Spatial data AI intelligent analysis system

Abstract

The invention provides a spatial data AI intelligent analysis system, and the system comprises an AI earth module which is used for uploading multi-period remote sensing image data through a platform interface, carrying out the change detection, extracting the similarities and differences of different time phase images in the same region, and obtaining a change pattern spot; the model construction module is used for constructing an initial detection model by using an automatic change detection algorithm and a deep learning model based on a coding-decoding structure, and performing a change detection task of the change pattern spots, model training and model prediction to obtain an initial detection result; the post-processing module is used for analyzing and detecting an initial detection result by using an unsupervised change detection algorithm to finish final detection; and the task management module is used for creating and executing a plurality of tasks. According to the invention, pixel-level rapid change detection can be carried out on the input multi-stage remote sensing image data, the extraction precision is improved, and more accurate change area information is obtained.

Description

Technical Field

[0001] The present invention relates to the field of spatial data analysis technology, and in particular to a spatial data AI intelligent analysis system. Background Art

[0002] Remote sensing data, also known as spatial data, is information about the earth's surface obtained from a long distance (usually from satellites, aircraft or drones) through remote sensing technology. Remote sensing technology uses sensors to capture signals reflected or emitted by different bands in the electromagnetic spectrum (such as visible light, infrared, microwaves, etc.) to generate images and data about the characteristics of the earth's surface. Remote sensing data can be used for environmental monitoring, agricultural management, urban planning, disaster response, climate research, etc.

[0003] Existing remote sensing data (spatial data) analysis technologies include image processing, deep learning, machine learning and other technologies, but the existing technologies still have the following problems: limited data processing capabilities and low efficiency when processing large-scale, high-resolution data. The model generalization ability is insufficient, and many models perform inconsistently in different environments or data sets, and lack generalization capabilities. Data labeling is difficult, and high-quality training data sets require a lot of manual labeling, which is time-consuming and expensive. Insufficient real-time performance, many analysis technologies cannot process data in real time, limiting their application in emergency response. Complexity and interpretability, although deep learning models are powerful, they are often regarded as "black boxes" and lack interpretability. Therefore, the present invention proposes a spatial data AI intelligent analysis system to solve the above problems. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a spatial data AI intelligent analysis system that can perform rapid pixel-level change detection on input multi-period remote sensing image data (spatial data), improve extraction accuracy, and obtain more accurate change area information.

[0005] To achieve the above-mentioned object, the present invention provides the following solutions: a spatial data AI intelligent analysis system, comprising an interconnected platform interface, an AI earth module, a model building module, a post-processing module and a task management module;

[0006] The AI ​​Earth module is used to upload multi-period remote sensing image data using the platform interface, and perform change detection on remote sensing images through layer management operations, data upload operations, online interpretation operations, and online editing operations, extracting the similarities and differences of images of different phases in the same area in the remote sensing image, and obtaining change spots;

[0007] The model building module is used to build an initial inspection model using an automatic change detection algorithm and a deep learning model based on an encoding-decoding structure, use the initial inspection model to perform the change detection task of the change pattern, and use the constructed test data set to perform model training and model prediction to obtain an initial inspection result;

[0008] The post-processing module is used to analyze and detect the initial inspection results using an unsupervised change detection algorithm to complete the final inspection;

[0009] The task management module is used to integrate the detection process of the remote sensing image into one task, and to create and execute multiple tasks on the platform interface.

[0010] Optionally, the AI ​​earth module includes:

[0011] A layer management unit, used to manage raster layers and vector layers in the uploaded remote sensing image using the platform interface, the layer management unit includes a background layer, a raster layer, a vector layer and a layer menu panel;

[0012] A data uploading unit, used to add new vector data and raster data using the platform interface, the data uploading unit includes a vector uploading function and a raster uploading function;

[0013] An online interpretation unit is used to draw AOI on the remote sensing image based on the operation results of the layer management unit and the data uploading unit to obtain the target area, perform change detection on the target area, extract the similarities and differences of different phases of the same area in the remote sensing image, and obtain the change spots. The online interpretation unit includes AOI drawing function, ground object extraction function, target recognition function and thematic detection function;

[0014] An online editing unit is used to edit and modify the vector layer and the vector data using the platform interface. The online editing unit includes an element adding option, an element selecting option, an element moving option, an element identifying option, an element pasting option, and an element deleting option.

[0015] Optionally, the thematic detection function includes a multi-source change detection task, an iterative multivariate change detection task, a principal component analysis change detection task, a change vector analysis change detection task, an object-based change detection task and a deep learning change detection task.

[0016] Optionally, the model building module includes:

[0017] The feature extraction unit is used to learn and extract the features of the dual-phase remote sensing image based on the twin network structure using ResNet50 and residual network to obtain a difference feature map;

[0018] A feature fusion unit is used to expand the size of the difference feature map by using a feature pyramid network to obtain a feature enlarged map, connect the feature enlarged map horizontally to the corresponding encoding layer, and fuse the features of the bottom layer and the high layer of the feature enlarged map to obtain a multi-level feature map with different sizes, and input the multi-level feature map into a decoder for multiple upsampling to obtain a new feature map with detail information and semantic features;

[0019] A deep supervision unit is used to introduce an intermediate loss function in the hidden layer for supervision according to the output of the encoding layer of each scale, and add an auxiliary classifier to supervise the backbone network, and then perform multiple upsampling on the new feature map according to the encoding structure of each scale, and use a Sigmoid classifier to map the new feature map that has completed multiple upsampling into a binary map, compare the binary map with the label map and calculate the loss;

[0020] A model training unit, used to construct a dynamic weighted cross entropy loss function, and use the test data set to iterate the training of the initial inspection model, and randomly select data enhancements for combination during the training iteration process;

[0021] The model prediction unit is used to input remote sensing image data and perform enhancement operations on the input data to generate multiple enhanced samples, input the enhanced samples into the initial inspection model after iterative training for prediction to obtain initial prediction results, and average or vote on the initial prediction results to obtain final prediction results; wherein the enhancement operations include random scaling and random flipping.

[0022] Optionally, the post-processing module includes:

[0023] An invariant feature analysis unit, used to process the initial inspection result by using deviation calculation, position correction and invariance measurement methods, eliminate pseudo changes caused by geometric displacement in the initial inspection result, and obtain a secondary inspection result;

[0024] The morphological operation unit is used to fill holes in the spots in the secondary detection results using morphological opening and closing operations, delete small spots according to a preset pixel threshold, and then use morphological dilation operations and erosion operations with 7 pixels to remove burrs and sharp grooves at the boundaries of the spots in the secondary detection results, and use the DP algorithm to smooth the boundaries of the spots in the secondary detection results, complete boundary optimization, and obtain the final detection result.

[0025] Optionally, the unsupervised change detection algorithm includes a multivariate change detection method, a principal component analysis method, a change vector analysis method, an object-based change spot analysis method and a segmented object-based constrained post-processing method; wherein the object-based change spot analysis method includes an image object overlay comparison method, an image object direct comparison method and a multi-phase image object comparison method, and the segmented object-based constrained post-processing method is used to analyze the proportion of changed pixels in each segmentation unit in the spot, set a threshold to filter out pseudo-changes, and introduce segmented objects for comparison with the final detection results to complete boundary optimization.

[0026] Optionally, the task management module includes:

[0027] Create a new task unit to input the task name, interpretation type and image dataset; the interpretation type includes ground objects, targets and topics, and the ground objects include deep learning change detection options, change vector analysis change detection options, iterative multivariate change detection options, object-based change detection options, principal component analysis change detection options, multivariate change detection functions and change detection accuracy evaluation options;

[0028] Task list unit, used to display list information and perform list operations;

[0029] List sorting unit, used to sort the list in ascending or descending order;

[0030] Task search unit, used to search the list by keyword.

[0031] The present invention discloses the following technical effects by providing a spatial data AI intelligent analysis system:

[0032] 1. The present invention can be deployed at the BS end and is easy to use.

[0033] 2. The present invention integrates automatic change detection algorithm and unsupervised change detection algorithm, and performs rapid pixel-level change detection by inputting multi-period remote sensing image data, thereby realizing the extraction of similarities and differences between images of different phases in the same area; at the same time, the extracted change spots can be published and displayed, and the change spots can be statistically analyzed.

[0034] 3. The present invention supports the setting of relevant parameters such as regional division restrictions and change fuzziness, and can achieve an extraction accuracy of more than 90%. It can serve application scenarios with changing demands such as satellite image law enforcement, urban construction, ecological protection and monitoring, and mine reclamation.

[0035] 4. The present invention performs double detection by establishing a model and post-processing, which can solve the problems of image holes, small image spots and uneven boundaries, so that the output results of change detection are more consistent with the edges of real changes, and more accurate change area information is obtained.

[0036] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 A schematic diagram of a system architecture provided by an embodiment of the present invention;

[0039] Figure 2 This is an operation flow chart of the platform interface provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] like Figure 1-2 As shown, the present invention provides a spatial data AI intelligent analysis system, comprising:

[0043] 1. AI Earth Module

[0044] It is used to upload multi-period remote sensing image data using the platform interface, and perform change detection on remote sensing images through layer management operations, data upload operations, online interpretation operations, and online editing operations, extract the similarities and differences of different phases in the same area of ​​remote sensing images, and obtain change spots. The AI ​​Earth module includes:

[0045] 1.1 Layer Management Unit

[0046] It is used to manage the raster layers and vector layers in the uploaded remote sensing images using the platform interface. The layer management unit includes a background layer, a raster layer, a vector layer and a layer menu panel (which can be displayed and hidden).

[0047] Layer management is to browse, edit, delete and other operations on layers. Layer management includes raster layer management and vector layer management.

[0048] 1.11 Background layers mainly include: Mapbox satellite, OSM map, Tiandi map and other place name layers.

[0049] 1.12 Raster layer can browse the current raster layer. Click on a specific raster layer to browse the related graphics and attribute information of the layer.

[0050] 1) View raster properties: Click [Properties] to view raster properties, including name, range, height, width, data type, band, pixel size and other information;

[0051] 2) Layer top or bottom: Click [Layer top] or [Layer bottom] to move the layer to the top or bottom;

[0052] 3) Layer rename: Click [Rename] to modify the layer name;

[0053] 4) Layer deletion: Click [Delete] to delete the layer.

[0054] 1.13 Vector layer You can browse the current vector layer. Click on a specific vector layer to browse the related graphics and attribute information of the layer.

[0055] 1) View statistical information: Click [Statistics] to view the number of elements in the vector and the total area of ​​the elements;

[0056] 2) Vector export: Click [Export] to export the saved vector to your local computer;

[0057] 3) View vector information: Click [Properties] to view vector properties, including: name, file type, geometry type, coordinate reference system, range, unit, and number of elements;

[0058] 4) Layer on top or bottom: Same as the operation of raster layer, click [Layer on top] or [Layer on bottom] to move the layer to the top or bottom;

[0059] 5) Layer rename: Same as the raster layer operation, click [Rename] to modify the layer name.

[0060] 1.2 Data upload unit

[0061] It is used to add new vector data and raster data using the platform interface. The data upload unit includes vector upload function and raster upload function. Adding data means uploading data to the platform. Data upload includes vector upload and raster upload (new data). There is also the upload of original data, including raster data set upload and vector data set upload.

[0062] 1.21 Newly added data upload

[0063] 1.211 Raster Upload

[0064] You can add new rasters by uploading them. The formats supported for uploaded rasters include: geotiff, tif, and the size cannot exceed 2G. The uploaded rasters can be viewed in the raster layer of the layer management.

[0065] On the grid upload page, click [Click to upload], a local folder panel pops up, click the grid to be uploaded, and click [OK] to complete the data upload. During the upload process, the uploaded file name, size, and upload progress will pop up in the lower left corner of the system.

[0066] 1.212 Vector Upload

[0067] You can add new vectors by uploading vectors. The uploaded vectors only support uploading shp zip compressed files, and the recommended size is no more than 2GB. The uploaded vectors can be viewed in the vector layer of layer management.

[0068] First, select the type to be uploaded as feature layer or target layer; then, the same as the raster upload operation, click [Click to Upload], a local folder panel will pop up, click the vector to be uploaded, and click [OK] to complete the data upload.

[0069] 1.22 Upload the original dataset

[0070] 1.221 Raster dataset upload

[0071] Raster dataset upload includes three functions: client local upload, server import, and display upload status:

[0072] Local upload: Click the [Client Upload] button to upload. It supports uploading local images of the user. It supports tif and tiff image formats. You can drag and drop to upload or click to upload. Drag and drop to upload means directly dragging the file to the dialog box, or click the [Click to Upload] button to select from the folder. After uploading, you can see the file name and file size of the uploaded file. Click OK to upload.

[0073] Server import: Click the [Server Import] button. Server import supports importing images from the server into the system. You can browse the files in the server and check the data to be uploaded.

[0074] Upload status: After confirming the upload, the file upload dialog box will pop up, showing the file name, size and upload status.

[0075] 1.222 Create a new folder

[0076] Click the [New Folder] button and enter a folder name to create a new folder. You can rename or delete it.

[0077] 1.223 File List

[0078] List information: The list information of the file list includes file name, data size, data format, geographic information, creation time and operation.

[0079] List operations: divided into file operations and folder operations. The operations on folders include renaming, and the operations on files include details, renaming, and uploading. Click the [Details] button to pop up the raster properties. On the left is the image thumbnail, and on the right is the raster attribute information, including the current range, current coordinate reference system (CRS), unit, width, starting point, bit depth, pixel size, name, number of bands, dimension, and height information. Click the [Rename] button, enter a new file name, and click the [OK] button to rename. Click the [Upload] button to upload the raster image to the AI ​​Earth display.

[0080] List sorting: You can sort files and folders by file name, data size, data format, geographic information, and creation time. Click the [Triangle] button next to the creation time to sort the list in ascending or descending order according to the creation time. Click the [More] button to pop up a drop-down menu containing three operations: download, move, and delete. Check the data you want to download and click the [Download] button to download the image. Check the data you want to move, click the [Move] button, select the file directory to move to, and click [OK] to complete the move. Check the data you want to delete and click the [Delete] button to delete the image.

[0081] 1.224 Search

[0082] Click the [Enter Keywords] button, enter the keywords to search, and press the [Enter] key to search in the grid list. The operations that can be performed on the results of the keyword search include details, rename, and upload.

[0083] 1.23 Upload vector dataset

[0084] 1.231 vector dataset upload module includes three functions: client local upload, server import and display upload status:

[0085] Local upload: Click the [Client Upload] button to upload. It supports uploading local vectors of the user and zip compressed packages in shp vector format. You can drag and drop to upload or click to upload. Drag and drop means directly dragging the file to the dialog box, or click the [Click to Upload] button to select from the folder. After selecting the vector file, the file name and file size of the uploaded file will be displayed. Click the [OK] button to upload.

[0086] Server import: Click the [Server Import] button. Server import supports importing vectors from the server into iBrain. You can browse the folders in the server and check the data to be uploaded.

[0087] Upload status: After confirming the upload, a file upload dialog box will pop up, which can display the file name, size and upload status.

[0088] 1.232 Create a new folder

[0089] Including list information, list operations, list sorting, etc.

[0090] 1.233 Results List

[0091] Including list information, list operations, list sorting, list search, deletion, etc.

[0092] 1.3 Online interpretation unit

[0093] It is used to draw AOI on the remote sensing image based on the operation results of the layer management unit and the data upload unit, obtain the target area, perform change detection on the target area, extract the similarities and differences between images of different phases in the same area in the remote sensing image, and obtain change spots. The online interpretation unit includes AOI drawing function, ground object extraction function, target recognition function and thematic detection function; the thematic detection function includes multi-source change detection task, iterative multivariate change detection task, principal component analysis change detection task, change vector analysis change detection task, object-based change detection task and deep learning change detection task.

[0094] Online interpretation is to obtain change information from remote sensing images based on existing layers and uploaded data. First, AOI drawing is to use a rectangular box to draw the target area for completing different tasks. Among them, the target area is drawn by left-clicking the mouse, dragging the matrix box, and left-clicking the mouse again after completing the area selection. Secondly, after completing the AOI drawing, different tasks will pop up on the right side of the system, including: object extraction, target identification, and thematic monitoring. Click on thematic monitoring, select the corresponding change detection algorithm, and click [Online Calculation] on the right toolbar to classify and intelligently identify roads, vegetation, water systems, aircraft, ships, buildings, etc. in the image base map. Click [Reset] to re-identify.

[0095] 1.31 Multivariate Change Detection

[0096] Draw a rectangular box. If you are not satisfied with the drawn rectangular box, click [Delete] in the upper right corner of the rectangular box to redraw the rectangular box. After completing the AOI drawing, click [Multivariate Change Detection], and the [Task Name] text box will automatically enter "Multivariate Change Detection Interpretation Task". Click [Online Calculation]. After clicking [Online Calculation], a progress bar will appear. When it is loaded to 100%, the multivariate change detection task is completed and the extraction results are displayed. After the extraction is completed, the extracted vector map and the extracted result statistics will appear, including: the number of elements and the total area of ​​the elements. Click [Save] in the upper right corner of the rectangular box, enter the layer name, and click [OK] to save the extracted vector map and view it in the vector layer.

[0097] 1.32 Iterative Multivariate Change Detection

[0098] Draw a rectangular box. If you are not satisfied with the drawn rectangular box, click [Delete] in the upper right corner of the rectangular box to redraw the rectangular box. After completing the AOI drawing, click [Iterative Multivariate Change Detection], and the [Task Name] text box will automatically enter "Iterative Multivariate Change Detection Interpretation Task". Click [Online Calculation]. After clicking [Online Calculation], a progress bar will appear. When it is loaded to 100%, the iterative multivariate change detection task is completed and the extraction results are displayed. After the extraction is completed, the extracted vector map and the extracted result statistics will appear, including: the number of elements and the total area of ​​the elements. Click [Save] in the upper right corner of the rectangular box, enter the layer name, and click [OK] to save the extracted vector map and view it in the vector layer.

[0099] 1.33 Composition Analysis Change Detection

[0100] Draw a rectangular box. If you are not satisfied with the drawn rectangular box, click [Delete] in the upper right corner of the rectangular box to redraw the rectangular box. After completing the AOI drawing, click [Principal Component Analysis Change Detection], and the [Task Name] text box will automatically enter "Principal Component Analysis Change Detection Interpretation Task". Click [Online Calculation]; after clicking [Online Calculation], a progress bar will appear. When it is loaded to 100%, the principal component analysis change detection task is completed and the extraction results are displayed; after the extraction is completed, the extracted vector map and the extracted result statistics will appear, including: number of elements and total area of ​​elements. Click [Save] in the upper right corner of the rectangular box, enter the layer name, and click [OK] to save the extracted vector map and view it in the vector layer.

[0101] 1.34 Change Detection Using Change Vector Analysis

[0102] Draw a rectangular box. If you are not satisfied with the drawn rectangular box, click [Delete] in the upper right corner of the rectangular box to redraw the rectangular box. After completing the AOI drawing, click [Change Vector Analysis Change Detection], and the [Task Name] text box will automatically enter "Change Vector Analysis Change Detection Interpretation Task", and click [Online Calculation]; after clicking [Online Calculation], a progress bar will appear. When it is loaded to 100%, the change vector analysis change detection task is completed and the extraction results are displayed; after the extraction is completed, the extracted vector map and the extracted result statistics will appear, including: number of elements and total area of ​​elements. Click [Save] in the upper right corner of the rectangular box, enter the layer name, and click [OK] to save the extracted vector map and view it in the vector layer.

[0103] 1.35 Object-based change detection

[0104] Draw a rectangular box. If you are not satisfied with the drawn rectangular box, click [Delete] in the upper right corner of the rectangular box to redraw the rectangular box. After completing the AOI drawing, click [Object-based Change Detection], and the [Task Name] text box will automatically enter "Object-based Change Detection Interpretation Task". Click [Online Calculation]; after clicking [Online Calculation], a progress bar will appear. When it is loaded to 100%, the object-based change detection task is completed and the extraction results are displayed; after the extraction is completed, the extracted vector map and the extracted result statistics will appear, including: number of elements and total area of ​​elements. Click [Save] in the upper right corner of the rectangular box, enter the layer name, and click [OK] to save the extracted vector map and view it in the vector layer.

[0105] 1.36 Deep Learning Change Detection

[0106] Draw a rectangular box. If you are not satisfied with the drawn rectangular box, click [Delete] in the upper right corner of the rectangular box to redraw the rectangular box. After completing the AOI drawing, click [Deep Learning Change Detection], and the [Task Name] text box will automatically enter "Deep Learning Change Detection Interpretation Task". Click [Online Calculation]; after clicking [Online Calculation], a progress bar will appear. When it is loaded to 100%, the deep learning change detection task is completed and the extraction results are displayed; after the extraction is completed, the extracted vector map and the extracted result statistics will appear, including: number of elements and total area of ​​elements. Click [Save] in the upper right corner of the rectangular box, enter the layer name, and click [OK] to save the extracted vector map and view it in the vector layer.

[0107] 1.4 Online Editing Unit

[0108] Used to edit and modify the vector layer and the vector data using the platform interface, the online editing unit includes an element adding option, an element selecting option, an element moving option, an element identifying option, an element pasting option, and an element deleting option.

[0109] 1.41 elements

[0110] Add new elements, including drawing polygons, drawing rectangles, and drawing circles. Select the [Draw Polygon] tool and draw along the outline of the object to draw a new element. After drawing, the element's attribute panel will appear on the left, including: category, detailed category, content description, element area, center longitude and latitude information. Click [Save] to complete the addition of the new element.

[0111] To select an element, move the mouse to the element to be selected and click the left button of the mouse, and the corresponding element will be selected.

[0112] To split a feature, you can use [Cut Spot] and [Split Spot] to split a whole feature. Select the [Cut Spot] tool and use the icon to draw a closed area to be cut to complete the feature cutting, which will remove the drawn area and split a feature into two. Select the [Split Spot] tool and use the icon to draw a closed area on the feature to split to complete the feature splitting, splitting a feature into three.

[0113] Moving an element is to move an element to a different location. Select the element, right-click the mouse, the element editing menu pops up, select the [Move] tool, and drag it to another location to complete the element move.

[0114] Element rotation is to rotate around the center point of the element. Select the element, right-click the mouse, the element editing menu pops up, select the [Rotate] tool, drag it to rotate, and you can complete the element rotation at any angle.

[0115] Feature flipping includes flipping along the short axis, flipping along the long axis, flipping along the X axis, and flipping along the Y axis. Select the feature, right-click the mouse, the feature editing menu pops up, select the required flip tool, and you can complete the feature flipping.

[0116] To circularize an element, select the element, right-click the mouse, the element editing menu will pop up, select the [Circularize] tool, and you can complete the circularization of the element.

[0117] To copy features, select the feature, right-click the mouse, the feature editing menu pops up, select the [Copy] tool, and the feature copying can be completed.

[0118] To delete an element, select the element, right-click the mouse, and the element editing menu will pop up. Select the [Delete] tool to complete the element deletion.

[0119] 1.42 Quick Tools

[0120] Rolling Shutter: Rolling Shutter allows you to view two images at the same time and compare them. First, click [Rolling Shutter] in the AI ​​Earth Tools menu, drag the two images you want to view in the raster layer into the image viewing area, and slide the middle button to slide left and right to view the difference between the two images.

[0121] Place name layer: Select [Place name layer] in the bottom bar to display all place names on the layer for quick positioning.

[0122] Measurement: The measurement tools include area measurement and length measurement. Select [Area] or [Length] in the bottom bar, drag the icon to select the target area to measure the area or length of the selected area, and display the measurement results, as shown in the figure below. Click [Delete] next to the measurement result to delete the selected area.

[0123] Eagle Eye: Select [Eagle Eye] in the bottom bar to see the overview and current location.

[0124] 2. Model building module

[0125] It is used to construct a preliminary inspection model using an automatic change detection algorithm and a deep learning model based on an encoding-decoding structure, use the preliminary inspection model to perform the change detection task of the change pattern, and use the constructed test data set to perform model training and model prediction to obtain preliminary inspection results. The model construction module includes:

[0126] 2.1 Feature Extraction Unit

[0127] It is used to learn and extract the features of dual-phase remote sensing images based on the twin network structure, using ResNet50 and residual network to obtain difference feature maps.

[0128] 2.11 Model construction: A deep learning model with an encoding-decoding structure is adopted. First, the weight-shared twin neural network is used to extract the features of the dual-phase remote sensing images, and the feature difference method is used to obtain the difference feature map. Then, the feature pyramid is used to establish the connection between the multi-level feature maps. Finally, the decoder is used to upsample the fused features to obtain the final change detection result.

[0129] 2.12 Automatic change detection algorithm: The change spot extraction work is based on the high-resolution remote sensing images of the previous and subsequent periods to extract the suspected illegal change information. Its essence is a change detection task for a specific target. For this change detection task, a dedicated sample set and deep learning model need to be established.

[0130] 2.13 Twin Network Structure

[0131] In addition to learning the shallow texture and structural features of the changing objects, the convolutional network can also learn the high-dimensional abstract features of the changing area. The twin structure adopts a weight sharing mechanism, which can reduce the information loss during channel fusion and greatly reduce the number of parameters.

[0132] ResNet50 is used as the feature extraction layer of the twin structure. It is composed of multiple residual modules in cascade, and jump connections are added to the convolutional modules. During the forward propagation process, the input signal of the residual module can be directly transmitted from any low layer to the high layer, forming a natural identity mapping to ensure that excellent features can be learned in the deep network. At the same time, through the reverse derivation of the residual network, it can be seen that the residual structure can directly transfer the model loss to the shallow network, avoiding the problem of gradient disappearance.

[0133] The ResNet50 network is based on the Bottleneck structure, which consists of two Block units, namely Conv Block and Identity Block. The input and output dimensions of Conv Block are different, so they cannot be connected in series. Its function is to change the dimension of the network; the input dimension and output dimension of Identity Block are the same, and they can be connected in series to deepen the network.

[0134] 2.2 Feature Fusion Unit

[0135] It is used to expand the size of the difference feature map by using a feature pyramid network to obtain a feature enlargement map, laterally connect the feature enlargement map to the corresponding encoding layer, and fuse the features of the bottom layer and the high layer of the feature enlargement map to obtain a multi-level feature map with different sizes, and input the multi-level feature map into a decoder for multiple upsampling to obtain a new feature map with detail information and semantic features.

[0136] During the feature extraction process of the residual network, feature maps of different sizes are generated. The bottom layer contains more detailed information of the image, and the deep layer contains more semantic feature information. The traditional convolutional neural network usually uses the top layer network to obtain the change map after multiple layers of convolution operations. Although the feature map contains rich information at this time, it is not conducive to the change detection of small volume targets due to its small size, reduced resolution and loss of spatial information. At the same time, it ignores the detailed information of feature maps of other layers, affecting the measurement accuracy. In order to solve this problem, the feature map pyramid network FPN is introduced.

[0137] FPN is a network proposed in 2017. It mainly solves the multi-scale problem in object detection. Through simple changes in network connections, it greatly improves the performance of small object detection without increasing the computational complexity of the original model. The principle is to upsample and gradually restore the feature map obtained by multiple convolution operations, so that the size of the feature map is increased while ensuring that there is no loss of high-level semantic information. Larger feature maps are used to detect small-volume targets, thereby improving the accuracy of the network's detection of changes in small-volume targets. In addition, FPN expands the size of the feature map in a top-to-bottom manner, and horizontally connects the corresponding coding layers to fuse the features of the bottom and high layers to obtain feature maps of different sizes, so that the feature map of the bottom layer also has the rich semantic information of the high-level feature map.

[0138] Decoder: The multi-level feature maps obtained after fusion are used as decoder input, and multiple upsampling is performed on them to obtain new feature maps with both detailed information and semantic features.

[0139] 2.3 Deep Supervision Unit

[0140] It is used to introduce an intermediate loss function in the hidden layer for supervision according to the output of the encoding layer at each scale, and add an auxiliary classifier to supervise the backbone network, and then perform multiple upsampling on the new feature map according to the encoding structure of each scale, and use the Sigmoid classifier to map the new feature map that has completed multiple upsampling into a binary map, compare the binary map with the label map and calculate the loss.

[0141] The deep supervision mechanism improves the directness and transparency of the hidden layer learning process by introducing an intermediate loss function supervision function for each hidden layer. The intermediate loss function can be regarded as an additional constraint on the network learning process. At the same time, it can solve the problems of gradient vanishing and slow convergence in deep neural network training. In the structure of this article, the deep supervision mechanism is added after the output of the encoding layer at each scale, and the backbone network is supervised by adding additional auxiliary classifiers. After the encoding structure at each scale, the available feature map is multiply upsampled, and then mapped to a binary map using a Sigmoid classifier, and compared with the label map to calculate the loss.

[0142] 2.4 Model Training Unit

[0143] Aiming at the problems of category imbalance, large differences in change patch areas, and uncertainty in edge labels in change detection model training, a dynamic weighted cross entropy loss function is constructed to improve model training and prediction accuracy.

[0144] Training strategy: The initial inspection model is trained using the constructed high-resolution remote sensing image change detection dataset (test dataset). In each training iteration, data enhancement is randomly selected for combination to increase the generalization performance of the model, including random cropping, random flipping, brightness, contrast, saturation, and chroma adjustment.

[0145] 2.5 Model prediction unit

[0146] It is used to input remote sensing image data, and perform enhancement operations on the input data to generate multiple enhanced samples, input the enhanced samples into the initial inspection model after iterative training for prediction, obtain initial prediction results, and average or vote on the initial prediction results to obtain final prediction results; wherein the enhancement operations include random scaling and random flipping.

[0147] 3. Post-processing module

[0148] Used to analyze and detect the initial inspection results using an unsupervised change detection algorithm to complete the final inspection; the post-processing module includes:

[0149] 3.1 Invariant Feature Analysis Unit

[0150] It is used to process the initial inspection result by using deviation calculation, position correction and invariance measurement methods, eliminate pseudo changes caused by geometric displacement in the initial inspection result, and obtain a secondary inspection result.

[0151] 3.2 Morphological Operation Unit

[0152] It is used to fill holes in the spots in the secondary detection results using morphological opening and closing operations, delete small spots according to a preset pixel threshold, and then use morphological dilation operations and erosion operations with 7 pixels to remove burrs and sharp grooves at the boundaries of the spots in the secondary detection results, and use the DP algorithm to smooth the boundaries of the spots in the secondary detection results, complete boundary optimization, and obtain the final detection results.

[0153] 3.3 The unsupervised change detection algorithm includes:

[0154] Multivariate change detection method: In the mining of bi-temporal, multivariate and hypervariate image data, it can efficiently capture the changes of unstable points, accurately obtain change information and is less affected by external factors. It is widely used in remote sensing image change detection. The core idea of ​​IR-MAD is that the initial weight of each pixel is 1. Each iteration assigns a new weight to each pixel in the two images. Through calculation, pixels that have not changed have a larger weight. The final weight is the basis for determining whether each pixel has changed. After several iterations, the weight of each pixel will tend to stabilize until the change is less than the set threshold or no longer changes, then the iteration is stopped.

[0155] Principal component analysis: Principal component analysis is a multi-band orthogonal linear transformation based on statistical characteristics. It obtains eigenvectors and eigenvalues ​​by calculating the covariance matrix between each band of the multispectral image, thereby obtaining the transformation matrix. The transformed bands are called principal components. In the feature domain, the principal components are orthogonal but unrelated. The energy is mainly concentrated in the first few principal components, and the amount of information in the subsequent principal components gradually decreases. Therefore, principal component analysis can achieve the purpose of separating change information in images, reducing data redundancy, and highlighting different types of objects.

[0156] Change vector analysis: Change vector analysis (CVA) is a method to interpret change information based on the magnitude and direction of the change. The dual-phase image is described as the following three components: a feature vector at time T1, a feature vector at time T2, and a vector connecting the feature vectors at time T1 and T2. The connected vector is also called a change vector, and its magnitude and direction information can help obtain the degree of change and distinguish the type of change. In standard CVA, the magnitude variable is obtained by calculating the length of the change feature. The larger the magnitude, the greater the possibility of change. The direction variable can be used to indicate the type of change, which is calculated by the direction cosines of each axis in the feature space. Therefore, the comprehensive use of magnitude and direction information is a means to effectively mark changes and minimize pseudo-change errors. However, in a feature space with n dimensions, n-1 direction cosines are required to describe the direction of change, which will make the output data matrix too complex and difficult to interpret. To solve this problem, a series of improved methods have been proposed and used in image classification, such as sequential spectral change vector analysis (S2CVA).

[0157] Object-based change pattern analysis: Object-based methods divide the image into more meaningful and uniform regions and organize them into image objects. An object is a group of pixels that are spatially adjacent and spectrally similar to each other. Object-based analysis units provide a suitable framework to model statistical features such as texture, and can be roughly divided into: image object overlay comparison method, image object direct comparison method, and multi-temporal image object comparison method.

[0158] Constrained post-processing method based on segmentation objects: By analyzing the proportion of changed pixels in each segmentation unit and setting a certain threshold, the pseudo-changes caused by displacement offset and spectral difference are filtered out to improve the accuracy of change detection. At the same time, by introducing segmentation objects, the output results of change detection are more consistent with the edges of real changes.

[0159] 3.4 Test data: The test data available are taken by the GF02 satellite in the Ordos area on November 19, 2019, covering longitude 109.74-109.78 degrees and latitude 39.62-39.64 degrees. The width is about 4.3km*3.65km.

[0160] 4. Task management module

[0161] It is used to integrate the detection process of the remote sensing image into one task, and create and execute multiple tasks on the platform interface. The task management module includes:

[0162] 4.1 Create a new task unit

[0163] Used to input the task name, interpretation type and image dataset; the interpretation types include ground objects, targets and topics; the ground objects include deep learning change detection options, change vector analysis change detection options, iterative multivariate change detection options, object-based change detection options, principal component analysis change detection options, multivariate change detection functions and change detection accuracy evaluation options.

[0164] New Task: Click the [New Task] button to pop up the New Task window. The required parameters are task name, interpretation type and image dataset. The optional parameter is task description. Interpretation types include three modules: ground objects, targets and topics. Ground objects include six functions: deep learning change detection, change vector analysis change detection, iterative multivariate change detection, object-based change detection, principal component analysis change detection, multivariate change detection and change detection accuracy evaluation. Click the [Please click to select] button to pop up the Select Image Dataset dialog box. Select from the dataset list on the left. The images in the dataset are displayed on the right, including the image name, data size and data format. Check the image and click [OK] to complete the new task.

[0165] 4.2 Task List Unit

[0166] Used to display list information and perform list operations.

[0167] List information: The task list displays information such as task name, task type, interpretation type, status, start time, end time, running time and operation. The status is divided into in progress, completed and terminated.

[0168] List operations: including details, rename, delete, and terminate.

[0169] 4.3 List Sorting Unit

[0170] In the list, you can sort by task name, task type, interpretation type, start time, end time and running time. Click the triangle button next to the task details to sort the list in ascending or descending order.

[0171] 4.4 Task Search Unit

[0172] Click the [Please enter keywords] dialog box, enter the search keywords, and press the [Enter] key to search in the task list.

[0173] 5. Added new models

[0174] Add a new model: Click the [Add a new model] button to pop up the Add a new model window. Select the compressed model file in the pop-up box.

[0175] Model list information: including model name, publisher, scenario, type, version number, upload time and operation.

[0176] The model list supports sorting by model name, publisher, scenario, type, version number, upload time, and other information.

[0177] The model list supports searching by category information, resolution and keyword combinations.

[0178] Model list operations include details, task association, update, and visualization. Click the [Details] button to jump to the details page. The list information on the details page includes model information, model accuracy, model architecture diagram, and model JSON code information. Click the [Task Association] button to jump to the task management page. The list information on the task management page is all the tasks run by the model. Click the [Update] button to update the model. Upload the new model compression package in the page that jumps out to complete the model update.

[0179] You can delete the model by clicking the [Delete] button in the model list. However, it is not recommended that users easily delete the model files uploaded by the administrator, otherwise the model functions cannot be called normally after deletion.

[0180] Therefore, the present invention provides a spatial data AI intelligent analysis system that can perform rapid pixel-level change detection on input multi-period remote sensing image data (spatial data), improve extraction accuracy, and obtain more accurate change area information.

[0181] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0182] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A spatial data AI intelligent analysis system, characterized in that: Includes interconnected platform interface, AI earth module, model building module, post-processing module and task management module; The AI ​​Earth module is used to upload multi-period remote sensing image data using the platform interface, and perform change detection on remote sensing images through layer management operations, data upload operations, online interpretation operations, and online editing operations, extracting the similarities and differences of images of different phases in the same area in the remote sensing image, and obtaining change spots; The model building module is used to build an initial inspection model using an automatic change detection algorithm and a deep learning model based on an encoding-decoding structure, use the initial inspection model to perform the change detection task of the change pattern, and use the constructed test data set to perform model training and model prediction to obtain an initial inspection result; The post-processing module is used to analyze and detect the initial inspection results using an unsupervised change detection algorithm to complete the final inspection; The task management module is used to integrate the detection process of the remote sensing image into one task, and to create and execute multiple tasks on the platform interface.

2. A spatial data AI intelligent analysis system according to claim 1, characterized in that: The AI ​​Earth module includes: A layer management unit, used to manage raster layers and vector layers in the uploaded remote sensing image using the platform interface, the layer management unit includes a background layer, a raster layer, a vector layer and a layer menu panel; A data uploading unit, used to add new vector data and raster data using the platform interface, the data uploading unit includes a vector uploading function and a raster uploading function; An online interpretation unit is used to draw AOI on the remote sensing image based on the operation results of the layer management unit and the data uploading unit to obtain the target area, perform change detection on the target area, extract the similarities and differences of different phases of the same area in the remote sensing image, and obtain the change spots. The online interpretation unit includes AOI drawing function, ground object extraction function, target recognition function and thematic detection function; An online editing unit is used to edit and modify the vector layer and the vector data using the platform interface. The online editing unit includes an element adding option, an element selecting option, an element moving option, an element identifying option, an element pasting option, and an element deleting option.

3. A spatial data AI intelligent analysis system according to claim 2, characterized in that: The thematic detection functions include multi-source change detection tasks, iterative multivariate change detection tasks, principal component analysis change detection tasks, change vector analysis change detection tasks, object-based change detection tasks and deep learning change detection tasks.

4. A spatial data AI intelligent analysis system according to claim 3, characterized in that: The model building module includes: The feature extraction unit is used to learn and extract the features of the dual-phase remote sensing image based on the twin network structure using ResNet50 and residual network to obtain a difference feature map; A feature fusion unit is used to expand the size of the difference feature map by using a feature pyramid network to obtain a feature enlarged map, connect the feature enlarged map horizontally to the corresponding encoding layer, and fuse the features of the bottom layer and the high layer of the feature enlarged map to obtain a multi-level feature map with different sizes, and input the multi-level feature map into a decoder for multiple upsampling to obtain a new feature map with detail information and semantic features; A deep supervision unit is used to introduce an intermediate loss function in the hidden layer for supervision according to the output of the encoding layer of each scale, and add an auxiliary classifier to supervise the backbone network, and then perform multiple upsampling on the new feature map according to the encoding structure of each scale, and use a Sigmoid classifier to map the new feature map that has completed multiple upsampling into a binary map, compare the binary map with the label map and calculate the loss; A model training unit, used to construct a dynamic weighted cross entropy loss function, and use the test data set to iterate the training of the initial inspection model, and randomly select data enhancements for combination during the training iteration process; The model prediction unit is used to input remote sensing image data and perform enhancement operations on the input data to generate multiple enhanced samples, input the enhanced samples into the initial inspection model after iterative training for prediction to obtain initial prediction results, and average or vote on the initial prediction results to obtain final prediction results; wherein the enhancement operations include random scaling and random flipping.

5. A spatial data AI intelligent analysis system according to claim 4, characterized in that: The post-processing module comprises: An invariant feature analysis unit, used to process the initial inspection result by using deviation calculation, position correction and invariance measurement methods, eliminate pseudo changes caused by geometric displacement in the initial inspection result, and obtain a secondary inspection result; The morphological operation unit is used to fill holes in the spots in the secondary detection results using morphological opening and closing operations, delete small spots according to a preset pixel threshold, and then use morphological dilation operations and erosion operations with 7 pixels to remove burrs and sharp grooves at the boundaries of the spots in the secondary detection results, and use the DP algorithm to smooth the boundaries of the spots in the secondary detection results, complete boundary optimization, and obtain the final detection result.

6. A spatial data AI intelligent analysis system according to claim 5, characterized in that: The unsupervised change detection algorithm includes a multivariate change detection method, a principal component analysis method, a change vector analysis method, an object-based change spot analysis method and a segmentation object-based constrained post-processing method; wherein the object-based change spot analysis method includes an image object overlay comparison method, an image object direct comparison method and a multi-phase image object comparison method; the segmentation object-based constrained post-processing method is used to analyze the proportion of changed pixels in each segmentation unit in the spot, set a threshold to filter out pseudo-changes, and introduce segmentation objects for comparison with the final detection results to complete boundary optimization.

7. A spatial data AI intelligent analysis system according to claim 6, characterized in that: The task management module includes: Create a new task unit to input the task name, interpretation type and image dataset; the interpretation type includes ground objects, targets and topics, and the ground objects include deep learning change detection options, change vector analysis change detection options, iterative multivariate change detection options, object-based change detection options, principal component analysis change detection options, multivariate change detection functions and change detection accuracy evaluation options; Task list unit, used to display list information and perform list operations; List sorting unit, used to sort the list in ascending or descending order; Task search unit, used to search the list by keyword.

Images