time, in order to enable the model to process image data uniformly, it is also necessary to normalize the size of ROI images of different resolutions and adjust them to the standard size required by the model input.

[0024] Building a visual multi-scale feature extractor based on the dynamic convolutional labeling model is the core step of the implementation. The dynamic convolutional labeling model combines the advantages of convolutional neural networks (CNNs) and Transformer models. Inside the model, the input reclamation area ROI image is first divided into multiple local blocks of appropriate size. This division is not arbitrary, but is determined based on the common size of the objects in the reclamation area and experience to ensure that each local block can contain valuable information. For example, for buildings in the reclamation area, each local block may contain part of the building structure, such as corners, doors and windows, etc.; for roads, it contains a representative section of road surface texture and boundary information. After the division is completed, CNN is used to process each local block. The convolution layer of CNN performs convolution operations on local blocks by designing convolution kernels of different sizes, such as 3×3, 5×5, etc. Smaller convolution kernels are good at capturing subtle texture features in local blocks, such as small cracks on the soil surface and texture details of vegetation; larger convolution kernels can obtain more macroscopic structural features, such as the overall outline of buildings and the boundary shape of large water areas. After multiple layers of convolution and pooling operations, each local block is converted into a feature map containing rich local features. Although these feature maps describe the local information in detail, they lack the expression of long-distance dependencies between local blocks.

[0025] Therefore, the model uses the Transformer model to make up for this deficiency. The Transformer model can calculate the attention weights between the feature maps of each local block. In this way, the model can capture the long-distance dependencies between different local blocks and learn richer global semantic information. For example, in the reclamation area, the spatial relationship between buildings and surrounding roads and waters, as well as the relationship between vegetation cover changes in different regions and land use patterns. The acquisition of this global information helps the model to fully understand the overall characteristics and changing trends of the reclamation area. The "dynamic" nature of the dynamic convolutional labeling model is also reflected in its ability to automatically adjust internal operations according to the characteristics of the input image data. When encountering complex reclamation area scenes, such as a large number of buildings and complex terrain, the model will automatically increase the analysis weight of key local blocks (such as densely built areas) and extract the features of these areas more carefully; in relatively simple areas, such as large tracts of empty farmland, the model will appropriately reduce the allocation of computing resources to improve processing efficiency.

[0026] In the feature extraction process, the first reclamation area ROI image and the second reclamation area ROI image are respectively input into the constructed visual multi-scale feature extractor. For each image, after the collaborative processing of the CNN and Transformer models, a feature map that integrates local detail features and global semantic information will be obtained, namely the feature map describing the visual features of the first reclamation area ROI and the feature map describing the visual features of the second reclamation area ROI. These feature maps contain rich multi-scale information, from microscopic texture changes of objects to macroscopic changes in topography and landforms, which are effectively encoded in them. For example, the feature map may contain both the texture features of tiny cracks on the ground (small-scale features) and the trend features of the overall terrain undulations of the reclamation area (large-scale features).

[0027] In order to ensure the accuracy and effectiveness of feature extraction, the extraction process also needs to be optimized and verified. On the one hand, by setting reasonable model parameters, such as the number of convolutional layers and the number of attention heads in the Transformer model, the performance of the model is continuously adjusted. The selection of these parameters requires repeated debugging and optimization based on a large amount of experimental data and actual monitoring results. On the other hand, the cross-validation method is used to divide different reclamation area ROI image data sets into training sets, validation sets, and test sets. During the training process, the model is trained using the training set, and then the performance of the model is evaluated on the validation set, and the model parameters are adjusted according to the evaluation results. Finally, the optimized model is finally tested using the test set to ensure that the model can stably and accurately extract multi-scale features on different reclamation area image data.

[0028] In step S140, the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map are respectively enhanced by local neighborhood perception to obtain a first enhanced reclamation area ROI visual feature description feature map and a second enhanced reclamation area ROI visual feature description feature map. In particular, the feature quasi-enhancement process of the first reclamation area ROI visual feature description feature map is taken as an example for explanation. Specifically, Figure 3 FIG. 1 is a flow chart of step S140 in the method for high-precision investigation and monitoring of land subsidence in a reclamation area according to an embodiment of the present application. Figure 3 As shown, the step S140 includes: S141, performing feature decoupling and feature flattening on the first reclamation area ROI visual feature description feature map to obtain a set of first reclamation area ROI visual local feature vectors; S142, performing visual importance measurement on each first reclamation area ROI visual local feature vector in the set of the first reclamation area ROI visual local feature vectors to obtain a set of first reclamation area ROI visual local feature importance score values; S143, based on the set of the first reclamation area ROI visual local feature importance score values, Arrange the set of the first reclamation area ROI visual local feature vectors in descending order to obtain a descending sequence of the first reclamation area ROI visual local feature vectors; S144, perform local neighborhood activity calculation and feature selection on the descending sequence of the first reclamation area ROI visual local feature vectors to obtain a descending sequence of the selected first reclamation area ROI visual local feature vectors; S145, reshape the descending sequence of the selected first reclamation area ROI visual local feature vectors to obtain the first enhanced reclamation area ROI visual feature description feature map.

[0029] It should be understood that the characteristics presented by ground subsidence in the reclamation area often contain key information in the local neighborhood. For example, ground subsidence may first produce fine cracks in the local area. The changes in the texture, shape and other characteristics of these cracks in the local neighborhood are crucial to judging the subsidence situation, and there are also mutual correlations between these local areas. Therefore, in order to highlight these key local features so that subsequent analysis can more keenly capture information related to subsidence, the present application introduces a local neighborhood-aware feature enhancement method to enhance the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map based on the activity level in the neighborhood to obtain the first enhanced reclamation area ROI visual feature description feature map and the second enhanced reclamation area ROI visual feature description feature map. In particular, this method can identify key features closely related to ground subsidence based on the activity level of the features in its neighborhood. For example, ground subsidence may cause changes in texture, color, or shape in local areas, which manifest as changes in feature activity in their neighborhoods. In this way, attention can be focused on these key features to accurately reflect the actual situation of ground subsidence in the reclamation area.

[0030] Specifically, in the embodiment of the present application, the step S141 includes: performing feature decoupling on the first reclamation area ROI visual feature description feature map along the channel dimension of the first reclamation area ROI visual feature description feature map to obtain a set of first reclamation area ROI visual feature encoding local feature matrices. The process can be expressed as: ;in, is the visual feature description feature map of the first reclamation area ROI, is a feature decoupling operation, and It is the first, second, and third in the set of local feature matrices of the visual feature encoding of the first reclamation area ROI. and The first reclamation area ROI visual feature encodes the local feature matrix.

[0031] Each first reclamation area ROI visual feature encoding local feature matrix in the set of the first reclamation area ROI visual feature encoding local feature matrix is ​​feature flattened to obtain a set of the first reclamation area ROI visual local feature vectors. This process can be expressed as: ;in, and It is the first, second, and third in the set of local feature matrices of the visual feature encoding of the first reclamation area ROI. and The first reclamation area ROI visual feature encoding local feature matrix, is the feature flattening operation, and They are the first, second, and third local feature vectors of the first reclamation area ROI. and The first reclamation area ROI visual local feature vector.

[0032] It should be understood that the first reclamation area ROI visual feature description feature map is a comprehensive expression of the visual features of a specific area of ​​the reclamation area, which contains a lot of complex and intertwined information. The features involved in the land subsidence in the reclamation area are diverse, such as the deformation of buildings, changes in surface texture, etc. These features are mixed together in the original feature map and are difficult to distinguish and use directly. Through feature decoupling, this complex comprehensive feature can be decomposed into a more targeted and independent set of local feature matrices of the first reclamation area ROI visual feature encoding, so as to facilitate the subsequent in-depth analysis of each local feature. Then, through the feature flattening operation, each local feature matrix of the first reclamation area ROI visual feature encoding is transformed into a one-dimensional first reclamation area ROI visual local feature vector to obtain a set of first reclamation area ROI visual local feature vectors. In the local feature vector obtained after flattening, each element represents a different but related feature dimension, which can provide a rich data source for subsequent feature selection.

[0033] Specifically, in the embodiment of the present application, the step S142 includes: performing a linear transformation on each first reclamation area ROI visual local feature vector in the set of the first reclamation area ROI visual local feature vectors to obtain a set of first reclamation area ROI visual local feature linear transformation vectors; multiplying the weight score vector with each first reclamation area ROI visual local feature linear transformation vector in the set of the first reclamation area ROI visual local feature linear transformation vector to obtain a set of the first reclamation area ROI visual local feature importance score values. The above process can be expressed as follows: ;in, It is the first in the set of visual local feature vectors of the first reclamation area ROI The first reclamation area ROI visual local feature vector, and They are The corresponding weight matrix and bias vector, is matrix multiplication, is the weight score vector, is the set of importance scores of the local visual features of the first reclamation area ROI, and They are the first, second, and third in the set of visual local feature importance scores of the first reclamation area ROI. and The importance score of the visual local features of the first reclamation area ROI.

[0034] It should be understood that the set of visual local feature vectors of the first reclamation area ROI contains many feature vectors, and the information carried by each vector contributes to the judgment of ground subsidence to a different extent. By using the visual importance measurement and a specific algorithm to calculate the importance score for each visual local feature vector of the first reclamation area ROI, the contribution of these features to the overall task (analyzing the degree of ground subsidence) can be quantified, so that the model can clearly understand the value of each feature. In other words, the obtained set of visual local feature importance score values ​​of the first reclamation area ROI clearly presents the contribution of each local feature vector to the overall task in a quantitative way. For example, a feature vector with a high score indicates that the information it carries plays a key role in judging the ground subsidence. For example, a feature vector with a high score for the foundation of a building indicates that the change in the foundation of the building has a significant impact on the judgment of ground subsidence; a feature vector with a low score indicates that its contribution to the task is relatively small. This quantitative representation can provide a clear basis for subsequent feature selection.

[0035] Then, based on the set of the first reclamation area ROI visual local feature importance score values, the set of the first reclamation area ROI visual local feature vectors is arranged in descending order to obtain a descending sequence of the first reclamation area ROI visual local feature vectors. The above process can be expressed as: ;in, Based on The set of the visual local feature vectors of the first reclamation area ROI is arranged in descending order, and They are the first, second, and third in the descending sequence of the visual local feature vectors of the first reclamation area ROI. and The first reclamation area ROI visual local feature vector.

[0036] It should be understood that in the task of monitoring ground subsidence in the reclamation area, the set of visual local feature vectors of the first reclamation area ROI contains many feature vectors, and their importance to judging the ground subsidence situation varies. By arranging the set of visual local feature vectors of the first reclamation area ROI in descending order based on the set of visual local feature importance score values ​​of the first reclamation area ROI, it is possible to quickly determine which features are most critical to the task objectives. For example, feature vectors directly related to building cracks, road deformations, etc. caused by ground subsidence have higher importance scores and will be at the front of the sequence after descending order, so that the model can focus on these key features and deeply analyze their relationship with ground subsidence.

[0037] Specifically, in the embodiment of the present application, the step S144 includes: based on the neighborhood features of each first reclamation area ROI visual local feature vector in the descending sequence of the first reclamation area ROI visual local feature vector, calculating the feature neighborhood activity of each first reclamation area ROI visual local feature vector to obtain a sequence of first reclamation area ROI visual local feature neighborhood activity, and the process can be expressed by the formula: ;in, yes Middle The eigenvalues ​​at the positions, yes The number of eigenvalues ​​in , and They are and The corresponding weighted average of the visual features of the first reclamation area ROI, yes The corresponding first reclamation area ROI visual local feature neighborhood activity.

[0038] Based on the sequence of the neighborhood activity of the first reclamation area ROI visual local feature, feature selection is performed on the descending sequence of the first reclamation area ROI visual local feature vectors to obtain the descending sequence of the selected first reclamation area ROI visual local feature vectors. This process can be expressed by the formula: ;in, is the first reclamation area ROI visual local feature vector in the descending sequence The first reclamation area ROI visual local feature vector, For Perform feature selection, is the preset threshold, and They are the first, second, and third in the descending sequence of the visual local feature vector of the first reclamation area ROI after selection. and The visual local feature vector of the first reclamation area ROI after selection.

[0039] It should be understood that ground subsidence in the reclamation area is a complex process, and its related features do not exist in isolation. The information represented by the visual local feature vector of the first reclamation area ROI, such as the deformation characteristics of a certain part of the building, often interacts with other surrounding features, such as the texture changes of the surrounding ground and the status of adjacent buildings. Calculating the degree of interaction between the visual local feature vectors of each first reclamation area ROI and the surrounding features and obtaining the feature neighborhood activity index can provide a more comprehensive understanding of the ground subsidence phenomenon. For example, ground subsidence causes the building to tilt, which may also cause cracks in the surrounding ground. There is an interactive relationship between the crack feature and the building tilt feature, and this relationship can be captured through the visual local feature neighborhood activity of the first reclamation area ROI.

[0040] Accordingly, considering that the original first reclamation area ROI visual local feature vector descending sequence has been sorted according to the importance score, there may be some redundant features or features that have little effect on reflecting the ground subsidence pattern. By combining the neighborhood activity of the first reclamation area ROI visual local feature for feature selection, those features that perform poorly in the neighborhood relationship and contribute little to the overall data pattern can be removed, making the feature set more compact. For example, although some features have a high importance score, they have a low correlation with the surrounding features and are of limited help in capturing the overall pattern of ground subsidence. They can be eliminated by screening based on neighborhood activity, which is conducive to optimizing the feature set.

[0041] Finally, the descending sequence of the selected first reclamation area ROI visual local feature vectors is reshaped to obtain the first enhanced reclamation area ROI visual feature description feature map. The above process can be expressed as: ;in, and They are the first, second, and third in the descending sequence of the visual local feature vector of the first reclamation area ROI after selection. and The visual local feature vector of the first reclamation area ROI after selection, For the reshape operation, It is the first enhanced reclamation area ROI visual feature description feature map.

[0042] It should be understood that in the feature selection process, in order to highlight the key features, the spatial relationship of the original features is broken. By reshaping the feature shape, the selected first reclamation area ROI visual local feature vector is recombined into the first enhanced reclamation area ROI visual feature description feature map, which helps to restore the spatial layout and mutual relationship of these features in the original reclamation area image. For example, in the reclamation area image, the features of different objects such as buildings and roads are spatially correlated, and ground subsidence may cause coordinated changes in the features of these objects. The reshaped feature map can re-cluster these related features in space, so that the model can learn the spatial dependencies between these features, such as the relationship between building settlement and nearby road deformation, so as to have a more comprehensive understanding of the ground settlement in the monitoring area.

[0043] In step S150, a reclamation area ROI visual difference feature description feature map between the first enhanced reclamation area ROI visual feature description feature map and the second enhanced reclamation area ROI visual feature description feature map is calculated. Accordingly, considering that the first enhanced reclamation area ROI visual feature description feature map and the second enhanced reclamation area ROI visual feature description feature map represent the enhanced features of the reclamation area at two different time points, respectively, and ground subsidence, as a dynamic change process, will cause the surface features of the reclamation area, such as topography, texture, and distribution of objects, to change. Therefore, in order to quantify the changes in the reclamation area at two time points, especially the degree of ground subsidence, the present application calculates the reclamation area ROI visual difference feature description feature map between the first enhanced reclamation area ROI visual feature description feature map and the second enhanced reclamation area ROI visual feature description feature map. In particular, here it is obtained by calculating the position difference of each position of the two features. In this way, it is possible to accurately capture which areas have undergone significant changes and the specific nature of the changes (for example, subsidence, displacement, etc.), so as to intuitively and accurately display the changes in the visual features of the reclamation area between the two time points.

[0044] In step S160, based on the feature map of the visual difference feature description of the reclamation area ROI, a settlement monitoring result is obtained. Specifically, in an embodiment of the present application, the step S160 includes: inputting the feature map of the visual difference feature description of the reclamation area ROI into a settlement monitoring module based on a decoder to obtain the settlement monitoring result, and the settlement monitoring result includes a decoded value of the settlement height. That is, the feature map of the visual difference feature description of the reclamation area ROI obtained by positional difference between the first enhanced feature map of the visual feature description of the reclamation area ROI and the second enhanced feature map of the visual feature description of the reclamation area ROI is decoded and processed, so as to convert these complex image features into actual physical quantities that are easy to understand and apply, such as the decoded value of the settlement height, by the settlement monitoring module based on the decoder. Specifically, the decoder decodes the information in the feature map through a specific algorithm and model structure, maps it to the actual parameter space related to settlement, and realizes the conversion from image features to actual monitoring data. In particular, settlement height is one of the key indicators for measuring the degree of ground settlement. Accurately obtaining this value helps to determine whether the ground settlement in the reclamation area exceeds the safety range and provides clear data support for subsequent management decisions. For example, when the settlement height exceeds a certain threshold, the relevant departments can take timely measures, such as strengthening the foundation, adjusting the water resource exploitation plan, etc., to prevent adverse consequences such as building damage and underground pipeline rupture caused by ground settlement. In a specific embodiment of the present application, the reclamation area ROI visual difference feature description feature map is input into a decoder-based settlement monitoring module to obtain the settlement monitoring result, and the settlement monitoring result includes a decoded value of the settlement height, including: multiplying the decoding weight matrix of the decoder with each feature matrix of the reclamation area ROI visual difference feature description feature map along the channel dimension to obtain the decoded reclamation area ROI visual difference feature description feature map, and accumulating and summing all the feature values ​​of the decoded reclamation area ROI visual difference feature description feature map to obtain the decoded value of the settlement height.

[0045] Here, when the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map respectively represent the dynamic ROI multi-scale image semantic features of the first reclamation area image and the second reclamation area image, when calculating the position-by-position difference after local neighborhood-aware feature enhancement, the insufficient correspondence of the neighborhood activity association of different image semantic features will cause the sparse selective differential association of the reclamation area ROI visual difference feature description feature map, thereby reducing the accuracy of the settlement monitoring results obtained by inputting the decoder-based settlement monitoring module due to the lack of decoding inference degree.

[0046] In a preferred example, inputting the reclamation area ROI visual difference feature description feature map into a decoder-based settlement monitoring module to obtain a settlement monitoring result includes: expanding the reclamation area ROI visual difference feature description feature map into a reclamation area ROI visual difference feature description feature vector, and calculating the first Eigenvalues ​​and The eigenvalues Distance and Distance, thus obtaining the first reclamation area ROI visual difference feature description distance matrix and the second reclamation area ROI visual difference feature description distance matrix: Calculate the weighted sum of the first reclamation area ROI visual difference feature description distance matrix and the second reclamation area ROI visual difference feature description distance matrix, and determine the eigenvalues ​​of the distance weighted sum matrix arrive , so as to arrange the eigenvalues ​​to obtain the reclamation area ROI visual difference feature description distance eigenvector : ; Describe the distance eigenvector of the visual difference feature of the reclamation area ROI Interpolation is performed to obtain an interpolated eigenvector of the visual difference feature description of the reclamation area ROI having the same length as the eigenvector of the visual difference feature description of the reclamation area ROI.

[0047] The reclamation area ROI visual difference feature description feature vector is matrix-multiplied with the first reclamation area ROI visual difference feature description distance matrix to obtain the reclamation area ROI visual difference feature description intermediate feature vector, and the second reclamation area ROI visual difference feature description distance matrix is ​​matrix-multiplied with the autocorrelation matrix of the reclamation area ROI visual difference feature description feature vector to obtain the reclamation area ROI visual difference feature description intermediate feature matrix.

[0048] After matrix multiplication of the intermediate feature vector of the reclamation area ROI visual difference feature description and the intermediate feature matrix of the reclamation area ROI visual difference feature description, a dot multiplication is further performed with the interpolation eigenvector of the reclamation area ROI visual difference feature description to obtain an optimized reclamation area ROI visual difference feature description coding vector.

[0049] The optimized reclamation area ROI visual difference feature description feature vector is input into the decoder-based settlement monitoring module to obtain the settlement monitoring result.

[0050] That is, the first distance matrix and the second distance matrix of the visual difference feature description feature vector of the reclamation area ROI after the visual difference feature description feature map of the reclamation area ROI are expanded are used as the fine-grained metric association cluster representation of the visual difference feature description feature vector of the reclamation area ROI, and the dynamic programming of the relationship between the association clusters of different association clusters is performed on the visual difference feature description feature vector of the reclamation area ROI and the self-association representation of the visual difference feature description feature vector of the reclamation area ROI respectively to simulate the sparse activation based on neuron clusters of the association system, and the fine-grained predictable sparsity of the visual difference feature description feature vector of the reclamation area ROI is coordinated with the intrinsic representation of the metric association cluster of the first distance matrix and the second distance matrix of the visual difference feature description feature vector of the reclamation area ROI, so as to avoid the lack of decoding inference degree affected by insufficient association caused by sparsity, and improve the accuracy of the settlement monitoring result obtained by inputting the visual difference feature description feature map of the reclamation area ROI into the settlement monitoring module based on the decoder.

[0051] In summary, the high-precision survey and monitoring method of land subsidence in the reclamation area based on the embodiment of the present application is explained, which uses the image recognition and analysis algorithm based on machine vision to generate ROIs and extract visual multi-scale features of the first reclamation area image and the second reclamation area image, and then performs local neighborhood perception enhancement on the extracted first reclamation area ROI visual feature description features and the second reclamation area ROI visual feature description features, so as to intelligently obtain the settlement monitoring results of the settlement height based on the visual difference feature representation between the enhanced first enhanced reclamation area ROI visual feature description features and the second enhanced reclamation area ROI visual feature description features. In this way, the accuracy of the calculation of the land subsidence height in the reclamation area can be effectively improved.

Claims

1. A high-precision survey and monitoring method for land subsidence in reclamation areas, characterized in that: include: Acquire a first reclamation area image and a second reclamation area image at a first time point and a second time point captured by a camera; Dynamic ROI generation is performed on the first reclamation area image and the second reclamation area image to obtain the first reclamation area ROI image and the second reclamation area ROI image; visual multi-scale feature extraction is performed on the first reclamation area ROI image and the second reclamation area ROI image respectively to obtain the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map; local neighborhood perception feature enhancement is performed on the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map respectively to obtain the first enhanced reclamation area ROI visual feature description feature map and the second enhanced reclamation area ROI visual feature description feature map; the reclamation area ROI visual difference feature description feature map between the first enhanced reclamation area ROI visual feature description feature map and the second enhanced reclamation area ROI visual feature description feature map is calculated; and the settlement monitoring result is obtained based on the reclamation area ROI visual difference feature description feature map.

2. The high-precision survey and monitoring method for land subsidence in reclamation areas according to claim 1 is characterized in that: Dynamic ROI generation is performed on the first reclamation area image and the second reclamation area image to obtain the first reclamation area ROI image and the second reclamation area ROI image, including: inputting the first reclamation area image and the second reclamation area image into a dynamic ROI generator based on a U-Net network to obtain the first reclamation area ROI image and the second reclamation area ROI image.

3. The high-precision survey and monitoring method for land subsidence in reclamation areas according to claim 2 is characterized in that: Performing visual multi-scale feature extraction on the first reclamation area ROI image and the second reclamation area ROI image respectively to obtain a first reclamation area ROI visual feature description feature map and a second reclamation area ROI visual feature description feature map, including: inputting the first reclamation area ROI image and the second reclamation area ROI image into a visual multi-scale feature extractor based on a dynamic convolutional labeling model respectively to obtain the first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map.

4. The high-precision survey and monitoring method for land subsidence in reclamation areas according to claim 3 is characterized in that: The first reclamation area ROI visual feature description feature map and the second reclamation area ROI visual feature description feature map are respectively subjected to local neighborhood perception feature enhancement to obtain a first enhanced reclamation area ROI visual feature description feature map and a second enhanced reclamation area ROI visual feature description feature map, including: performing feature decoupling and feature flattening on the first reclamation area ROI visual feature description feature map to obtain a set of first reclamation area ROI visual local feature vectors; performing visual importance measurement on each first reclamation area ROI visual local feature vector in the set of the first reclamation area ROI visual local feature vectors to obtain the first reclamation area ROI visual local feature vector. The method comprises the steps of: calculating a set of visual local feature importance score values ​​of the first reclamation area ROI; arranging the set of visual local feature vectors of the first reclamation area ROI in descending order based on the set of visual local feature importance score values ​​of the first reclamation area ROI to obtain a descending sequence of visual local feature vectors of the first reclamation area ROI; performing local neighborhood activity calculation and feature selection on the descending sequence of visual local feature vectors of the first reclamation area ROI to obtain a descending sequence of visual local feature vectors of the first reclamation area ROI after selection; and reshaping the descending sequence of visual local feature vectors of the first reclamation area ROI after selection to obtain the first enhanced reclamation area ROI visual feature description feature map.

5. The high-precision survey and monitoring method for land subsidence in reclamation areas according to claim 4 is characterized in that: The first reclamation area ROI visual feature description feature map is feature decoupled and feature flattened to obtain a set of first reclamation area ROI visual local feature vectors, including: feature decoupling the first reclamation area ROI visual feature description feature map along the channel dimension of the first reclamation area ROI visual feature description feature map to obtain a set of first reclamation area ROI visual feature encoding local feature matrices; feature flattening each first reclamation area ROI visual feature encoding local feature matrix in the set of the first reclamation area ROI visual feature encoding local feature matrix to obtain a set of first reclamation area ROI visual local feature vectors.

6. The high-precision survey and monitoring method for land subsidence in reclamation areas according to claim 5 is characterized in that: Performing visual importance measurement on each first reclamation area ROI visual local feature vector in the set of the first reclamation area ROI visual local feature vectors to obtain a set of first reclamation area ROI visual local feature importance score values, including: performing linear transformation on each first reclamation area ROI visual local feature vector in the set of the first reclamation area ROI visual local feature vectors to obtain a set of first reclamation area ROI visual local feature linear transformation vectors; multiplying the weight score vector with each first reclamation area ROI visual local feature linear transformation vector in the set of the first reclamation area ROI visual local feature linear transformation vector to obtain a set of first reclamation area ROI visual local feature importance score values.

7. The high-precision survey and monitoring method for land subsidence in reclamation areas according to claim 6 is characterized in that: The method comprises: performing local neighborhood activity calculation and feature selection on the descending sequence of the first reclamation area ROI visual local feature vectors to obtain a descending sequence of the first reclamation area ROI visual local feature vectors after selection, comprising: calculating the feature neighborhood activity of each first reclamation area ROI visual local feature vector based on the neighborhood features of each first reclamation area ROI visual local feature vector in the descending sequence of the first reclamation area ROI visual local feature vectors to obtain a sequence of first reclamation area ROI visual local feature neighborhood activities; and performing feature selection on the descending sequence of the first reclamation area ROI visual local feature vectors based on the sequence of the first reclamation area ROI visual local feature neighborhood activities to obtain a descending sequence of the first reclamation area ROI visual local feature vectors after selection.

8. The high-precision survey and monitoring method for land subsidence in reclamation areas according to claim 7 is characterized in that: Based on the reclamation area ROI visual difference feature description feature map, a settlement monitoring result is obtained, including: inputting the reclamation area ROI visual difference feature description feature map into a decoder-based settlement monitoring module to obtain the settlement monitoring result, and the settlement monitoring result includes a decoded value of the settlement height.

Images