GIS surveying and mapping data analysis processing system and method based on AI

Abstract

The invention discloses an AI-based GIS surveying and mapping data analysis processing system and method, and relates to the technical field of artificial intelligence. Original surveying and mapping image data are acquired and preprocessed to obtain surveying and mapping image data; performing shadow segmentation and binaryzation on the surveying and mapping image data to obtain a shadow mask pattern, and performing shadow removal on the surveying and mapping image data based on the shadow mask pattern to obtain shadow-free image data; based on the shadow-free image data and the time weighting function, extracting a time weighted feature map, presetting a change probability threshold, and converting the time weighted feature map into a change detection image; and inputting the surveying and mapping image data into the constructed surface feature classification task, outputting the surface feature category probability graph, and optimizing the change detection image based on the surface feature category probability graph to obtain the optimized change detection image, thereby improving the geographic change detection precision based on the surveying and mapping data.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and particularly to an AI-based GIS mapping data analysis and processing system and method. Background Art

[0002] GIS mapping data is data on geospatial information collected and processed using Geographic Information System (GIS) technology. The collection methods of GIS mapping data include field measurement, remote sensing, map digitization, etc. These data provide an important basis for geospatial change analysis, decision-making, etc.

[0003] Chinese Patent Application with Publication No. CN111339225A discloses a method and system for dynamic updating of urban geospatial data. The method includes discovering change events, annotating, evidence collection and reporting of change events; update management, comparing change event information to confirm whether to establish an update task, and publishing and tracking the update task; information recovery, receiving the update task published in the update management step, and collecting, mapping and recovering the update task information that needs to change; incremental update, extracting, updating and publishing thematic incremental information from the recovered information.

[0004] Buildings in the urban geographical environment are diverse and dense, with mutual occlusion and obvious shadow effects among buildings. The shadows between ground objects change over time. Existing algorithms are difficult to accurately distinguish shadow changes from actual ground object changes, resulting in a decrease in change detection accuracy. Summary of the Invention

[0005] This application aims to solve at least one of the technical problems in the related art to some extent. To this end, an object of this application is to propose an AI-based GIS mapping data analysis and processing system and method, which improves the accuracy of geographical change detection based on mapping data.

[0006] One aspect of this application provides an AI-based GIS mapping data analysis and processing method, including:

[0007] Step S100: Obtain the original mapping image data and preprocess it to obtain the mapping image data;

[0008] Step S200: Perform shadow segmentation and binarization on the mapping image data to obtain a shadow mask map, and remove the shadow from the mapping image data based on the shadow mask map to obtain shadow-free image data;

[0009] Step S300: Extract a time-weighted feature map based on the shadow-free image data and a time weight function, preset a change probability threshold, and convert the time-weighted feature map into a change detection image;

[0010] Step S400: Input the surveyed image data into the constructed ground object classification task, output the ground object category probability map, and optimize the change detection image based on the ground object category probability map to obtain the optimized change detection image.

[0011] The specific method for obtaining the original surveyed image data and preprocessing it to obtain the surveyed image data is as follows:

[0012] Step S110: Determine the time range and spatial range of the monitoring area;

[0013] Step S120: Obtain the original surveyed image data within the time range and spatial range, and preprocess it to obtain the surveyed image data I t 。

[0014] The specific method for performing shadow segmentation and binarization on the surveyed image data to obtain the shadow mask map, and removing the shadow from the surveyed image data based on the shadow mask map to obtain the shadow-free image data is as follows:

[0015] Step S210: Select a deep learning model as the initial shadow segmentation model, use the surveyed image data as the input, and use the shadow probability map as the output;

[0016] Step S220: Obtain the shadow segmentation training set, where the shadow segmentation training set includes the surveyed image data and its corresponding shadow annotation mask. Use the cross-entropy loss function to train the shadow segmentation model, minimize the cross-entropy loss function between the true value of the shadow annotation mask of the pixel and the predicted shadow probability of the pixel, optimize the model parameters, and when the value of the cross-entropy loss function converges, obtain the trained shadow segmentation model;

[0017] Step S230: Use the trained shadow segmentation model to perform shadow segmentation on the surveyed image data I t to output the shadow probability map P t ;

[0018] Step S240: Preset the shadow probability threshold τ, binarize the shadow probability map, determine the pixels with a shadow probability greater than or equal to the shadow probability threshold as shadow pixels, and set their shadow mask values equal to 1, determine the pixels with a shadow probability less than the shadow probability threshold as non-shadow pixels, and set their shadow mask values equal to 0 to obtain the shadow mask map of the shadow probability map The shadow pixels form the shadow area;

[0019] Step S250: Remove the shadow from the shadow area in the surveyed image data I t to obtain the shadow-free image data

[0020] The shadow removal of the shaded area in the surveyed image data I according to the shadow mask image means that: when the shadow mask value of the k-th pixel in the shadow probability map t is equal to 0, it indicates that the shadow probability of this pixel is less than the shadow probability threshold, so this pixel is judged not to belong to the shaded area, and thus its pixel value remains unchanged; otherwise, this pixel is judged to belong to the shaded area, and the original pixel value of the shaded area is restored through shadow removal.

[0021] The method for obtaining the shadow-free image data is as follows:

[0022] Step S251: Perform connected component labeling on the shadow mask image, label the connected shaded pixel groups as the same shadow connected component, and extract attribute features for each shadow connected component S a The attribute features include the area A a of the shadow connected component, the perimeter P a and the shape complexity PAR a and the image statistical features. The image statistical features include the mean μ a of the pixel values of the shaded pixels within the shadow connected component, the variance

[0023] Step S252: According to the attribute features of the shadow connected component, preset the area threshold, complexity threshold and variance threshold, design a decision function, select the removal strategy for the shaded area according to the decision function, and execute the corresponding removal strategy for each shadow connected component to obtain the shadow-free image data.

[0024] The decision rule of the decision function is: when the area of the shadow connected component is less than the area threshold and the shape complexity is less than the complexity threshold, perform shadow correction on the shaded area; when the area of the shadow connected component is greater than or equal to the area threshold and the variance is less than the variance threshold, perform shadow compensation on the shaded area; otherwise, perform shadow matching on the shaded area;

[0025] The so-called shadow correction specifically means: using the pixel values inside and outside the shaded area, establishing a shadow correction function, and mapping the pixel values inside the shaded area to the shadow-free state;

[0026] The calculation formula of the shadow correction function is: where, and are the pixel values of the k-th shaded pixel before and after shadow correction in the a-th shadow connected component respectively, μ a and σ a are the mean and standard deviation of the pixel values of the shaded pixels within the shadow connected component, μ o and σ ois the mean and standard deviation of the pixel values of the neighboring pixels outside the shadow area, and (k) represents the k-th pixel;

[0027] The shadow compensation adopts the Poisson image editing method. By solving the Poisson equation and using the pixels outside the shadow area as boundary conditions, the shadow-free pixel values within the shadow area are estimated. The calculation formula of the Poisson image editing method is:

[0028]

[0029]

[0030] Where is the Laplace operator, is the boundary of the shadow area;

[0031] The shadow matching adopts the patch-based shadow matching method. The calculation formula of the patch-based shadow matching method is: Where P(k) and P(k') are the image patches centered on the k-th pixel and the k'-th pixel respectively, D(·) is the Euclidean distance between the image patches, Ω is the image spatial domain, represents finding the pixel (k') when the objective function takes the minimum value, and Ω\S a represents the part of the image spatial domain after removing the a-th shadow connected component;

[0032] The method for obtaining the image patch is: P(k) = {I(x + xi, y + yj)| -r ≤ xi, yj ≤ r}, where x and y represent the coordinate positions of the k-th pixel in the mapping image data, I(x + xi, y + yj) represents the pixel value at the coordinate position (x + xi, y + yj) in the mapping image data, xi and yj represent the coordinate offsets of the image patch relative to the k-th pixel, and r represents the half side length of the image patch.

[0033] The specific method for extracting the time-weighted feature map based on the shadow-free image data and the time weight function is:

[0034] Step S310: Construct a spatio-temporal feature extraction network, use the shadow-free image data as the input, and use the spatio-temporal feature map as the output F t (k);

[0035] Step S320: Define the time weight function, add the time weight function to the spatio-temporal feature extraction network, and apply the time weight to the spatio-temporal feature map to obtain the time-weighted feature map The time weight function selects the exponential decay function.

[0036] The specific method for converting the time-weighted feature map into a change detection image using the preset change probability threshold is:

[0037] Step S330: Set the change probability threshold δ, and convert the time-weighted feature map into a change region mask B t , when the value of the k-th pixel in the time-weighted feature map is greater than or equal to the change probability threshold, set its corresponding change region mask B t (k) to 1, otherwise, set the corresponding change region mask B t (k) to 0;

[0038] Step S340: For the k-th pixel, if the value of its change region mask is 1, it belongs to the change region, and assign the value of the pixel in the time-weighted feature map to the change probability C t (k) in the change detection image. Otherwise, assign the change probability of the pixel in the change detection image to 0;

[0039] Step S350: Output the change region mask B t and the change detection image C with change probabilities t .

[0040] The specific method of inputting the surveying and mapping image data into the constructed land cover classification task, outputting the land cover class probability map, and optimizing the change detection image based on the land cover class probability map is as follows:

[0041] Step S410: Establish a multi-task learning network, including a change detection task and a land cover classification task. The input of the change detection task is the change detection image, and the output is the optimized change detection image. The input of the land cover classification task is the surveying and mapping image data, and the output is the land cover class probability map;

[0042] The multi-task learning network consists of a shared feature extraction layer, a task-specific layer, a total loss function, and an optimization algorithm;

[0043] Step S420: The total loss function of the multi-task learning network is obtained by weighted summation of the change detection loss function and the land cover classification loss function. Define the change detection loss function as binary cross-entropy loss, and define the land cover classification loss function as cross-entropy loss;

[0044] Step S430: Optimize the change detection image with the land cover class probability map output by the multi-task learning network. For each change region, count the land cover classes therein, obtain the change characteristics of different land cover classes, and judge whether the change region is a real change according to the change characteristics. Update the change probability C t (k) of the false change region to 0 to obtain the optimized change detection image.

[0045] One aspect of the present application provides an AI-based GIS surveying and mapping data analysis and processing system, including:

[0046] A surveying and mapping data acquisition module is used to acquire original surveying and mapping image data and pre-process it to obtain surveying and mapping image data;

[0047] A shadow segmentation and removal module is used to perform shadow segmentation and binarization on the surveying and mapping image data to obtain a shadow mask map, and to remove shadows from the surveying and mapping image data based on the shadow mask map to obtain shadow-free image data;

[0048] A change detection conversion module is used to extract a time-weighted feature map based on the shadow-free image data and the time weight function, preset a change probability threshold, and convert the time-weighted feature map into a change detection image;

[0049] The change detection optimization module is used to input the surveying and mapping image data into the constructed object classification task, output the object category probability map, optimize the change detection image based on the object category probability map, and obtain the optimized change detection image.

[0050] Compared with the existing technology, the AI-based GIS surveying and mapping data analysis and processing system and method proposed in this application have the following advantages:

[0051] This application uses a deep learning model for shadow segmentation, automatically learns and extracts the features of shadow areas, improves the accuracy of shadow segmentation, adaptively selects different shadow removal strategies, adopts the optimal removal method for different types of shadow areas, improves the effect of shadow removal, effectively restores the original pixel value of the shadow area, obtains high-quality shadow-free image data, suppresses the interference of shadow areas in change detection, and improves detection accuracy.

[0052] This application constructs a spatiotemporal feature extraction network, takes shadow-free image data as input, extracts spatiotemporal feature maps, captures change information of surveying and mapping data in the time dimension, and assigns different weights to feature maps at different time points, highlighting the importance of recent changes, suppressing the interference of long-term changes, and improving the time sensitivity of change detection. Through the time weighting mechanism, the spatiotemporal characteristics of multi-phase image data are effectively integrated to enhance the ability to capture actual changes in land objects.

[0053] This application establishes a multi-task learning network to perform change detection and object classification tasks simultaneously, fully utilizes the complementary information between the two tasks, improves the accuracy of change detection, optimizes the change detection results using the object classification results, judges the authenticity of the changed area according to the change characteristics of different object categories, effectively removes false changes, and improves the reliability of change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A method flow chart of the AI-based GIS surveying and mapping data analysis and processing method provided for this application;

[0055] Figure 2 Flowchart of the method for obtaining shadow-free image data provided for this application;

[0056] Figure 3 Flowchart of the method for obtaining change detection images provided for this application;

[0057] Figure 4 Functional module diagram of the AI-based GIS mapping data analysis and processing system provided for this application. Detailed implementation manners

[0058] To better understand this application, various aspects of this application will be described in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are only descriptions of the exemplary embodiments of this application and do not limit the scope of this application in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0059] In the accompanying drawings, for ease of illustration, the sizes, dimensions, and shapes of the elements have been slightly adjusted. The drawings are only examples and are not drawn to an exact scale. As used herein, terms such as "substantially", "about", and similar terms are used as terms indicating approximation and not as terms indicating degree, and are intended to account for the inherent deviations in measured or calculated values that would be recognized by a person of ordinary skill in the art. Additionally, in this application, the order in which the steps are described does not necessarily represent the order in which these processes occur in actual operation, unless there is a clear other limitation or can be deduced from the context.

[0060] It should also be understood that expressions such as "including", "including having", "having", "containing", and / or "containing having" are open-ended rather than closed-ended expressions in this specification, which mean the presence of the stated features, elements, and / or components, but do not exclude the presence of one or more other features, elements, components, and / or their combinations. Furthermore, when an expression such as "at least one of..." appears after a list of listed features, it modifies the entire list of features rather than just an individual element in the list. Additionally, when describing the embodiments of this application, the use of "may" means "one or more embodiments of this application". And the term "exemplary" is intended to refer to an example or illustration.

[0061] Unless otherwise defined, all terms used herein (including engineering and scientific terms) shall have the same meaning as commonly understood by those of ordinary skill in the art to which this application belongs. It should also be understood that, unless otherwise clearly stated in this application, words defined in common dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art, and should not be interpreted in an idealized or overly formal sense.

[0062] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will describe this application in detail with reference to the drawings and in combination with the embodiments.

[0063] Embodiment 1

[0064] As Figure 1 shown, the AI-based GIS mapping data analysis and processing method provided by this application includes:

[0065] Step S100: Obtain the original mapping image data and preprocess it to obtain the mapping image data;

[0066] The specific method for obtaining the original mapping image data and preprocessing it to obtain the mapping image data is:

[0067] Step S110: Determine the time range and spatial range of the monitoring area;

[0068] Step S120: Obtain the original mapping image data within the time range and spatial range, and preprocess it to obtain the mapping image data.

[0069] The original mapping image data comes from different data sources, including remote sensing images, aerial images, and ground measurement data; the mapping image data is I t , where t is the timestamp;

[0070] The preprocessing includes: performing geometric correction and spatial registration on the original mapping image data to make it have a consistent spatial reference and pixel alignment; converting the data format of the original mapping image data into a standard format, and unifying the coordinate systems of different data sources into the target coordinate system.

[0071] The above step S100 provides a data basis for subsequent tasks such as shadow removal and change detection by obtaining the original mapping image data.

[0072] Step S200: Perform shadow segmentation and binarization on the mapping image data to obtain a shadow mask map, and remove the shadow from the mapping image data based on the shadow mask map to obtain shadow-free image data;

[0073] As Figure 2As shown in the figure, it is a flowchart of the method for obtaining shadow-free image data provided by this application;

[0074] The specific method for performing shadow segmentation and binarization on the surveying and mapping image data to obtain a shadow mask map, and then removing the shadow from the surveying and mapping image data based on the shadow mask map to obtain shadow-free image data is as follows:

[0075] Step S210: Select a deep learning model as the initial shadow segmentation model, use the surveying and mapping image data as the input, and use the shadow probability map as the output;

[0076] The shadow probability map refers to an image output by the shadow segmentation model and having the same size as the input surveying and mapping image data, where the value of each pixel represents the probability that the pixel belongs to the shadow;

[0077] Step S220: Obtain a shadow segmentation training set, where the shadow segmentation training set includes the surveying and mapping image data and its corresponding shadow annotation mask. Use the cross-entropy loss function to train the shadow segmentation model, minimize the cross-entropy loss function between the true value of the shadow annotation mask of the pixel and the predicted shadow probability of the pixel, optimize the model parameters, and when the value of the cross-entropy loss function converges, obtain the trained shadow segmentation model;

[0078] The calculation formula of the cross-entropy loss function is: where, L s is the cross-entropy loss function, K is the total number of pixels, k is the pixel index, M t (k) is the true value of the shadow annotation mask of the kth pixel, P t (k) is the predicted shadow probability of the kth pixel, and log(·) is the logarithmic function;

[0079] Step S230: Use the trained shadow segmentation model to perform shadow segmentation on the surveying and mapping image data I t and output the shadow probability map P t ;

[0080] Step S240: Preset a shadow probability threshold τ, binarize the shadow probability map, determine the pixels with a shadow probability greater than or equal to the shadow probability threshold as shadow pixels, set their shadow mask value equal to 1, determine the pixels with a shadow probability less than the shadow probability threshold as non-shadow pixels, set their shadow mask value equal to 0, and obtain the shadow mask map of the shadow probability map The shadow pixels form a shadow area;

[0081] The calculation formula of the shadow mask value is:

[0082] The value of the shadow probability threshold is set by those skilled in the art according to experience.

[0083] Step S250: Remove the shadows from the shaded areas in the surveying and mapping image data I t to obtain shadow-free image data

[0084] The pixel value of the shadow-free image data is expressed as: Where, is the pixel value of the k-th pixel in the shadow-free image data after shadow removal, φ(·) is the shadow removal function, and I t (k) is the pixel value of the k-th pixel in the surveying and mapping image data;

[0085] Removing the shadows from the shaded areas in the surveying and mapping image data I t means that when the shadow mask value of the k-th pixel in the shadow probability map is equal to 0, it indicates that the shadow probability of this pixel is less than the shadow probability threshold, so this pixel is judged not to belong to the shaded area, and thus its pixel value remains unchanged; otherwise, this pixel is judged to belong to the shaded area, and the original pixel value of the shaded area is restored through shadow removal.

[0086] The method for obtaining the shadow-free image data is:

[0087] Step S251: Perform connected component labeling on the shadow mask map, label the interconnected groups of shadow pixels as the same shadow connected component, and extract the attribute features for each shadow connected component S a The attribute features include the area A of the shadow connected component a , perimeter P a , shape complexity PAR a and image statistical features. The image statistical features include the mean μ of the pixel values of the shadow pixels within the shadow connected component a , variance

[0088] The area of the shadow connected component is the number of pixels in the shadow connected component, and A a represents the area of the a-th shadow connected component;

[0089] The perimeter of the shadow connected component is the number of boundary pixels in the shadow connected component, and P a represents the perimeter of the a-th shadow connected component;

[0090] The calculation formula for the shape complexity of the shadow connected component is: Where, PAR a represents the shape complexity of the a-th shadow connected component;

[0091] The method for marking the shadow connected regions adopts a two-pass scanning algorithm, including the first pass scanning and the second pass scanning. The specific method of the first pass scanning is as follows: Initialize the shadow mask image, mark the label values of all pixels as 0, traverse the shadow mask image, for each shadow pixel, check the label values of its upper and left neighboring pixels. If the neighboring pixel is a marked shadow pixel, mark the current shadow pixel with an equivalent label value; if there are multiple different label values among the neighboring pixels, record these label values as the same label value; if there is no marked shadow pixel among the neighboring pixels, mark the current shadow pixel with a new label value.

[0092] The specific method of the second pass scanning is as follows: Traverse the shadow mask image, for each non-zero label value, according to the record of its equivalent label value, replace it with the smallest equivalent label value.

[0093] In the shadow mask image after connected region marking, different label values correspond to different shadow connected regions;

[0094] Step S252: According to the attribute characteristics of the shadow connected regions, preset an area threshold, a complexity threshold, and a variance threshold, design a decision function, select a removal strategy for the shadow region according to the decision function, and execute the corresponding removal strategy for each shadow connected region to obtain shadow-free image data;

[0095] The decision rule of the decision function is: When the area of the shadow connected region is less than the area threshold and the shape complexity is less than the complexity threshold, perform shadow correction on the shadow region; when the area of the shadow connected region is greater than or equal to the area threshold and the variance is less than the variance threshold, perform shadow compensation on the shadow region; otherwise, perform shadow matching on the shadow region;

[0096] The values of the area threshold, the complexity threshold, and the variance threshold are set by those skilled in the art according to experience.

[0097] The so-called shadow correction specifically refers to: Using the radiation information inside and outside the shadow region, establish a shadow correction function, and map the pixel values in the shadow region to the shadow-free state;

[0098] The calculation formula of the shadow correction function is: Where, and are the pixel values of the k-th shadow pixel in the a-th shadow connected region before and after shadow correction respectively, μ a and σ a are the mean and standard deviation of the pixel values of the shadow pixels in the shadow connected region, μ o and σ ois the mean and standard deviation of the pixel values of the neighboring pixels outside the shadow area, where (k) represents the k-th pixel;

[0099] The shadow compensation specifically refers to: using the pixel information outside the shadow area to estimate the non-shadow pixel values inside the shadow area; specifically, the non-shadow pixel values inside the shadow area can be estimated by interpolation or filling methods. For example, using the Poisson image editing method, by solving the Poisson equation with the pixels outside the shadow area as the boundary conditions, the non-shadow pixel values inside the shadow area are estimated.

[0100] The calculation formula of the Poisson image editing method is:

[0101]

[0102] where is the Laplacian operator, is the boundary of the shadow area;

[0103] The shadow matching specifically refers to: finding a matching area outside the shadow area that is similar to the image information inside the shadow area, and filling the pixel values of the matching area into the shadow area by copying or synthesizing. For example, using the patch-based shadow matching method.

[0104] The calculation formula of the patch-based shadow matching method is: where P(k) and P(k') are image patches centered on the k-th pixel and the k'-th pixel respectively, D(·) is the Euclidean distance between image patches, Ω is the image spatial domain, represents finding the pixel (k') that minimizes the objective function, and Ω\S a represents the part of the image spatial domain after removing the a-th shadow connected component;

[0105] The image patch refers to a local influence area centered on a pixel; the method for obtaining the image patch is: P(k) = {I(x + xi, y + yj) | -r ≤ xi, yj ≤ r}, where x and y represent the coordinate positions of the k-th pixel in the surveying and mapping image data, I(x + xi, y + yj) represents the pixel value at the coordinate position (x + xi, y + yj) in the surveying and mapping image data, xi and yj represent the coordinate offsets of the image patch relative to the k-th pixel, and r represents the half side length of the image patch;

[0106] The half side length of the image patch is set by those skilled in the art according to experience.

[0107] The image spatial domain refers to the entire image of the surveying and mapping image data. In digital image processing, the image is regarded as a two-dimensional function, and the spatial coordinates of the pixels are the domain of the two-dimensional function, that is, the image spatial domain.

[0108] In the surveying and mapping image data, there may be multiple unconnected shadow regions, and their characteristics such as size, shape, and texture may vary. Dividing the shadow regions into shadow connected components can distinguish these different shadow regions, facilitating subsequent analysis and processing.

[0109] For each shadow connected component, extract its attribute features, input the attribute features of the shadow connected component into the decision function, and the decision function selects the optimal shadow removal method for each shadow connected component according to the attribute features. Through this connected-component-based adaptive shadow removal strategy, the optimal removal method is selected according to the characteristics of different shadow regions, improving the effect and accuracy of shadow removal. Compared with treating the entire shadow region as a whole, the connected-component-based processing method can better adapt to the diversity and complexity of shadow regions in surveying and mapping image data, achieving more intelligent and efficient shadow removal.

[0110] The above step S200 realizes accurate shadow segmentation through a deep learning model, adopts an adaptive shadow removal strategy, effectively removes the shadow effect, and obtains high-quality shadow-free image data. This step solves the interference caused by the shadow effect and lays a foundation for accurately distinguishing shadow changes and actual ground object changes.

[0111] Step S300: Based on the shadow-free image data and the time weight function, extract the time-weighted feature map, preset the change probability threshold, and convert the time-weighted feature map into a change detection image;

[0112] The specific method for extracting the time-weighted feature map based on the shadow-free image data and the time weight function is as follows:

[0113] Step S310: Construct a spatio-temporal feature extraction network, use the shadow-free image data as the input, and use the spatio-temporal feature map as the output F t (k);

[0114] The spatio-temporal feature extraction network selects ConvLSTM, which combines the CNN model and the LSTM model. The CNN model is responsible for extracting the spatial features of the shadow-free image data, and the LSTM model is responsible for capturing the change features of the time series. Input multi-temporal shadow-free image data, and through the forward propagation process of the spatio-temporal feature extraction network, output the spatio-temporal feature map;

[0115] The construction process of the spatio-temporal feature extraction network is as follows:

[0116] The spatio-temporal feature extraction network selects the ConvLSTM model, including an input layer, a ConvLSTM layer, a pooling layer, and a fully connected layer;

[0117] The input layer is used to receive multi-temporal shadow-free image data;

[0118] The ConvLSTM layer is composed of n ConvLSTM units. The input of the ConvLSTM layer is the shadow-free image data at the current timestamp t and the hidden state at the previous timestamp, and the output is the hidden state at the current timestamp. The hidden state contains the change information of the shadow-free image data in the time dimension;

[0119] The pooling layer performs spatial downsampling on the output of the ConvLSTM layer;

[0120] The fully connected layer transforms the output of the pooling layer through the fully connected layer to generate the final spatio-temporal feature map.

[0121] Collect multi-temporal shadow-free image data as training data, label the changed areas of the training data, generate a changed area mask as the training label, use the binary cross-entropy loss function to measure the difference between the predicted change probability and the true changed area mask training label, input the multi-temporal shadow-free image data into the ConvLSTM model, calculate the spatio-temporal feature map output through forward propagation, and calculate the binary cross-entropy loss function. Update the model parameters through the backpropagation algorithm. When the value of the binary cross-entropy loss function converges, the training is completed, and the trained ConvLSTM model is used to extract the spatio-temporal feature map of the shadow-free image data.

[0122] Step S320: Define a time weight function, add the time weight function to the spatio-temporal feature extraction network, and apply the time weight to the spatio-temporal feature map to obtain a time-weighted feature map The time weight function selects an exponential decay function;

[0123] The calculation formula for the time weight is: w(t) = exp(-λ × (T - t)); where λ is the decay factor, T is the latest timestamp, and w(t) is the time weight;

[0124] The time weight function is used to measure the importance of timestamp t for change detection. The farther the time point is from the latest timestamp, the smaller its weight. Incorporate the time weight information into the spatio-temporal feature map so that the feature contributions of different time points are different.

[0125] The decay factor is used to control the decay rate of the time weight and is set by those skilled in the art according to actual needs.

[0126] The calculation formula for the time-weighted feature map is: where F t (k) represents the value of the spatio-temporal feature map at the k-th pixel at timestamp t, Denote the value of the time-weighted feature map at timestamp \(t\) at the \(k\)-th pixel;

[0127] As Figure 3 shown, it is the flowchart of the method for obtaining the change detection image provided by this application;

[0128] The specific method of converting the time-weighted feature map into a change detection image by using the preset change probability threshold is:

[0129] Step S330: Set the change probability threshold \(\delta\) to convert the time-weighted feature map into a change region mask \(B\) t When the value of the \(k\)-th pixel in the time-weighted feature map is greater than or equal to the change probability threshold, set the corresponding change region mask to 1, otherwise, set the corresponding change region mask to 0;

[0130] The calculation formula of the change region mask is: where \(B\) t (k) represents the value of the change region mask of the \(k\)-th pixel in the change region at timestamp \(t\);

[0131] The value of the change probability threshold is set by those skilled in the art according to experience.

[0132] When the value of a pixel in the time-weighted feature map is greater than or equal to the change probability threshold, it means that the pixel belongs to the change region, otherwise, it means that it does not belong to the change region.

[0133] Step S340: For the \(k\)-th pixel, if the value of its change region mask is 1, it belongs to the change region, assign the value of the pixel in the time-weighted feature map to the change probability in the change detection image, otherwise, assign the change probability of the pixel in the change detection image to 0;

[0134] The calculation formula of the change detection image is: where \(C\) t (k) is the change probability of the \(k\)-th pixel in the change detection image at timestamp \(t\);

[0135] In the change detection image, only the pixels belonging to the change region retain their change probability values, and the change probability values of other pixels are set to 0 to highlight the change region and retain the change probability information for subsequent analysis.

[0136] Furthermore, extract the contour and area of the change region according to the change region mask; the contour information can be used to calculate the geometric features of the change region, and the area information can be used to evaluate the significance of the change, exclude small-area noises and false changes, and improve the reliability of change detection. To extract the contour and area of the change region according to the change region mask, first perform morphological processing on the change region mask, such as dilation and erosion. The dilation operation can expand the change region and fill small holes; the erosion operation can shrink the change region and eliminate small noises; then extract the contour and area of the change region. The morphological processing can be implemented using image processing libraries such as OpenCV.

[0137] Step S350: Output the change region mask B t and the change detection image C with change probability t .

[0138] In the above step S300, through the spatio-temporal feature extraction network and the time weighting mechanism, the spatio-temporal features of multi-temporal image data are fused, enhancing the change detection model's ability to capture actual ground object changes. This step effectively solves the problem that the shadows between ground objects change over time, improving the time sensitivity and accuracy of change detection.

[0139] Step S400: Input the surveying and mapping image data into the constructed ground object classification task, output the ground object category probability map, and optimize the change detection image based on the ground object category probability map to obtain the optimized change detection image;

[0140] The specific method of inputting the surveying and mapping image data into the constructed ground object classification task, outputting the ground object category probability map, and optimizing the change detection image based on the ground object category probability map to obtain the optimized change detection image is as follows:

[0141] Step S410: Establish a multi-task learning network, including a change detection task and a ground object classification task. The input of the change detection task is the change detection image, and the output is the optimized change detection image. The input of the ground object classification task is the surveying and mapping image data, and the output is the ground object category probability map. The multi-task learning network consists of a shared feature extraction layer, a task-specific layer, a total loss function, and an optimization algorithm;

[0142] The shared feature extraction layer is responsible for extracting shared feature representations from the input change detection images and mapping image data, and a convolutional neural network can be used for the shared feature extraction layer; the task-specific layers are designed as independent task-specific layers for the change detection task and the land cover classification task respectively. The task-specific layer corresponding to each task performs corresponding transformation and prediction on the shared feature representation output by the shared feature extraction layer. For the change detection task, the task-specific layer includes a convolutional layer and a fully connected layer to convert the shared feature representation into an optimized change detection image. For the land cover classification task, the task-specific layer includes a convolutional layer, a fully connected layer, and a softmax activation function to convert the shared feature representation into a land cover class probability map; the loss functions for the change detection task and the land cover classification task are binary cross-entropy loss and cross-entropy loss respectively. By performing weighted summation on the loss functions of the change detection task and the land cover classification task, the total loss function of the multi-task learning network is obtained; the multi-task learning network minimizes the loss function through an optimization function to update the model parameters of the multi-task learning network, and the optimization algorithm uses the Adam optimization algorithm.

[0143] Step S420: The total loss function of the multi-task learning network is obtained by weighted summation of the change detection loss function and the land cover classification loss function. The change detection loss function is defined as binary cross-entropy loss, and the land cover classification loss function is defined as cross-entropy loss;

[0144] The total loss function is: L = L b + α × L d , where α is a balance factor used to control the importance of the change detection task and the land cover classification task, L b is the change detection loss function, and L d is the land cover classification loss function;

[0145] The calculation formula for the change detection loss function is: where K is the total number of pixels, G t (k) is the true change detection label, is the change detection probability predicted by the change detection task;

[0146] The calculation formula for the land cover classification loss function is: where J is the number of land cover classes, Y t (k, j) represents the true label of the k-th pixel belonging to land cover class j in the mapping image data at timestamp t, taking values of 0 or 1, and S t (k, j) represents the predicted probability value of the k-th pixel belonging to land cover class j in the mapping image data at timestamp t, with a value range of [0, 1];

[0147] Step S430: Optimize the change detection image using the land cover class probability map output by the multi-task learning network. For each change area, count the land cover classes therein, obtain the change characteristics of different land cover classes, and determine whether the change area is a real change according to the change characteristics. Update the change probability C t (k) of the false change area to 0 to obtain the optimized change detection image.

[0148] Exemplarily, for the change in the vegetation area, judge whether it is a real change according to the change amplitude of the vegetation index. The calculation formula for the change amplitude of the vegetation index is: where t1 and t2 are the timestamps of the change detection images before and after the change, are the vegetation indices before and after the change respectively; if the change amplitude of the vegetation index is less than the preset change amplitude threshold, the corresponding change area is a false change, and the change probability of the corresponding change area on the change detection image needs to be updated to 0.

[0149] The above steps use the land cover classification results to optimize and improve the change detection results. The land cover classification information can help identify and screen real change areas, exclude false change areas, and improve the accuracy and reliability of change detection.

[0150] Embodiment 2

[0151] As Figure 4 shown, the AI-based GIS mapping data analysis and processing system provided by this application includes:

[0152] A mapping data acquisition module for acquiring original mapping image data and preprocessing it to obtain mapping image data;

[0153] A shadow segmentation and removal module for performing shadow segmentation and binarization on the mapping image data to obtain a shadow mask map, and removing the shadow from the mapping image data based on the shadow mask map to obtain shadow-free image data;

[0154] A change detection conversion module for extracting a time-weighted feature map based on the shadow-free image data and a time weight function, presetting a change probability threshold, and converting the time-weighted feature map into a change detection image;

[0155] A change detection optimization module for inputting the mapping image data into the constructed land cover classification task, outputting a land cover class probability map, and optimizing the change detection image based on the land cover class probability map to obtain an optimized change detection image.

[0156] In addition, parts of the above technical solutions provided in the embodiments of this application that are the same as the corresponding technical solutions in the prior art in terms of implementation principles are not described in detail to avoid excessive elaboration.

[0157] The specific embodiments described above further elaborate on the object, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only the specific embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An AI-based GIS mapping data analysis and processing method, characterized in that, Including: Obtain the original surveying and mapping image data and preprocess it to obtain the surveying and mapping image data; Perform shadow segmentation and binarization on the surveying and mapping image data to obtain a shadow mask map, and remove the shadow from the surveying and mapping image data based on the shadow mask map to obtain shadow-free image data; Extract a time-weighted feature map based on the shadow-free image data and a time weight function, preset a change probability threshold, and convert the time-weighted feature map into a change detection image; Input the surveying and mapping image data into the constructed ground object classification task, output a ground object category probability map, and optimize the change detection image based on the ground object category probability map to obtain an optimized change detection image.

2. The AI-based GIS mapping data analysis and processing method according to claim 1, wherein, The specific method for performing shadow segmentation and binarization on the surveying and mapping image data to obtain a shadow mask map, and removing the shadow from the surveying and mapping image data based on the shadow mask map to obtain shadow-free image data is as follows: Select a deep learning model as the initial shadow segmentation model, use the surveying and mapping image data as the input, and use the shadow probability map as the output; Obtain a shadow segmentation training set, where the shadow segmentation training set includes the surveying and mapping image data and its corresponding shadow annotation mask, use the cross-entropy loss function to train the shadow segmentation model, minimize the cross-entropy loss function between the true value of the shadow annotation mask of the pixel and the predicted shadow probability of the pixel, optimize the model parameters, and when the value of the cross-entropy loss function converges, obtain a trained shadow segmentation model; Use the trained shadow segmentation model to segment the surveying and mapping image data I t to perform shadow segmentation and output the shadow probability map P t ; Preset a shadow probability threshold τ, binarize the shadow probability map, determine pixels with a shadow probability greater than or equal to the shadow probability threshold as shadow pixels, and set their shadow mask values equal to 1, determine pixels with a shadow probability less than the shadow probability threshold as non-shadow pixels, and set their shadow mask values equal to 0, to obtain the shadow mask map of the shadow probability map The shadow pixels constitute a shadow area; Remove the shadows in the surveying and mapping image data I according to the shadow mask image t to obtain shadow-free image data 3. The AI-based GIS mapping data analysis and processing method according to claim 2, characterized in that, The shadow removal of the shadow area in the mapping image data I according to the shadow mask image means that: when the shadow mask value of the k-th pixel in the shadow probability map t is equal to 0, it indicates that the shadow probability of this pixel is less than the shadow probability threshold, so this pixel is judged not to belong to the shadow area, and thus its pixel value remains unchanged; When the shadow mask value of the k-th pixel in the shadow probability map is equal to 0, it means that the shadow probability of this pixel is less than the shadow probability threshold, so this pixel is judged not to belong to the shadow area, and thus its pixel value remains unchanged; Otherwise, the pixel is judged to belong to the shadow area, and the original pixel value of the shadow area is restored by shadow removal.

4. The AI-based GIS mapping data analysis and processing method according to claim 3, wherein, The method for obtaining the shadow-free image data is as follows: Perform connected component labeling on the shadow mask image, label the interconnected groups of shadow pixels as the same shadow connected component, and for each shadow connected component S a Extract attribute features, where the attribute features include the area A of the shadow connected component a , perimeter P a , shape complexity PAR a and image statistical features, where the image statistical features include the mean μ of the pixel values of the shadow pixels within the shadow connected component a , variance σ a 2 ; According to the attribute characteristics of the shadow connected component, preset an area threshold, a complexity threshold, and a variance threshold, design a decision function, select a removal strategy for the shadow area according to the decision function, and execute the corresponding removal strategy for each shadow connected component to obtain shadow-free image data.

5. The AI-based GIS mapping data analysis and processing method according to claim 4, wherein, The decision rule of the decision function is: when the area of the shadow connected component is less than the area threshold and the shape complexity is less than the complexity threshold, perform shadow correction on the shadow area; when the area of the shadow connected component is greater than or equal to the area threshold and the variance is less than the variance threshold, perform shadow compensation on the shadow area; otherwise, perform shadow matching on the shadow area.

6. The AI-based GIS mapping data analysis and processing method according to claim 5, characterized in that The specific method for extracting a time-weighted feature map based on the shadow-free image data and a time weight function is as follows: Construct a spatio-temporal feature extraction network, taking the shadowless image data as input and the spatio-temporal feature map as output F t (k); Define a time weight function, add the time weight function to the spatio-temporal feature extraction network, apply the time weight to the spatio-temporal feature map, and obtain a time-weighted feature map The time weight function selects an exponential decay function.

7. The AI-based GIS mapping data analysis and processing method according to claim 6, wherein The specific method for presetting the change probability threshold and converting the time-weighted feature map into a change detection image is as follows: Set the change probability threshold δ, and convert the time-weighted feature map into a change region mask B t , when the value of the k-th pixel in the time-weighted feature map is greater than or equal to the change probability threshold, then set its corresponding change region mask B t (k) to 1, otherwise, set the corresponding change region mask B t (k) to 0; For the k-th pixel, if the value of its change region mask is 1, it belongs to the change region, and the value of it in the time-weighted feature map is assigned to the change probability C t (k) in the change detection image. Otherwise, the change probability of the pixel in the change detection image is assigned a value of 0; Output change region mask B t and change detection image C with change probability t .

8. The AI-based GIS mapping data analysis and processing method according to claim 7, characterized in that The specific method for inputting the surveying and mapping image data into the constructed ground object classification task, outputting a ground object category probability map, and optimizing the change detection image based on the ground object category probability map to obtain an optimized change detection image is as follows: Establish a multi-task learning network, including a change detection task and a ground object classification task. The input of the change detection task is the change detection image, and the output is the optimized change detection image. The input of the ground object classification task is the surveying and mapping image data, and the output is the ground object category probability map; The total loss function of the multi-task learning network is obtained by weighted summation of the change detection loss function and the ground object classification loss function. Define the change detection loss function as the binary cross-entropy loss, and define the ground object classification loss function as the cross-entropy loss; The land cover class probability map output by the multi-task learning network optimizes the change detection image. For each change region, the land cover classes therein are counted to obtain the change characteristics of different land cover classes. Whether the change region is a real change is judged according to the change characteristics, and the change probability C t (k) is updated to 0 to obtain the optimized change detection image.

9. An AI-based GIS mapping data analysis and processing system for implementing the AI-based GIS mapping data analysis and processing method according to any one of claims 1-8, characterized in that Including: A surveying and mapping data acquisition module, which is used to acquire original surveying and mapping image data and preprocess it to obtain surveying and mapping image data; A shadow segmentation and removal module, which is used to perform shadow segmentation and binarization on the surveying and mapping image data to obtain a shadow mask map, and remove the shadow from the surveying and mapping image data based on the shadow mask map to obtain shadow-free image data; A change detection conversion module, which is used to extract a time-weighted feature map based on the shadow-free image data and a time weight function, preset a change probability threshold, and convert the time-weighted feature map into a change detection image; A change detection optimization module, which is used to input the surveying and mapping image data into a constructed ground object classification task, output a ground object category probability map, and optimize the change detection image based on the ground object category probability map to obtain an optimized change detection image.

Images