Change detection method and device

Abstract

The invention provides a change detection method and device, and relates to the technical field of image processing. The method comprises the steps of firstly obtaining a remote sensing image of a to-be-predicted region, then reading the remote sensing image of the to-be-predicted region, then generating a change region result by using the read remote sensing image of the to-be-predicted region and a twin neural network model, then obtaining categories of a front and back time phase change region by using the change region result and a scene classification network, and finally, obtaining a time phase change region. And completing semantic change detection drawing by using the change region result and the categories of the previous and later time phase change regions. In this way, scene classification is performed after detection, and change detection can be completed accurately.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a change detection method and device. Background Art

[0002] Human activities have caused serious interference to the ecological environment. Obtaining the areas and categories of interference is of great significance for restoring the ecological environment. The human activities that are the focus of attention in the field of ecological and environmental supervision mainly include resource exploitation such as quarrying and sand mining, energy, industry, transportation, tourism development, and other development and construction activities that cause ecological damage, such as villas and pond excavation. Multi-category change detection based on high-resolution remote sensing images has the advantages of high efficiency, wide range, and low cost compared to field-based multi-category change detection. Therefore, remote sensing technology is increasingly used in multi-category change detection.

[0003] The purpose of change detection is to locate and analyze the areas where land cover changes on the earth's surface. Most of the existing change detection methods focus on detecting the location of pixels that change between images, and express the task as a binary change detection (BCD) problem. However, BCD ignores the detection of the type of pixel changes, while semantic information is a very important part of the information in change detection. Therefore, it is very important to design a semantic change detection (SCD) method that can simultaneously extract the change area and identify the type of change. At present, there are two solutions to the semantic change detection task, the classification-then-comparison method and the detection-then-classification method. The classification-then-comparison method semantically segments the remote sensing images of the previous and next phases, determines the type of change, and then determines the changed area by the difference between the two images. The detection-then-classification method first performs binary change detection, and then semantically segments the changed area to obtain the changed area and category. However, binary change detection can only obtain the changed area, but not the type of pixel changes; the classification-then-comparison method relies on the classification accuracy, and is prone to pseudo-changes caused by factors such as lighting, season, and terrain, as well as the problem of imbalanced categories; the detection-then-classification method needs to complete the two tasks of change area extraction and semantic segmentation in sequence, and the computational efficiency is not high.

[0004] In summary, how to accurately complete change detection is a problem that those skilled in the art need to solve urgently. Summary of the invention

[0005] In view of this, the present application provides a change detection method and device, aiming to accurately complete change detection.

[0006] In a first aspect, the present application provides a change detection method, comprising:

[0007] Obtain remote sensing images of the area to be predicted;

[0008] Reading the remote sensing image of the area to be predicted;

[0009] Generate a change area result by using the read remote sensing image of the area to be predicted and the twin neural network model;

[0010] Using the change region result and the scene classification network, the categories of the change regions before and after the phases are obtained;

[0011] The semantic change detection mapping is completed by using the change region results and the categories of the change regions before and after the phases.

[0012] Optionally, before generating a change area result by using the read remote sensing image of the area to be predicted and the twin neural network model, the method further includes:

[0013] Acquisition of remote sensing images;

[0014] Dividing the remote sensing images into a training set and a validation set;

[0015] Using the training set to train the twin neural network model;

[0016] Using the training set to train the scene classification network;

[0017] The twin neural network model and the scene classification network are evaluated for accuracy.

[0018] Optionally, before dividing the remote sensing image into a training set and a validation set, the method further includes:

[0019] The remote sensing image is subjected to resampling processing, projection conversion processing, radiation calibration processing, atmospheric correction processing and geographic registration processing to obtain a preprocessed remote sensing image.

[0020] Optionally, dividing the remote sensing image into a training set and a validation set includes:

[0021] Obtaining the intersection area of ​​the two remote sensing images;

[0022] Marking the contour of the changed area and the category to which each image spot belongs on the intersection area of ​​the two remote sensing images;

[0023] The processed remote sensing image is cropped and divided into a training set and a validation set.

[0024] Optionally, before reading the remote sensing image of the area to be predicted, the method further includes:

[0025] The remote sensing image of the area to be predicted is subjected to resampling processing, projection conversion processing, radiation calibration processing, atmospheric correction processing and geographic registration processing to obtain a preprocessed remote sensing image of the area to be predicted.

[0026] Optionally, reading the remote sensing image of the area to be predicted includes:

[0027] The remote sensing image of the area to be predicted is read in a range reading manner.

[0028] In a second aspect, the present application provides a change detection device, comprising:

[0029] The first acquisition module is used to acquire the remote sensing image of the area to be predicted;

[0030] A reading module, used for reading the remote sensing image of the area to be predicted;

[0031] A generation module, used to generate a change area result by using the read remote sensing image of the area to be predicted and the twin neural network model;

[0032] A second acquisition module is used to obtain the categories of the regions of change before and after the phases by using the change region results and the scene classification network;

[0033] The semantic change detection module is used to complete semantic change detection mapping by using the change area results and the categories of the previous and next phase change areas.

[0034] Optionally, the device further comprises:

[0035] The third acquisition module is used to acquire remote sensing images;

[0036] A division module, used for dividing the remote sensing image into a training set and a verification set;

[0037] A first training module, used to train the twin neural network model using the training set;

[0038] A second training module, used to train the scene classification network using the training set;

[0039] An accuracy evaluation module is used to use the twin neural network model and the scene classification network to perform accuracy evaluation.

[0040] Optionally, the device further comprises:

[0041] The second preprocessing module is used to perform resampling processing, projection conversion processing, radiation calibration processing, atmospheric correction processing and geographic registration processing on the remote sensing image to obtain a preprocessed remote sensing image.

[0042] Optionally, the division module includes:

[0043] An acquisition unit, used for acquiring an intersection area of ​​the two remote sensing images;

[0044] A marking unit, used for marking the contour of the changed area and the category to which each image spot belongs on the intersection area of ​​the two remote sensing images;

[0045] The division unit is used to cut the processed remote sensing image and divide it into a training set and a verification set.

[0046] Optionally, the device further comprises:

[0047] The second preprocessing module is used to perform resampling processing, projection conversion processing, radiation calibration processing, atmospheric correction processing and geographic registration processing on the remote sensing image of the area to be predicted, so as to obtain the preprocessed remote sensing image of the area to be predicted.

[0048] Optionally, the reading module includes:

[0049] The reading unit is used to read the remote sensing image of the area to be predicted in a range reading manner.

[0050] The present application provides a change detection method. When executing the method, first obtain the remote sensing image of the area to be predicted, then read the remote sensing image of the area to be predicted, and then use the read remote sensing image of the area to be predicted and the twin neural network model to generate the change area result, and then use the change area result and the scene classification network to obtain the category of the before and after phase change area, and finally, use the change area result and the category of the before and after phase change area to complete the semantic change detection mapping. In this way, detection is performed first and scene classification is performed later, which can improve the accuracy of change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 A flow chart of a change detection method provided in an embodiment of the present application;

[0053] Figure 2 A schematic diagram of the structure of a change detection device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. The present application provides a change detection method and device for use in the field of image processing technology. The above is only an example and does not limit the application field of the method and device provided by the present application.

[0055] Human activities have caused serious interference to the ecological environment. Obtaining the areas and categories of interference is of great significance for restoring the ecological environment. The human activities that are the focus of attention in the field of ecological and environmental supervision mainly include resource exploitation such as quarrying and sand mining, energy, industry, transportation, tourism development, and other development and construction activities that cause ecological damage, such as villas and pond excavation. Multi-category change detection based on high-resolution remote sensing images has the advantages of high efficiency, wide range, and low cost compared to field-based multi-category change detection. Therefore, remote sensing technology is increasingly used in multi-category change detection.

[0056] The purpose of change detection is to locate and analyze the areas where land cover changes on the earth's surface. Most of the existing change detection methods focus on detecting the location of pixels that change between images, and express the task as a binary change detection (BCD) problem. However, BCD ignores the detection of the type of pixel changes, while semantic information is a very important part of the information in change detection. Therefore, it is very important to design a semantic change detection (SCD) method that can simultaneously extract the change area and identify the type of change. At present, there are two solutions to the semantic change detection task, the classification-then-comparison method and the detection-then-classification method. The classification-then-comparison method semantically segments the remote sensing images of the previous and next phases, determines the type of change, and then determines the changed area by the difference between the two images. The detection-then-classification method first performs binary change detection, and then semantically segments the changed area to obtain the changed area and category. However, binary change detection can only obtain the changed area, but not the type of pixel changes; the classification-then-comparison method relies on the classification accuracy, and is prone to pseudo-changes caused by factors such as lighting, season, and terrain, as well as the problem of imbalanced categories; the detection-then-classification method needs to complete the two tasks of change area extraction and semantic segmentation in sequence, and the computational efficiency is not high.

[0057] After research, the inventor proposed the technical solution of this application, which first obtains the remote sensing image of the area to be predicted, then reads the remote sensing image of the area to be predicted, and then uses the read remote sensing image of the area to be predicted and the twin neural network model to generate the change area result, and then uses the change area result and the scene classification network to obtain the category of the area of ​​change before and after the time phase, and finally uses the change area result and the category of the area of ​​change before and after the time phase to complete the semantic change detection mapping. In this way, detection is performed first and scene classification is performed later, which can improve the accuracy of change detection.

[0058] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application. It should be noted that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0059] See also Figure 1 , Figure 1 A flow chart of a change detection method provided in an embodiment of the present application includes:

[0060] S101: Acquire a remote sensing image of the area to be predicted.

[0061] After obtaining the remote sensing image of the area to be predicted, it is necessary to pre-process the remote sensing image first, and the specific method is as follows: First, the remote sensing image is a GF-1 image, and different bands of the image have different resolutions, including 4 multispectral bands with a resolution of 8 meters and 1 band with a resolution of 2 meters. It is necessary to resample the bands with a resolution less than 8 meters so that the resolution of all bands is 8 meters. Secondly, the projection method used by different remote sensing images is related to the longitude where they are located, which leads to the projections between different images being often different, hindering the unified processing of the images. It is necessary to perform a unified projection conversion on all images. In an embodiment of the present application, all images are transferred to the WGS-84 coordinate system. Radiation calibration is the process of converting the digital quantization value of an image into physical quantities such as radiation brightness value and reflectivity. In some embodiments, the radiation calibration tool provided in the mature remote sensing data processing platform ENVI can be used to automatically perform image radiation calibration. Atmospheric correction can eliminate the radiation error caused by atmospheric images and better reflect the real situation of the surface. In some implementations, the atmospheric correction tool provided by the mature remote sensing data processing platform ENVI can be used to automatically perform image atmospheric correction. Geo-referencing is to align remote sensing images of the same area at different times to facilitate visual interpretation and automatic interpretation. Geo-referencing can be implemented based on ENVI, and the registration error is controlled within 1 pixel.

[0062] S102: Read the remote sensing image of the area to be predicted.

[0063] Since a single remote sensing image is large, directly reading the remote sensing image will take up a lot of memory and waste a lot of computing resources. That is, in the embodiment of the present application, the image is read by range reading, that is, only part of the impact is stored in the memory each time, and the block reading is cyclic, which can effectively reduce the memory usage in the prediction process.

[0064] S103: Generate a change area result using the read remote sensing image of the area to be predicted and the twin neural network model.

[0065] Before executing step S103, it is necessary to first train the twin neural network model and the scene classification network. The specific process is as follows:

[0066] First, it is necessary to obtain several remote sensing images, which are used for training and verifying the model. Then, it is necessary to preprocess these remote sensing images. The specific preprocessing method is the same as the preprocessing method in step S101, which will not be repeated here.

[0067] Next, the preprocessed remote sensing images are used to generate sample data, that is, the intersection area of ​​the two remote sensing images is obtained, the contour of the changed area is marked on the intersection area of ​​the two remote sensing images, and the category to which each patch belongs is marked. Then the remote sensing images are cut into 512×512 pixels and divided into training set and validation set. The specific steps are as follows:

[0068] First, remote sensing images of two phases are obtained to obtain the geographical scope of the remote sensing images, and it is determined whether the geographical scopes of the two remote sensing images have an intersecting area. If there is no intersecting area, an error is directly reported; if there is an intersecting area, the geographical scope of the intersecting area is calculated, and then the geographical scope of the intersecting area is respectively converted into the image range of the front phase image and the image range of the back phase image, and the front phase image and the back phase image data are respectively cropped according to the two range information to obtain the front and back phase images of the intersecting area.

[0069] Next, the polygon marking of the contour of the changed area is performed under the interactive interface of the image marking software. In the embodiment of the present application, the marking software ArcMap 10.2 can be used for marking. The software supports reading remote sensing images and displaying them in the form of rolling curtains. It supports outlining the contour of the object with polygons based on the image display, and the outlined polygons are stored as vectors in .shp file format. The semantic change detection task requires providing a marked image with the same size as the original image, and each pixel value on the marked image represents whether the pixel has changed. The vector-to-raster tool provided by ArcMap is used to realize the polygon-to-image format marking, and the contour marking of the changed area in raster format is obtained. Then, the category to which the change spots before and after the annotation belong is visually interpreted, and the scene classification samples centered on the change spots are cut out, and the sample size is 256×256. In some embodiments, in order to facilitate the training and testing of change detection, the remote sensing image and its label are cut into a size of 512×512. Finally, after cutting, the samples are randomly divided into a training set and a test set according to a division ratio of 3:1, wherein the training set is used for network training and the test set is used for accuracy evaluation.

[0070] Among them, the training process of the twin neural network model is as follows:

[0071] In the embodiment of the present application, the network model is a twin network structure. This network structure takes a pair of images as input, and extracts image features in parallel by two identical network paths. Since the network structures used in the two parallel network paths are the same and the parameters are shared, it is called a twin network structure. In addition, the network structure is of the Encoder-Decoder type, which follows the process of first stepwise downsampling and then stepwise upsampling.

[0072] In the downsampling stage, the basic structure of EfficientNet-B1 is used. Compared with EfficientNet-B0, EfficientNet-B1 expands the network depth to 1.1 times the original, and expands the network image resolution from 224×224 to 240×240, while the network width remains unchanged. After downsampling, the output feature map size is 1 / 32 of the original image. In the downsampling stage, the features of the dual-phase images are extracted separately through the network structure with shared weights, and all the features of the 5 downsamplings are retained.

[0073] In the upsampling stage, the high-low layer fusion structure of U-Net is used to fuse the features extracted from the previous and next phases. While upsampling, the feature maps of the same size of the previous and next phases are fused. The feature fusion is stacked by channel, and then convolution is performed with a convolution kernel of size 3×3, and one row is filled in the top, bottom, left and right to keep the size of the feature map unchanged.

[0074] In the training process of the twin neural network model, a multi-scale attention mechanism is introduced: first, the two remote sensing images are reduced by 0.5 times respectively to form a set of data with the original size and a set of data with 0.5 times the original image size. Then, the two sets of data are input into the twin network respectively, and each set of data obtains 5 feature maps of different scales. Then, the feature maps of different scales of the previous and next phase images are fused by upsampling + channel splicing + convolution to obtain a fused feature map with the same size as the original image. Then, the fused features of the 0.5 times data are input into the attention structure to generate a 0.5 times attention mask α, (1-α) is the large-scale attention mask. The fused features of the large and small scale data are input into the semantic segmentation structure respectively to generate change detection features. The 0.5 times attention mask and the change detection feature are upsampled to the original image size, and the attention mask and the fused features of different scales are correspondingly multiplied and added to obtain the final change detection result of the changed area.

[0075] Finally, the loss function used is the cross entropy loss function and the Dice loss function weighted at a ratio of 1:1. The Adam optimizer is used, the initial learning rate is set to 0.002, the learning rate strategy selects the adaptive learning rate strategy ReduceLROnPlateau, and the random dropout probability Dropout is 0.1 during training. The adaptive learning rate strategy can adaptively reduce the learning rate according to the current network training accuracy and the time it stays at this accuracy. Setting the random dropout probability can speed up training and improve the generalization ability of the network to a certain extent.

[0076] The process of training the scene classification network includes:

[0077] The task of identifying the attributes of the change spots is carried out based on a two-stage scheme of binary change extraction and spot scene classification. First, the scene classification samples are produced: 1-2 meter resolution, image size 256×256. The sample categories include: ore mining, tailings ponds, residential areas, hydropower stations / dams, wind turbines, thermal power plants, photovoltaics, airports, railways, roads, ditches, and greenhouses. Based on the above scene classification samples, the MobileNet scene classification model is trained to achieve the scene classification task. Secondly, the scene classification prediction module is integrated with the binary change detection module to achieve the change spot extraction and change attribute classification tasks.

[0078] The scene classification network uses the MobileNet model. After extracting the features of the image, the extracted features are converted into the probabilities of each category through the softmax function, and the output with the highest probability is selected to represent the scene of the picture.

[0079] In some implementations, the test set can also be used to evaluate the accuracy of the training model, specifically in the following manner:

[0080] In the change detection stage, four accuracy evaluation parameters, namely, overall classification progress OA, accuracy Precision, recall rate Recall, and F1 score, are used as accuracy evaluation indicators. OA is the ratio of the number of correctly classified category pixels to the total number of category pixels, and its specific calculation formula is:

[0081]

[0082] Among them, p i,i represents the pixels classified into and belonging to the i-th class, p i,j OA represents pixels belonging to the i-th class and classified as the j-th class. By comparing with the sample labels, we can get the total number of correctly extracted true positives TP, the total number of incorrectly extracted false positives FP, and the total number of missed false negatives FN of the classified pixels, so that the precision and recall can be calculated as follows:

[0083]

[0084]

[0085] The F1 score is an indicator used in statistics to measure the accuracy of pixel classification models. It takes into account both the precision and recall of the classification model. It is a reconciled evaluation of the precision and recall. The specific calculation formula is:

[0086]

[0087] In the scene classification stage, the four accuracy evaluation parameters of volume classification progress OA, accuracy Precision, recall rate Recall, and F1 score are still used as accuracy evaluation indicators. However, a single pixel is no longer regarded as a sample, but a patch is regarded as a sample.

[0088] The specific implementation method of generating the change area result by using the read remote sensing image of the area to be predicted and the twin neural network model is as follows:

[0089] The image tensor obtained by reading the image is used, and three scales of 0.5, 1.0, and 1.5 are used to form three sets of data. The twin network model is called three times to output three change detection features and three attention masks. The combination of fused feature maps follows the principle of large first and small later, that is, the change detection features of 1.0 times and 1.5 times the size are combined first, and then combined with the change detection features of 0.5 times the size, and finally the change area result is generated.

[0090] S104: Using the change region result and the scene classification network, the categories of the change regions before and after the phases are obtained.

[0091] Centered on the changed area, the before and after phase images are cropped to a size of 256×256, and the category of the before and after phase changed area is obtained using the scene classification network.

[0092] S105: Using the change region results and the categories of the change regions before and after the time phases, complete semantic change detection mapping.

[0093] The changed area obtained in step S103 is converted into a vector patch, and the category information of the previous and next phases is assigned to the changed patch to complete the semantic change detection mapping.

[0094] In the embodiment of the present application, the remote sensing image of the area to be predicted is first obtained, and then the remote sensing image of the area to be predicted is read, and then the change area result is generated by using the read remote sensing image of the area to be predicted and the twin neural network model, and then the change area result and the scene classification network are used to obtain the category of the area of ​​change before and after the time phase, and finally, the change area result and the category of the area of ​​change before and after the time phase are used to complete the semantic change detection mapping. In this way, detection is performed first and scene classification is performed later, which can improve the accuracy of change detection.

[0095] The above are some specific implementations of the method provided in the embodiment of the present application. Based on this, the present application also provides a corresponding device. The device provided in the embodiment of the present application will be introduced from the perspective of functional modularization.

[0096] See also Figure 2 , Figure 2 The present invention provides a schematic diagram of a change detection device according to an embodiment of the present invention. The change detection device 200 includes:

[0097] The first acquisition module 210 is used to acquire a remote sensing image of the area to be predicted;

[0098] A reading module 220 is used to read the remote sensing image of the area to be predicted;

[0099] A generation module 230 is used to generate a change area result by using the read remote sensing image of the area to be predicted and the twin neural network model;

[0100] A second acquisition module 240 is used to obtain the categories of the regions of change before and after the phases by using the change region result and the scene classification network;

[0101] The semantic change detection module 250 is used to complete semantic change detection mapping by using the change region results and the categories of the previous and next phase change regions.

[0102] Optionally, the device 200 further includes:

[0103] The third acquisition module is used to acquire remote sensing images;

[0104] A division module, used for dividing the remote sensing image into a training set and a verification set;

[0105] A first training module, used to train the twin neural network model using the training set;

[0106] A second training module, used to train the scene classification network using the training set;

[0107] An accuracy evaluation module is used to use the twin neural network model and the scene classification network to perform accuracy evaluation.

[0108] Optionally, the device 200 further includes:

[0109] The second preprocessing module is used to perform resampling processing, projection conversion processing, radiation calibration processing, atmospheric correction processing and geographic registration processing on the remote sensing image to obtain a preprocessed remote sensing image.

[0110] Optionally, the division module includes:

[0111] An acquisition unit, used for acquiring an intersection area of ​​the two remote sensing images;

[0112] A marking unit, used for marking the contour of the changed area and the category to which each image spot belongs on the intersection area of ​​the two remote sensing images;

[0113] The division unit is used to cut the processed remote sensing image and divide it into a training set and a verification set.

[0114] Optionally, the device 200 further includes:

[0115] The second preprocessing module is used to perform resampling processing, projection conversion processing, radiation calibration processing, atmospheric correction processing and geographic registration processing on the remote sensing image of the area to be predicted, so as to obtain the preprocessed remote sensing image of the area to be predicted.

[0116] Optionally, the reading module 220 includes:

[0117] The reading unit is used to read the remote sensing image of the area to be predicted in a range reading manner.

[0118] The embodiments of the present application also provide corresponding devices and computer storage media for implementing the solutions provided by the embodiments of the present application.

[0119] The device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes the method described in any embodiment of the present application.

[0120] The computer storage medium stores codes, and when the codes are executed, a device executing the codes implements the method described in any embodiment of the present application.

[0121] Through the description of the above implementation methods, it can be known that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a read-only memory (ROM) / RAM, a magnetic disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in each embodiment of the present application or some parts of the embodiments.

[0122] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0123] It should also be noted that the various embodiments in this specification are described in a progressive manner, and the same and similar parts between the various embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the equipment and device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The device embodiments described above are merely schematic, in which the units described as separate components may or may not be physically separated, and the components prompted as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0124] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A change detection method, characterized in that: include: Obtain remote sensing images of the area to be predicted; Reading the remote sensing image of the area to be predicted; Generate a change area result by using the read remote sensing image of the area to be predicted and the twin neural network model; Using the change region result and the scene classification network, the categories of the change regions before and after the phases are obtained; The semantic change detection mapping is completed by using the change region results and the categories of the change regions before and after the phases.

2. The method according to claim 1, characterized in that Before generating the change area result by using the read remote sensing image of the area to be predicted and the twin neural network model, the method further includes: Acquisition of remote sensing images; Dividing the remote sensing images into a training set and a validation set; Using the training set to train the twin neural network model; Using the training set to train the scene classification network; The twin neural network model and the scene classification network are evaluated for accuracy.

3. The method according to claim 2, characterized in that Before dividing the remote sensing image into a training set and a validation set, the method further includes: The remote sensing image is subjected to resampling processing, projection conversion processing, radiation calibration processing, atmospheric correction processing and geographic registration processing to obtain a preprocessed remote sensing image.

4. The method according to claim 2, characterized in that: The step of dividing the remote sensing image into a training set and a validation set comprises: Obtaining the intersection area of ​​the two remote sensing images; Marking the contour of the changed area and the category to which each image spot belongs on the intersection area of ​​the two remote sensing images; The processed remote sensing image is cropped and divided into a training set and a validation set.

5. The method according to claim 1, characterized in that Before reading the remote sensing image of the area to be predicted, the method further includes: The remote sensing image of the area to be predicted is subjected to resampling processing, projection conversion processing, radiation calibration processing, atmospheric correction processing and geographic registration processing to obtain a preprocessed remote sensing image of the area to be predicted.

6. The method according to claim 1, characterized in that The step of reading the remote sensing image of the area to be predicted includes: The remote sensing image of the area to be predicted is read in a range reading manner.

7. A change detection device, characterized in that: include: The first acquisition module is used to acquire the remote sensing image of the area to be predicted; A reading module, used for reading the remote sensing image of the area to be predicted; A generation module, used to generate a change area result by using the read remote sensing image of the area to be predicted and the twin neural network model; A second acquisition module is used to obtain the categories of the regions of change before and after the phases by using the change region results and the scene classification network; The semantic change detection module is used to complete semantic change detection mapping by using the change area results and the categories of the previous and next phase change areas.

8. The device according to claim 7, characterized in that The device also includes: The third acquisition module is used to acquire remote sensing images; A division module, used for dividing the remote sensing image into a training set and a verification set; A first training module, used to train the twin neural network model using the training set; A second training module, used to train the scene classification network using the training set; An accuracy evaluation module is used to use the twin neural network model and the scene classification network to perform accuracy evaluation.

9. The device according to claim 8, characterized in that The device also includes: The second preprocessing module is used to perform resampling processing, projection conversion processing, radiation calibration processing, atmospheric correction processing and geographic registration processing on the remote sensing image to obtain a preprocessed remote sensing image.

10. The device according to claim 8, characterized in that The division module comprises: An acquisition unit, used for acquiring an intersection area of ​​the two remote sensing images; A marking unit, used for marking the contour of the changed area and the category to which each image spot belongs on the intersection area of ​​the two remote sensing images; The division unit is used to cut the processed remote sensing image and divide it into a training set and a verification set.

Images