River and lake shoreline monthly monitoring remote sensing image automatic interpretation and achievement management system

Abstract

A river and lake shoreline monthly monitoring remote sensing image automatic interpretation and result management system comprises a data resource management and access module, a remote sensing image intelligent preprocessing module, an intelligent interpretation and change identification module, a man-machine collaborative interaction and correction module and a database and report generation module. The data resource management and access module is used for accessing image data, the remote sensing image intelligent preprocessing module is used for data preprocessing, the intelligent interpretation and change recognition module is used for element automatic extraction and change discovery, and the man-machine collaborative interaction and correction module is used for man-machine interaction and result correction. And the database and report generation module is used for data storage and monitoring result generation. The invention provides an improved self-adaptive gating network sub-pixel shoreline interpretation algorithm for automatic interpretation of remote sensing images, and provides an improved cross-temporal attention and uncertainty collaborative change detection algorithm for monthly dynamic analysis of river and lake shorelines. A better scheme is provided for a river and lake shoreline monthly monitoring remote sensing image automatic interpretation and result management system.

Description

Technical Field

[0002] This invention relates to the fields of geographic information extraction and spatiotemporal change analysis, specifically to an automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines. Background Technology

[0004] Geographic information extraction (GIS) technology utilizes computer vision and deep learning methods to automatically identify various geographic features from remote sensing images and transform their location, boundaries, and attributes into structured vector data and semantic information. This technology leverages the unprecedented data foundation provided by high-resolution satellite and UAV remote sensing. Deep learning—especially convolutional neural networks, Transformer, and U-Net architectures—has become the core engine for automatically learning geographic features from massive amounts of imagery and achieving pixel-level recognition. Cloud computing and distributed computing platforms provide the necessary computing power and storage support for processing TB / PB-level image data. Simultaneously, automatic annotation algorithms, open-source deep learning frameworks, and crowdsourced geographic data have jointly promoted the construction of high-quality sample libraries and the rapid iteration of algorithms. Furthermore, the development of multimodal data fusion and edge computing devices is driving this technology towards greater precision, real-time performance, and automation, enabling it to accurately extract vectorized geographic feature information from complex scenes. The synergy of these technologies lays a solid technical foundation for the automatic interpretation of remote sensing images in a monthly monitoring and management system for river and lake shorelines.

[0005] Spatiotemporal change analysis technology is a technique that automatically compares and intelligently interprets multi-temporal remote sensing images to identify, quantify, and interpret the dynamic evolution of land cover and geographic elements. This technology is based on the continuous image streams provided by increasingly abundant high-spatiotemporal resolution satellite constellations and UAV remote sensing networks. It relies on advanced change detection models such as deep learning-based twin networks and temporal segmentation models, as well as hidden Markov chain temporal analysis algorithms. Cloud computing and geographic big data platforms provide crucial computing power and storage support for processing massive amounts of time-series data, enabling spatiotemporal change analysis to support long-term, large-scale dynamic monitoring. Simultaneously, multi-source information fusion technology effectively overcomes the limitations of single data sources, enhancing the robustness of change identification and all-weather monitoring capabilities. These technologies collectively propel this field from traditional manual interpretation to a new stage of automated, quantitative, and knowledge-based analysis, achieving accurate and timely perception and understanding of dynamic land surface processes. The continuous development of these technologies lays a solid technical foundation for the monthly dynamic analysis of river and lake shorelines in an automatic interpretation and results management system for monthly monitoring remote sensing images. Summary of the Invention

[0007] To address the aforementioned issues, this invention aims to provide an automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines, comprising a data resource management and access module, a remote sensing image intelligent preprocessing module, an intelligent interpretation and change recognition module, a human-computer collaborative interaction and correction module, and a database and report generation module. The data resource management and access module is used to access multi-source heterogeneous data; the remote sensing image intelligent preprocessing module is used for standardization processing of raw images; and the intelligent interpretation and change recognition module includes intelligent shoreline element... The system includes an interpretation unit and a monthly change intelligent monitoring unit. The shoreline element intelligent interpretation unit proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm for automatic interpretation of remote sensing images. The monthly change intelligent monitoring unit proposes an improved cross-temporal attention and uncertainty collaborative change detection algorithm for monthly dynamic analysis of river and lake shorelines. The human-computer collaborative interaction and correction module is used for human-computer interaction and result correction. The database and report generation module includes a monitoring results database unit and a monitoring report automatic generation unit. The monitoring results database unit is used for spatiotemporal structured storage, and the monitoring report automatic generation unit is used for automatic output of monitoring results.

[0009] Furthermore, the data resource management and access module is used to access monthly image data and basic geographic data. It automatically pulls the latest remote sensing images covering the target river and lake area from the designated satellite / aerial remote sensing data platform and records its metadata, managing and maintaining all static / quasi-static reference data.

[0010] Furthermore, the remote sensing image intelligent preprocessing module is used for the standardization of raw images, eliminating physical and geometric distortions of the images. After receiving the raw images from the data resource management and access module, it eliminates the effects of atmospheric scattering, absorption, and differences in sensor response. It also uses ground control points and high-precision DEMs to correct geometric deformations caused by sensor attitude and terrain undulations, so that the images have accurate geographic coordinates and can be accurately overlaid with basic geographic data.

[0011] Furthermore, the shoreline element intelligent interpretation unit proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm for automatic interpretation of remote sensing images, and automatically identifies various land features in the pre-processed images.

[0012] Furthermore, the improved adaptive gating network sub-pixel shoreline interpretation algorithm is as follows: After receiving image data processed by the remote sensing image intelligent preprocessing module, to address the issue of uneven perception of large and small features by fixed-size tiles, a lightweight network is first used to analyze the image, quickly evaluating the suggested scale and key score map for each pixel location, i.e.: , ,in, For suggested standards, For key score plots, It is a lightweight neural network model. For complete remote sensing images from the preprocessing module, For pixels Recommended standards For the membership operator, This is the first slice size grade. This is the second slice size grade. For the first The image is then divided into several slice size levels, and a non-uniform grid is proposed for adaptive slicing. The slicing rule function is defined as follows: ,in, This results in a final output set of image tiles of varying sizes. For the tile index, For the first Image patch For adaptive slicing rule functions, A key score threshold is used to control the granularity of the slices, and a corresponding downsampled version is generated for each slice, which together serve as the input to subsequent models, thus constructing a multi-scale input stream. ,in, For the first A multi-scale input stream constructed from individual tiles For the tiles Downsampling operations are performed to obtain a low-resolution version. To simultaneously capture both global context and local details of the image, a hybrid model fusing Transformer and CNN is proposed for pixel-level classification. The Transformer branch captures global context, while the CNN branch extracts local details. Adaptive fusion is achieved through a gating mechanism. Feature extraction and fusion are then performed as follows: , ,in, The output features of the Transformer encoder branch This is an encoder module based on the Transformer architecture. The output features of the CNN encoder branch, For the encoder module based on a convolutional neural network, gating fusion is... , ,in, For the gating weight graph, It is the Sigmoid activation function. This is a learnable linear transformation weight matrix used to map the concatenated features from dimensionality to 1D. To achieve adaptive weighted fusion of features through a gating mechanism, The final output is the result of element-wise multiplication. ,in, For pixels The uncertainty of the forecast, Pixels predicted by the model Belongs to the The probability values ​​of each category, To perform logarithmic operations, a global optimization model is constructed to solve the problems of noise, small holes, and boundary smoothing in one go, resulting in an optimized classification raster, namely: , ,in, For the label field of the entire image, To define the labels in the entire image Gibbs energy on top For pixels The tag, For pixels The tag, Give the model pixels The corresponding tag value is Time probability value, The weighting coefficients of the smoothing term are used to control the strength of the smoothing constraint. For adjacent pixel pairs Adaptive smoothing weights, Proportional to the operator, For exponential operations, For pixels Color vectors in the original image For pixels Color vectors in the original image To control the coefficient of color similarity sensitivity, To control the impact of uncertainty on the smoothing weights, For pixels To address the prediction uncertainty and obtain shoreline accuracy superior to pixel-level precision, a smooth and accurate sub-pixel-level shoreline vector is finally extracted from the optimized water body probability map using an active contour model. ,in, For parametric contour curves, To define on the contour curve Total energy of the active contour model on the surface, For along the curve The arc length integral, For curves For arc length The first derivative, For curves For arc length The second derivative, As an energy field defined by the image data itself, this paper proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm. First, a lightweight network is used to analyze the image to address the uneven perception of large and small features by fixed-size slices. Then, a non-uniform grid is proposed to adaptively slice the image, enabling intelligent processing by using large slices for large areas to improve efficiency and small slices for key details to preserve accuracy. Next, a gated fusion Transformer-CNN hybrid model is proposed to simultaneously capture the global context and local details of the image. Finally, prediction uncertainty quantification and joint optimization strategies are deeply fused to effectively remove noise and artifacts and obtain shorelines with higher accuracy than pixels. This enables automatic interpretation of remote sensing images and automatic identification of various land features in the preprocessed images.

[0013] Furthermore, the monthly change intelligent monitoring unit proposes an improved cross-temporal attention and uncertainty collaborative change detection algorithm to conduct monthly dynamic analysis of river and lake shorelines. By comparing the shoreline vectors extracted in the current period with those extracted in the previous period, it automatically calculates the horizontal advance and retreat distance and area changes of the shoreline, identifies erosion sections, siltation sections and stable sections, and quantifies the amount of change.

[0014] Furthermore, the improved algorithm for detecting cross-temporal attention and uncertainty co-change is as follows: Two registered images from the remote sensing image intelligent preprocessing module are received. First, a pre-trained Siamese encoder network is used to simultaneously extract depth features from both images, i.e. ,in, This is the depth feature map of this image. This is the depth feature map of the previous image. For twin encoder networks, This is the registered remote sensing imagery for this period. Based on the registered remote sensing images from the previous issue, a cross-temporal attention module is proposed to weight features to focus on the content that has truly changed, i.e. ,in, For cross-temporal attention weights, It is the Sigmoid activation function. For convolution operations in convolutional layers, To integrate the two phases in the channel dimension and The feature maps are concatenated, and then attention-enhanced feature differences are generated, along with the initial probability map. ,in, This is the probability diagram of the first generation of changes. For decoder networks, For element-wise multiplication, the algorithm then performs synchronous estimation to predict uncertainty, followed by uncertainty-guided morphological segmentation. Reliable change feature cores are generated preferentially in low-uncertainty regions. Monte Carlo Dropout sampling is used to estimate the prediction variance to reliably extract change feature cores in high-confidence regions with lower uncertainty. , ,in, To predict uncertainty graphs, For variance calculation, For the first The initial change probability map obtained from Monte Carlo Dropout sampling. The total number of Monte Carlo Dropout samples, with uncertain guided threshold segmentation. ,in, To change the binary mask of the core region, For the mathematical symbol of the indicator function, The threshold for the probability of change. For uncertainty threshold, Using the logical AND operator, we guide the initial extraction of changed patches. Then, we perform joint verification of spectral feature distance and temporal persistence for each candidate patch to effectively eliminate spurious changes caused by registration errors, shading, and seasonal fluctuations. The average distance of the patch in the two-period depth feature space is calculated. ,in, This represents the average distance of the patch in the depth feature space. Let be the average feature vector of the current period's depth features within a candidate changed patch. Let be the average feature vector of the previous period's depth features within a candidate changed patch, and let the time-series persistence be verified as follows: ,in, As a time series persistence indicator, It is a mean function. From Expected Historical change probability diagram of the period For the first The algorithm identifies candidate variation patches and then iteratively optimizes the boundaries of the validated patches to achieve sub-pixel level precision. ,in, To define at the boundary Total energy on For the probability diagram of change In pixels The gradient vector at that point, The smoothing term weighting coefficients are used to control the smoothness of the boundary. For the boundary The total length, and automatically calculates the class inference and comprehensive confidence score of the intelligent attributes before and after the change, the main class inference before and after the change is , ,in, The inferred area of ​​this patch is in The main land cover categories of the period For clustering and decision functions, The inferred area of ​​this patch is in The main land cover categories during the period, with a comprehensive confidence score of ,in, To achieve a comprehensive confidence score, the final output is a change detection layer with high-precision vector boundaries and rich semantic attributes. An improved cross-temporal attention and uncertainty-based collaborative change detection algorithm is proposed. First, a cross-temporal attention module is introduced to weight features, enabling the model to dynamically focus on truly changing land features. Then, the prediction uncertainty of Monte Carlo Dropout estimation is transformed into a reliable basis for guiding change area extraction, prioritizing the identification of high-confidence change cores. Next, a dual constraint of spectral feature distance and temporal persistence index is proposed to effectively distinguish between real surface changes and pseudo-changes such as seasonal fluctuations and shadows. Finally, iterative boundary optimization for sub-pixel accuracy and intelligent attribute generation incorporating multi-dimensional information are proposed to ensure that the output is not only a high-precision vector boundary but also includes change category inference and a comprehensive confidence score. This allows for monthly dynamic analysis of river and lake shorelines, automatically triggering precise measurement of the vectors of corresponding shoreline segments in two consecutive periods, producing quantitative results on shoreline advance / retreat distance, erosion and sedimentation area, and spatial distribution.

[0015] Furthermore, the human-computer collaborative interaction and correction module is used for human-computer interaction and result correction. Through necessary human intervention, it corrects the limitations of the AI ​​model and ensures that every shoreline, every map patch, and every change entered into the database is real and reliable.

[0016] Furthermore, through the design of a spatiotemporal data model, the monitoring results database unit assigns a timestamp to each shoreline and map patch, forming a complete historical version chain, and categorizes and stores all types of results data.

[0017] Furthermore, the monitoring report automatic generation unit is used to automatically output monitoring results, automatically extract key data for the month from the results database, and generate a monthly shoreline monitoring report with a standardized format.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] 1. This invention proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm for automatic interpretation of remote sensing images. The innovation of this invention lies in the following: First, the improved adaptive gating network sub-pixel shoreline interpretation algorithm analyzes the image through a lightweight network to solve the problem of uneven perception of large and small features by fixed-size slices. Then, it proposes a non-uniform grid to adaptively slice the image to achieve intelligent processing by using large slices for large areas to improve efficiency and small slices for key details to preserve accuracy. Next, it proposes a gated fusion Transformer-CNN hybrid model that can simultaneously capture the global context and local details of the image. Finally, it deeply fuses prediction uncertainty quantification and joint optimization strategies to effectively remove noise and artifacts and obtain shorelines with higher accuracy than pixels. This enables automatic interpretation of remote sensing images and automatic identification of various features in the preprocessed image.

[0020] 2. This invention proposes an improved algorithm for detecting cross-temporal attention and uncertainty-based collaborative changes in river and lake shorelines for monthly dynamic analysis. The innovation lies in the following: First, the improved algorithm uses a cross-temporal attention module to weight features, enabling the model to dynamically focus on truly changing land features. Then, the prediction uncertainty of Monte Carlo Dropout estimation is transformed into a reliable basis for guiding the extraction of change areas, prioritizing the identification of high-confidence change cores. Next, a dual constraint of spectral feature distance and temporal persistence index is proposed to effectively distinguish between real surface changes and pseudo-changes such as seasonal fluctuations and shadows. Finally, it proposes iterative boundary optimization for sub-pixel accuracy and intelligent attribute generation that integrates multi-dimensional information, ensuring that the output is both a high-precision vector boundary and includes change category prediction and a comprehensive confidence score. This allows for monthly dynamic analysis of river and lake shorelines, automatically triggering precise measurement of the vectors of the corresponding shoreline segments in the preceding and following periods, producing quantitative results on shoreline advance / retreat distance, erosion and sedimentation area, and spatial distribution. Attached Figure Description

[0022] The invention will be further illustrated with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the invention. For those skilled in the art, other drawings can be obtained based on the following drawings without any creative effort.

[0023] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] An automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines is disclosed. This system comprises a data resource management and access module, a remote sensing image intelligent preprocessing module, an intelligent interpretation and change recognition module, a human-computer collaborative interaction and correction module, and a database and report generation module. The data resource management and access module is used to access multi-source heterogeneous data. The remote sensing image intelligent preprocessing module is used for standardization processing of raw images. The intelligent interpretation and change recognition module includes a shoreline element intelligent interpretation unit and a monthly change intelligent monitoring unit. The shoreline element intelligent interpretation unit proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm for automatic interpretation of remote sensing images. The monthly change intelligent monitoring unit proposes an improved cross-temporal attention and uncertainty collaborative change detection algorithm for monthly dynamic analysis of river and lake shorelines. The human-computer collaborative interaction and correction module is used for human-computer interaction and result correction. The database and report generation module includes a monitoring results database unit and a monitoring report automatic generation unit. The monitoring results database unit is used for spatiotemporal structured storage, and the monitoring report automatic generation unit is used to automatically output monitoring results.

[0027] Preferably, the data resource management and access module is used to access monthly image data and basic geographic data. It automatically pulls the latest remote sensing images covering the target river and lake area from the designated satellite / aerial remote sensing data platform and records its metadata. It manages and maintains all static / quasi-static reference data. The monthly image data includes the latest periodic satellite / aerial remote sensing images acquired according to the plan. The basic geographic data includes administrative division vectors, historical river and lake shorelines and water areas, locations of water conservancy facilities, digital elevation models, and ecological protection red line range data.

[0028] Preferably, the remote sensing image intelligent preprocessing module is used for the standardization processing of the original image, eliminating the physical and geometric distortions of the image. After receiving the original image from the data resource management and access module, it eliminates the influence of atmospheric scattering, absorption and sensor response differences, so that the image pixel values ​​truly reflect the spectral reflectance of the ground objects. It also uses ground control points and high-precision DEM to correct the geometric deformation caused by sensor attitude and terrain undulation, so that the image has accurate geographic coordinates and achieves accurate overlay with basic geographic data.

[0029] Preferably, the shoreline feature intelligent interpretation unit proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm for automatic interpretation of remote sensing images, and automatically identifies various land features in the preprocessed images.

[0030] Specifically, the improved adaptive gating network sub-pixel shoreline interpretation algorithm is as follows: After receiving image data processed by the remote sensing image intelligent preprocessing module, to address the issue of uneven perception of large and small features by fixed-size tiles, a lightweight network is first used to analyze the image, quickly evaluating the suggested scale and key score map for each pixel location, i.e.: , ,in, For suggested standards, For key score plots, It is a lightweight neural network model. For complete remote sensing images from the preprocessing module, For pixels Recommended standards For the membership operator, This is the first slice size grade. This is the second slice size grade. For the first The paper proposes several slice size levels and then uses a non-uniform grid for adaptive slicing of the image. Larger slices are used in large-scale, low-criticality areas to improve efficiency, while smaller slices are used in small-scale, high-criticality areas to improve accuracy. The slicing rule function is defined as follows: ,in, This results in a final output set of image tiles of varying sizes. For the tile index, For the first Image patch For adaptive slicing rule functions, A key score threshold is used to control the granularity of the slices, and a corresponding downsampled version is generated for each slice, which together serve as the input to subsequent models, thus constructing a multi-scale input stream. ,in, For the first A multi-scale input stream constructed from individual tiles For the tiles Downsampling operations are performed to obtain a low-resolution version, providing explicit multi-scale contextual information for subsequent models. This allows the model to grasp both the global water system structure and distinguish local details, simultaneously capturing both global context and local details. A hybrid model fusing Transformer and CNN is then proposed for pixel-level classification. The Transformer branch captures global context to understand the overall water system structure, while the CNN branch extracts local details to distinguish subtle edges. Adaptive fusion is achieved through a gating mechanism. Feature extraction and fusion are... , ,in, The output features of the Transformer encoder branch This is an encoder module based on the Transformer architecture. The output features of the CNN encoder branch, For the encoder module based on a convolutional neural network, gating fusion is... , ,in, For the gating weight graph, It is the Sigmoid activation function. This is a learnable linear transformation weight matrix used to map the concatenated features from dimensionality to 1D. To achieve adaptive weighted fusion of features through a gating mechanism, The final output is the result of element-wise multiplication. ,in, For pixels The uncertainty of the forecast, Pixels predicted by the model Belongs to the The probability values ​​of each category, To perform logarithmic operations, a global optimization model is constructed to solve the problems of noise, small holes, and boundary smoothing in one go, resulting in an optimized classification raster, namely: , ,in, For the label field of the entire image, To define the labels in the entire image Gibbs energy on top For pixels The tag, For pixels The tag, Give the model pixels The corresponding tag value is Time probability value, The weighting coefficients of the smoothing term are used to control the strength of the smoothing constraint. For adjacent pixel pairs Adaptive smoothing weights, Proportional to the operator, For exponential operations, For pixels Color vectors in the original image For pixels Color vectors in the original image To control the coefficient of color similarity sensitivity, To control the impact of uncertainty on the smoothing weights, For pixels To address the prediction uncertainty and obtain shoreline accuracy superior to pixel-level precision, a smooth and accurate sub-pixel-level shoreline vector is finally extracted from the optimized water body probability map using an active contour model. ,in, For parametric contour curves, To define on the contour curve Total energy of the active contour model on the surface, For along the curve The arc length integral, For curves For arc length The first derivative, For curves For arc length The second derivative, As an energy field defined by the image data itself, this paper proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm. First, a lightweight network is used to analyze the image to address the uneven perception of large and small features by fixed-size slices. Then, a non-uniform grid is proposed to adaptively slice the image, enabling intelligent processing by using large slices for large areas to improve efficiency and small slices for key details to preserve accuracy. Next, a gated fusion Transformer-CNN hybrid model is proposed to simultaneously capture the global context and local details of the image. Finally, prediction uncertainty quantification and joint optimization strategies are deeply fused to effectively remove noise and artifacts and obtain shorelines with higher accuracy than pixels. This enables automatic interpretation of remote sensing images and automatic identification of various land features in the preprocessed images.

[0031] Preferably, the monthly change intelligent monitoring unit proposes an improved cross-temporal attention and uncertainty collaborative change detection algorithm to conduct monthly dynamic analysis of river and lake shorelines. By comparing the shoreline vectors extracted in the current period with those extracted in the previous period, it automatically calculates the horizontal advance and retreat distance and area change of the shoreline, identifies erosion sections, siltation sections and stable sections, and quantifies the amount of change.

[0032] Specifically, the improved algorithm for detecting cross-temporal attention and uncertainty co-change is as follows: Two registered images from the remote sensing image intelligent preprocessing module are received. First, a pre-trained Siamese encoder network is used to simultaneously extract depth features from both images, i.e. ,in, This is the depth feature map of this image. This is the depth feature map of the previous image. For twin encoder networks, This is the registered remote sensing imagery for this period. Based on the registered remote sensing images from the previous issue, a cross-temporal attention module is proposed to weight features to focus on the content that has truly changed, i.e. ,in, For cross-temporal attention weights, It is the Sigmoid activation function. For convolution operations in convolutional layers, To integrate the two phases in the channel dimension and The feature maps are concatenated, and then attention-enhanced feature differences are generated, along with the initial probability map. ,in, This is the initial probability map, representing the model's preliminary assessment of the probability that the pixel position has changed. For decoder networks, For element-wise multiplication, the algorithm then performs synchronous estimation to predict uncertainty, followed by uncertainty-guided morphological segmentation. Reliable change feature cores are generated preferentially in low-uncertainty regions. Monte Carlo Dropout sampling is used to estimate the prediction variance to reliably extract change feature cores in high-confidence regions with lower uncertainty. , ,in, The uncertainty map represents the degree of uncertainty the model's judgment on changes in each pixel value in the map. A larger variance indicates greater inconsistency in the model's predictions across multiple iterations, signifying higher uncertainty. For variance calculation, For the first The initial change probability map obtained from Monte Carlo Dropout sampling. The total number of Monte Carlo Dropout samples, with uncertain guided threshold segmentation. ,in, To change the binary mask of the core region, The mathematical symbol for an indicator function, used to determine combination conditions. Whether it is valid, The threshold for the probability of change. For uncertainty threshold, Using the logical AND operator, we guide the initial extraction of changed patches. Then, we perform joint verification of spectral feature distance and temporal persistence for each candidate patch to effectively eliminate spurious changes caused by registration errors, shading, and seasonal fluctuations. The average distance of the patch in the two-period depth feature space is calculated. ,in, This represents the average distance of the patch in the depth feature space. Let be the average feature vector of the current period's depth features within a candidate changed patch. Let be the average feature vector of the previous period's depth features within a candidate changed patch, and let the time-series persistence be verified as follows: ,in, As a time series persistence indicator, It is a mean function. From Expected Historical change probability diagram of the period For the first The algorithm identifies candidate variation patches and then iteratively optimizes the boundaries of the validated patches to achieve sub-pixel level precision. ,in, To define at the boundary Total energy on For the probability diagram of change In pixels The gradient vector at that point, The smoothing term weighting coefficients are used to control the smoothness of the boundary. For the boundary The total length, and automatically calculates the class inference and comprehensive confidence score of the intelligent attributes before and after the change, the main class inference before and after the change is , ,in, The inferred area of ​​this patch is in The main land cover categories of the period For clustering and decision functions, The inferred area of ​​this patch is in The main land cover categories during the period, with a comprehensive confidence score of ,in, To achieve a comprehensive confidence score, a change detection layer with high-precision vector boundaries and rich semantic attributes is output. An improved cross-temporal attention and uncertainty-based collaborative change detection algorithm is proposed. First, a cross-temporal attention module is introduced to weight features, enabling the model to dynamically focus on truly changing land cover content. Then, the prediction uncertainty of Monte Carlo Dropout estimation is transformed into a reliable basis for guiding change area extraction, prioritizing the identification of high-confidence change cores. Finally, a dual constraint of spectral feature distance and temporal persistence index is proposed to effectively distinguish between real surface changes and pseudo-changes such as seasonal fluctuations and shading. Finally, an iterative boundary optimization approach oriented towards sub-pixel accuracy and intelligent attribute generation that integrates multi-dimensional information are proposed to ensure that the output results are not only high-precision vector boundaries, but also include change category prediction and comprehensive confidence scores. This approach is used for monthly dynamic analysis of river and lake shorelines, automatically triggering precise measurement of the vectors of the corresponding shoreline segments in two periods, producing quantitative results of shoreline advance and retreat distance, erosion and siltation area, and spatial distribution. These results are then transmitted to the downstream human-computer collaborative interaction and correction module, providing professionals with accurate and reliable judgment targets, thus forming a complete intelligent closed loop from automatic image processing to precise quantification of shoreline dynamics.

[0033] Preferably, the human-computer collaborative interaction and correction module is used for human-computer interaction and result correction. It performs vector drawing, editing, deletion and reclassification of shorelines / plots that are mistakenly extracted or missed by AI. At the same time, it reviews each changed plot automatically detected by the system to confirm its authenticity. Through necessary manual intervention, it corrects the limitations of the AI ​​model to ensure that every shoreline, every plot and every change in the database is real and reliable.

[0034] Preferably, the monitoring results database unit, through the design of a spatiotemporal data model, assigns a timestamp to each shoreline and map patch, forming a complete historical version chain, and classifies and stores all types of results data. At the same time, it provides powerful spatiotemporal query capabilities, integrating scattered monthly results into a systematic and traceable digital archive of river and lake shorelines.

[0035] Preferably, the monitoring report automatic generation unit is used to automatically output monitoring results, automatically extract key data for the month from the monitoring results database, generate a monthly shoreline monitoring report with standardized format and rich graphics, and automatically complete the summary and presentation of monitoring results, greatly reducing the workload of manual report preparation.

[0036] This paper proposes an automatic interpretation and results management system for monthly remote sensing images of river and lake shorelines. This system enables intelligent interpretation of monthly remote sensing images and intelligent detection of monthly dynamic changes, supporting routine and precise monitoring of river and lake shorelines. It integrates a data resource management and access module, a remote sensing image intelligent preprocessing module, an intelligent interpretation and change recognition module, a human-computer collaborative interaction and correction module, and a database and report generation module. The system provides an improved adaptive gating network sub-pixel shoreline interpretation algorithm for automatic interpretation of remote sensing images. The innovation of this invention lies in the fact that the improved adaptive gating network sub-pixel shoreline interpretation algorithm first uses a lightweight network to automatically interpret the images. This paper analyzes images to address the uneven perception of large and small features by fixed-size tiling. It then proposes a non-uniform grid for adaptive tiling, using large tiling for large areas to improve efficiency and small tiling for key details to preserve precision. Next, a gated fusion Transformer-CNN hybrid model is proposed to simultaneously capture global context and local details of the image. Finally, prediction uncertainty quantification and joint optimization strategies are deeply fused to effectively remove noise and artifacts and obtain shoreline accuracy superior to pixel-level precision. This enables automatic interpretation of remote sensing images and automatic identification of various land features in preprocessed images. An improved algorithm for cross-temporal attention and uncertainty co-change detection is also proposed for monthly dynamic analysis of river and lake shorelines. Analysis reveals that the innovation of this invention lies in its improved cross-temporal attention and uncertainty collaborative change detection algorithm. Firstly, it proposes a cross-temporal attention module to weight features, enabling the model to dynamically focus on truly changing land cover content. Then, it transforms the prediction uncertainty of Monte Carlo Dropout estimation into a reliable basis for guiding change area extraction, prioritizing the identification of high-confidence change cores. Next, it proposes dual constraints of spectral feature distance and temporal persistence indices to effectively distinguish between real surface changes and pseudo-changes such as seasonal fluctuations and shadows. Finally, it proposes iterative boundary optimization for sub-pixel accuracy and intelligent attribute generation that integrates multi-dimensional information, ensuring that the output is both a high-precision vector boundary and includes change category inferences. This invention utilizes measurement and comprehensive confidence scoring to conduct monthly dynamic analysis of river and lake shorelines. It automatically triggers precise calculations of vectors for corresponding shoreline segments in two consecutive periods, producing quantitative results on shoreline advance / retreat distances, erosion and sedimentation areas, and spatial distribution. This effectively improves the performance of a monthly remote sensing image interpretation and results management system for river and lake shorelines, providing more comprehensive and accurate technical support. Furthermore, this invention relates to geographic information extraction and spatiotemporal change analysis technologies, offering a precise and efficient one-stop solution for intelligent data interpretation and dynamic change detection in the routine monitoring of river and lake shorelines, contributing significant application value to the intelligent monitoring of river and lake shorelines.

[0037] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines, characterized in that, This invention includes an automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines. The system comprises a data resource management and access module, a remote sensing image intelligent preprocessing module, an intelligent interpretation and change recognition module, a human-computer collaborative interaction and correction module, and a database and report generation module. The data resource management and access module is used to access multi-source heterogeneous data. The remote sensing image intelligent preprocessing module is used for standardization processing of raw images. The intelligent interpretation and change recognition module includes a shoreline element intelligent interpretation unit and a monthly change intelligent monitoring unit. The shoreline element intelligent interpretation unit proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm for automatic interpretation of remote sensing images. The monthly change intelligent monitoring unit proposes an improved cross-temporal attention and uncertainty collaborative change detection algorithm for monthly dynamic analysis of river and lake shorelines. The human-computer collaborative interaction and correction module is used for human-computer interaction and result correction. The database and report generation module includes a monitoring results database unit and a monitoring report automatic generation unit. The monitoring results database unit is used for spatiotemporal structured storage, and the monitoring report automatic generation unit is used to automatically output monitoring results.

2. The automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines according to claim 1, characterized in that, The data resource management and access module is used to access monthly image data and basic geographic data. It automatically pulls the latest remote sensing images covering the target river and lake area from the designated satellite / aerial remote sensing data platform and records its metadata, and manages and maintains all static / quasi-static reference data.

3. The automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines according to claim 1, characterized in that, The remote sensing image intelligent preprocessing module is used for the standardization of raw images, eliminating physical and geometric distortions. After receiving raw images from the data resource management and access module, it eliminates the effects of atmospheric scattering, absorption, and differences in sensor response. It also uses ground control points and high-precision DEMs to correct geometric deformations caused by sensor attitude and terrain undulations, giving the images accurate geographic coordinates and enabling precise overlay with basic geographic data.

4. The automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines according to claim 1, characterized in that, The shoreline element intelligent interpretation unit proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm for automatic interpretation of remote sensing images, and automatically identifies various land features in the pre-processed images.

5. The automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines according to claim 4, characterized in that, The improved sub-pixel shoreline interpretation algorithm of the adaptive gating network is as follows: After receiving image data processed by the intelligent preprocessing module of remote sensing imagery, in order to solve the problem of uneven perception of large and small features by fixed-size tiles, a lightweight network is first used to analyze the image, quickly evaluating the suggested scale and key score map of each pixel location, i.e.: , ,in, For suggested standards, For key score plots, It is a lightweight neural network model. For complete remote sensing images from the preprocessing module, For pixels Recommended standards For the membership operator, This is the first slice size grade. This is the second slice size grade. For the first The image is then divided into several slice size levels, and a non-uniform grid is proposed for adaptive slicing. The slicing rule function is defined as follows: ,in, This results in a final output set of image tiles of varying sizes. For the tile index, For the first Image patch For adaptive slicing rule functions, A key score threshold is used to control the granularity of the slices, and a corresponding downsampled version is generated for each slice, which together serve as the input to subsequent models, thus constructing a multi-scale input stream. ,in, For the first A multi-scale input stream constructed from individual tiles For the tiles Downsampling operations are performed to obtain a low-resolution version. To simultaneously capture both global context and local details of the image, a hybrid model fusing Transformer and CNN is proposed for pixel-level classification. The Transformer branch captures global context, while the CNN branch extracts local details. Adaptive fusion is achieved through a gating mechanism. Feature extraction and fusion are then performed as follows: , ,in, The output features of the Transformer encoder branch This is an encoder module based on the Transformer architecture. The output features of the CNN encoder branch, For the encoder module based on a convolutional neural network, gating fusion is... , ,in, For the gating weight graph, It is the Sigmoid activation function. This is a learnable linear transformation weight matrix used to map the concatenated features from dimensionality to 1D. To achieve adaptive weighted fusion of features through a gating mechanism, The final output is the result of element-wise multiplication. ,in, For pixels The uncertainty of the forecast, Pixels predicted by the model Belongs to the The probability values ​​of each category, To perform logarithmic operations, a global optimization model is constructed to solve the problems of noise, small holes, and boundary smoothing in one go, resulting in an optimized classification raster, namely: , ,in, For the label field of the entire image, To define the labels in the entire image Gibbs energy on top For pixels The tag, For pixels The tag, Give the model pixels The corresponding tag value is Time probability value, The weighting coefficients of the smoothing term are used to control the strength of the smoothing constraint. For adjacent pixel pairs Adaptive smoothing weights, Proportional to the operator, For exponential operations, For pixels Color vectors in the original image For pixels Color vectors in the original image To control the coefficient of color similarity sensitivity, To control the impact of uncertainty on the smoothing weights, For pixels To address the prediction uncertainty and obtain shoreline accuracy superior to pixel-level precision, a smooth and accurate sub-pixel-level shoreline vector is finally extracted from the optimized water body probability map using an active contour model. ,in, For parametric contour curves, To define on the contour curve Total energy of the active contour model on the surface, For along the curve The arc length integral, For curves For arc length The first derivative, For curves For arc length The second derivative, As an energy field defined by the image data itself, this paper proposes an improved adaptive gating network sub-pixel shoreline interpretation algorithm. First, a lightweight network is used to analyze the image to address the uneven perception of large and small features by fixed-size slices. Then, a non-uniform grid is proposed to adaptively slice the image, enabling intelligent processing by using large slices for large areas to improve efficiency and small slices for key details to preserve accuracy. Next, a gated fusion Transformer-CNN hybrid model is proposed to simultaneously capture the global context and local details of the image. Finally, prediction uncertainty quantification and joint optimization strategies are deeply fused to effectively remove noise and artifacts and obtain shorelines with higher accuracy than pixels. This enables automatic interpretation of remote sensing images and automatic identification of various land features in the preprocessed images.

6. The automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines according to claim 1, characterized in that, The monthly change intelligent monitoring unit proposes an improved cross-temporal attention and uncertainty collaborative change detection algorithm to conduct monthly dynamic analysis of river and lake shorelines. By comparing the shoreline vectors extracted in the current period with those extracted in the previous period, it automatically calculates the horizontal advance and retreat distance and area changes of the shoreline, identifies erosion sections, siltation sections and stable sections, and quantifies the amount of change.

7. The automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines according to claim 6, characterized in that, The improved algorithm for detecting co-current attention and uncertainty changes across time is as follows: Two registered images from the remote sensing image intelligent preprocessing module are received. First, a pre-trained Siamese encoder network is used to simultaneously extract depth features from both images. ,in, This is the depth feature map of this image. This is the depth feature map of the previous image. For twin encoder networks, This is the registered remote sensing imagery for this period. Based on the registered remote sensing images from the previous issue, a cross-temporal attention module is proposed to weight features to focus on the content that has truly changed, i.e. ,in, For cross-temporal attention weights, It is the Sigmoid activation function. For convolution operations in convolutional layers, To integrate the two phases in the channel dimension and The feature maps are concatenated, and then attention-enhanced feature differences are generated, along with the initial probability map. ,in, This is the probability diagram of the first generation of changes. For decoder networks, For element-wise multiplication, the algorithm then performs synchronous estimation to predict uncertainty, followed by uncertainty-guided morphological segmentation. Reliable change feature cores are generated preferentially in low-uncertainty regions. Monte Carlo Dropout sampling is used to estimate the prediction variance to reliably extract change feature cores in high-confidence regions with lower uncertainty. , ,in, To predict uncertainty graphs, For variance calculation, For the first The initial change probability map obtained from Monte Carlo Dropout sampling. The total number of Monte Carlo Dropout samples, with uncertain guided threshold segmentation. ,in, To change the binary mask of the core region, For the mathematical symbol of an indicator function, The threshold for the probability of change. For uncertainty threshold, Using the logical AND operator, we guide the initial extraction of changed patches. Then, we perform joint verification of spectral feature distance and temporal persistence for each candidate patch to effectively eliminate spurious changes caused by registration errors, shading, and seasonal fluctuations. The average distance of the patch in the two-period depth feature space is calculated. ,in, This represents the average distance of the patch in the depth feature space. Let be the average feature vector of the current period's depth features within a candidate changed patch. Let be the average feature vector of the previous period's depth features within a candidate changed patch, and let the time-series persistence be verified as follows: ,in, As a time series persistence indicator, It is a mean function. From Expected Historical change probability diagram of the period For the first The algorithm identifies candidate variation patches and then iteratively optimizes the boundaries of the validated patches to achieve sub-pixel level precision. ,in, To define at the boundary Total energy on For the probability diagram of change In pixels The gradient vector at that point, The smoothing term weighting coefficients are used to control the smoothness of the boundary. For the boundary The total length, and automatically calculates the class inference and comprehensive confidence score of the intelligent attributes before and after the change, the main class inference before and after the change is , ,in, The inferred area of ​​this patch is in The main land cover categories of the period For clustering and decision functions, The inferred area of ​​this patch is in The main land cover categories during the period, with a comprehensive confidence score of ,in, To achieve a comprehensive confidence score, the final output is a change detection layer with high-precision vector boundaries and rich semantic attributes. An improved cross-temporal attention and uncertainty-based collaborative change detection algorithm is proposed. First, a cross-temporal attention module is introduced to weight features, enabling the model to dynamically focus on truly changing land features. Then, the prediction uncertainty of Monte Carlo Dropout estimation is transformed into a reliable basis for guiding change area extraction, prioritizing the identification of high-confidence change cores. Next, a dual constraint of spectral feature distance and temporal persistence index is proposed to effectively distinguish between real surface changes and pseudo-changes such as seasonal fluctuations and shadows. Finally, iterative boundary optimization for sub-pixel accuracy and intelligent attribute generation incorporating multi-dimensional information are proposed to ensure that the output is not only a high-precision vector boundary but also includes change category inference and a comprehensive confidence score. This allows for monthly dynamic analysis of river and lake shorelines, automatically triggering precise measurement of the vectors of corresponding shoreline segments in two consecutive periods, producing quantitative results on shoreline advance / retreat distance, erosion and sedimentation area, and spatial distribution.

8. The automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines according to claim 1, characterized in that, The human-computer collaborative interaction and correction module is used for human-computer interaction and result correction. Through necessary human intervention, it corrects the limitations of the AI ​​model and ensures that every shoreline, every patch, and every change entered into the database is real and reliable.

9. The automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines according to claim 1, characterized in that, The monitoring results database unit uses a spatiotemporal data model to assign timestamps to each shoreline and map patch, forming a complete historical version chain and storing all types of results data in a categorized manner.

10. The automatic interpretation and results management system for monthly monitoring remote sensing images of river and lake shorelines according to claim 1, characterized in that, The automatic monitoring report generation unit is used to automatically output monitoring results, automatically extract key data for the month from the results database, and generate a monthly shoreline monitoring report in a standardized format.

Images