A remote sensing land class change detection method based on an optimized Unet++ neural network deep learning

Abstract

This invention discloses a remote sensing land use change detection method based on optimized Unet++ neural network deep learning. The method includes: acquiring an image to be detected; inputting the image to be detected into a remote sensing land use change detection model for detection to obtain a detection result image; and outputting the detection result image. The remote sensing land use change detection model is a Unet++ neural network model trained using annual land use update results as a sample database. By implementing the method of this invention, the efficiency of land use patch updating can be improved, and the accuracy, precision, recall, and correctness of land use change detection can be increased, with strong practicality.

Description

TECHNICAL FIELD

[0001] The present application relates to a land class change detection method, more particularly to a remote sensing land class change detection method based on an optimized Unet++ neural network deep learning. BACKGROUND

[0002] With the continuous acceleration of urbanization, the change of land cover categories is more and more frequent, and the change is more and more complex. At the same time, the remote sensing technology including unmanned aerial vehicle remote sensing is developing continuously, the image resolution is further improved, and the observation accuracy of the ground is also continuously improved. These conditions have put forward higher requirements for the current land class updating and change detection technology.

[0003] At present, county area map spot updating mainly relies on semi-manual drawing under human-computer interaction. Although the introduction of computers reduces manual work to a certain extent, the continuous human-computer interaction mode restricts the operation personnel by their own experience and operation habit, the spot result has strong subjectivity, the spot missing condition is difficult to estimate, and a large amount of time and manpower is needed. The current land class detection methods include post-classification comparison method and direct comparison method. The change detection result of the post-classification comparison method is mainly restricted by the insufficient classification accuracy. When the change detection is carried out after classification, the accuracy error will be multiplied in the detection process, so that the detection result accuracy is greatly reduced. Especially in the case of multiple land classes, the accuracy error of multi-classification and the complexity of cross comparison will multiply the error of change detection after classification. The direct comparison method refers to a method of directly extracting the change area in the image without pre-classifying the multi-temporal images. This method is direct and efficient, but usually cannot distinguish the change type. After introducing various classification and extraction algorithms, manual subjective intervention still has a great influence on the extraction result. Especially in the direct classification method, the extraction accuracy of the change part is greatly dependent on the determination of the threshold value. The threshold value of various methods is usually manually specified with the assistance of statistics, and the controllability of the extraction accuracy of the change is low, and it is difficult to realize full automation processing. Although there are also deep learning models for land class change detection, the models are highly targeted and have insufficient universality. The current change detection methods and models are often designed and experimented according to a specific task, and it is difficult to maintain effectiveness after changing the data source or experimental environment. In the direct extraction method, the threshold selection method depends on the statistical calculation result to a certain extent. When the task target and the statistical target are inconsistent, statistics cannot automatically select the threshold value, and there is no way to assist the manual threshold value. Specifically, the change detection based on the U-shaped network according to the difference map loses the rich original information in the image, resulting in low change detection accuracy. The change detection based on the U-shaped network according to the multi-temporal data directly retains as much rich information as possible in the front and rear temporal images, and the change detection accuracy is improved to a certain extent. However, the detection ability of the model in details is still weak, a large amount of detailed information is lost in the predicted change result, there is a certain gap between the change area boundary and the labeled result, and part of the small range spot cannot be detected.

[0004] Therefore, it is necessary to design a new method to improve the efficiency of land class spot updating, improve the precision, precision, recall rate and accuracy of land class change detection, and have strong practicability. SUMMARY

[0005] The purpose of the present application is to overcome the defects of the prior art, and provide a remote sensing land class change detection method based on an optimized Unet++ neural network deep learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a remote sensing land cover change detection method based on optimized Unet++ neural network deep learning, comprising:

[0007] Acquire the image to be detected;

[0008] The image to be detected is input into the remote sensing land cover change detection model for detection to obtain the detection result image;

[0009] Output the detection result image;

[0010] The remote sensing land use change detection model refers to the model obtained by training a Unet++ neural network model using annual land use update results as a sample database.

[0011] Its further technical solution is as follows: the remote sensing land use change detection model refers to the model obtained by training a Unet++ neural network model using annual land use update results as a sample database, including:

[0012] A sample library of land use changes required for deep learning was created based on the annual land use update results.

[0013] Training, validation, and test sets were created based on the aforementioned land use change sample database.

[0014] Construct a Unet++ neural network model;

[0015] Set the loss function, and set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model. Take a weighted average of the loss functions of the backbone and branches to obtain the final model loss function; set the initial learning rate and optimization strategy.

[0016] The training set is input into the optimized Unet++ neural network model for training, with the loss function as the metric to obtain the training results. The test set is input into the trained model to extract features at each level. The segmentation results corresponding to the features extracted at each layer during the downsampling process are fused to obtain the output result image. This image is then compared with the ground truth in the test set. The model that passes the test can be used as a remote sensing land cover change detection model.

[0017] The further technical solution is as follows: the construction of the Unet++ neural network model includes:

[0018] The downsampling channels of the four U-shaped networks are merged, and the four U-shaped models are integrated into one model to obtain the Unet++ neural network model;

[0019] The Unet++ neural network model consists of a deep U-shaped network as its backbone, with branches consisting of U-shaped networks of different depths. Each branch is derived from a corresponding downsampling layer. The loss function for each branch is selected and its weights are adjusted based on the branch's characteristics. The loss functions of each branch are then combined to form the final model loss function.

[0020] The further technical solution is as follows: The step of inputting the training set into the Unet++ neural network model for training to obtain training results includes:

[0021] The training set is input into the Unet++ neural network model for training, and the feature extraction results of the same layer are interconnected to obtain the training results.

[0022] The further technical solution is as follows: the loss function convergence strategy is that if the final model loss function does not decrease for three consecutive times, the most recently saved training result is returned, the learning rate is halved, and the Adam optimizer is used to calculate the extreme value.

[0023] The present invention also provides a remote sensing land cover change detection device, comprising:

[0024] The image acquisition unit is used to acquire the image to be detected;

[0025] The detection unit is used to input the image to be detected into the remote sensing land cover change detection model for detection, so as to obtain the detection result image;

[0026] The output unit is used to output the detection result image.

[0027] Its further technical solution includes a model generation unit, which is used to train the Unet++ neural network model using the annual land use update results as a sample library to obtain a remote sensing land cover change detection model.

[0028] The further technical solution is as follows: the model generation unit includes:

[0029] The sample library creation sub-unit is used to create a land use change sample library required for deep learning based on the annual land use update results.

[0030] The dataset creation subunit is used to create training, validation, and test sets based on the land use change sample library.

[0031] Construct sub-units for building Unet++ neural network models;

[0032] The loss function construction sub-unit is used to set the loss function and the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model. The loss functions of the backbone and branches are weighted and averaged to obtain the final model loss function.

[0033] The optimization subunit is used to set the initial learning rate and optimization strategy; the training subunit is used to input the training set into the Unet++ neural network model for training to obtain the training results.

[0034] Save sub-units, used to save training results with the loss function as the metric;

[0035] The test subunit is used to input the test set into the trained model, extract features at each level, and fuse the segmentation results corresponding to the features extracted at each layer during the downsampling process to obtain the output result image. The model that passes the test can be used as a remote sensing land cover change detection model.

[0036] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described above.

[0037] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0038] The beneficial effects of this invention compared with the prior art are as follows: This invention obtains the detection result image by inputting the image to be detected into the remote sensing land cover change detection model. The remote sensing land cover change detection model is obtained by training the Unet++ neural network model. The Unet++ model is applied to the remote sensing land cover change detection process to improve the efficiency of land cover patch updating, improve the accuracy, precision, recall and correctness of land cover change detection, and has strong practicality.

[0039] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram illustrating an application scenario of the remote sensing land cover change detection method provided in this embodiment of the invention;

[0042] Figure 2 This is a flowchart illustrating the remote sensing land cover change detection method provided in an embodiment of the present invention.

[0043] Figure 3This is a schematic diagram of a sub-process of the remote sensing land cover change detection method provided in an embodiment of the present invention;

[0044] Figure 4 A schematic diagram of the Unet++ neural network model provided in an embodiment of the present invention;

[0045] Figure 5 Image detection illustration of the remote sensing land cover change detection method provided in this embodiment of the invention. Figure 1 ;

[0046] Figure 6 Image detection illustration of the remote sensing land cover change detection method provided in this embodiment of the invention. Figure 2 ;

[0047] Figure 7 Image detection illustration of the remote sensing land cover change detection method provided in this embodiment of the invention. Figure 3 ;

[0048] Figure 8 This is a schematic block diagram of a remote sensing land cover change detection device provided in an embodiment of the present invention;

[0049] Figure 9 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0052] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0053] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0054] Please see Figure 1 and Figure 2 , Figure 1 This is a schematic diagram illustrating an application scenario of the remote sensing land cover change detection method provided in an embodiment of the present invention. Figure 2 This is a schematic flowchart illustrating the remote sensing land cover change detection method provided in an embodiment of the present invention. This method is applied in a server. The server interacts with a camera, capturing images to be detected and inputting them into the remote sensing land cover change detection model for land cover change detection. By analyzing the characteristics of land cover patches, such as random size, diverse types, and complex morphological features, a deep learning neural network with powerful learning capabilities can identify and segment target land covers. Deep learning is directly applied to the detection of change areas in multi-temporal remote sensing images, enabling automatic extraction of land cover change patches by computer, reducing the subjective influence of manual patch updates, and further improving the efficiency of county-level land cover patch updates.

[0055] Figure 2 This is a flowchart illustrating a remote sensing land cover change detection method based on optimized Unet++ neural network deep learning, provided by an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S130.

[0056] S110. Obtain the image to be detected.

[0057] In this embodiment, the image to be detected refers to a multi-temporal remote sensing image captured by a camera.

[0058] S120. Input the image to be detected into the remote sensing land cover change detection model for detection to obtain the detection result image.

[0059] In this embodiment, the detection result image refers to the image formed by fusing the features of each layer of the image after segmentation.

[0060] The remote sensing land use change detection model refers to the model obtained by training a Unet++ neural network model using annual land use update results as a sample database.

[0061] In one embodiment, please refer to Figure 3 The aforementioned remote sensing land use change detection model refers to the model obtained by training a Unet++ neural network model using annual land use update results as a sample database, including the following steps S121 to S128.

[0062] S121. Based on the annual land use update results, create a sample library of land use changes required for deep learning.

[0063] In this embodiment, the land use change sample database refers to a database consisting of change images corresponding to the annual land use update results.

[0064] Based on the annual land use update results, a land use change sample library was created for deep learning. Transfer learning was introduced to improve the generalization ability of the deep learning model, accelerate the model training rate, and make full use of each dataset. Overfitting problems in model training were reduced by data augmentation, L2 norm regularization, adding BN (Batch Normalization), Dropou layers, and early termination of training. These methods are all existing technologies and will not be elaborated here.

[0065] In terms of the construction of the change detection sample library, it is produced by analyzing land use data from different years. It has the characteristics of complete data, wide selection range and rich sources, which solves the problems of change detection samples relying on the experience of labelers, high manpower input and inconsistent labeling standards.

[0066] S122. Based on the land use change sample database, create a training set, a validation set, and a test set.

[0067] In this embodiment, the training set refers to the annual land use update results used for model training, the validation set refers to the annual land use update results used to validate the trained model, and the test set refers to the annual land use update results used to test the trained model.

[0068] S123. Construct the structure of the Unet++ neural network model.

[0069] Specifically, such as Figure 4 As shown, the downsampling channels of the four U-shaped networks are merged, and the four U-shaped models are integrated into one model to obtain the Unet++ neural network model;

[0070] The Unet++ neural network model consists of a deep U-shaped network as its backbone, with branches consisting of U-shaped networks of different depths. Each branch is derived from a corresponding downsampling layer. The loss function of each branch is adjusted by weights, and the loss functions of each branch are combined to form the final model loss function.

[0071] Specifically, the Unet++ neural network model is constructed based on the U-shaped neural network model, i.e., the U-shaped network optimization. Although the U-shaped neural network integrates shallow and deep features by completely feeding back the feature information extracted during downsampling into the upsampling path through the pruning and copying process, some feature maps lose or distort information after multiple upsamplings after being integrated into the heatmap. Furthermore, due to the different positions of the intermediate feature layers integrated into the upsampling channel, their influence on the final result varies, and the further away from the final output, the more significantly the feature map's influence on the result is weakened. Therefore, the following method is used to optimize the U-shaped neural network model to construct the Unet++ neural network model. Specifically, the downsampling channels of the four U-shaped models are merged, and a group of neural networks are integrated into one model, i.e., the Unet++ neural network model, which enables the direct output of information extracted from different depths through the construction of shallow networks. Drawing on the idea of ​​deep supervision, the deepest U-shaped network among the four is used as the backbone, and U-shaped networks of different depths are regarded as branches of the backbone. The branches are derived from the corresponding downsampling layers, and the loss function of each branch is combined into the loss function of the overall model through weight adjustment. The model is trained using only the combination of downsampling channel merging and deep supervision, i.e., only the Unet++ neural network model is trained. During training, the error propagates back along different U-shaped networks. In order to increase the backpropagation path of the error and improve the training rate of the model, the relationship between each U-shaped network is fully utilized to connect the feature extraction results of the same layer at the macro level.

[0072] Based on the U-shaped neural network, the model structure is improved by adding model parameters to form an optimized Unet++ neural network, which enhances the model's learning ability and improves the accuracy, precision, recall, and correctness of land use change detection.

[0073] Because the Unet++ model contains four U-shaped networks of different depths, each network has a corresponding channel to output segmentation results during upsampling. Therefore, during downsampling, information extracted from any layer of the model can be processed by the corresponding U-shaped network to output segmentation results, improving the model's ability to perceive shallow feature information. Applying deep supervision to perform relay analysis on the loss results of U-shaped networks of different depths allows direct understanding of the model training status for extracting corresponding features at different network levels. Weight adjustments can then be made to adjust the model's detection focus.

[0074] Specifically, a loss function suitable for the usage scenario is set, and the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model are set. The initial learning rate and optimization strategy are also set to complete the optimization of the Unet++ model.

[0075] In this embodiment, a loss function suitable for the usage scenario is set, and the weights of different depth U-shaped network branches in the comprehensive loss function of the Unet++ neural network model are set. An initial learning rate and optimization strategy are also set to optimize the Unet++ model, including:

[0076] Set a loss function that suits the use case, and set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model, as well as the initial learning rate and optimization strategy.

[0077] The downsampling channels of the four U-shaped networks are merged, and the four U-shaped models are integrated into one model to obtain the Unet++ neural network model;

[0078] The Unet++ neural network model consists of a deep U-shaped network as its backbone, with branches consisting of U-shaped networks of different depths. Each branch is derived from a corresponding downsampling layer. The loss function for each branch is selected based on the characteristics of the branch model and adjusted by weights. The loss functions of each branch are then combined to form the final model loss function.

[0079] S124. Input the training set into the Unet++ neural network model for training, and save the training results with the loss function set as the indicator that it can no longer be optimized.

[0080] Specifically, the training set is input into the Unet++ neural network model for training, and the feature extraction results of the same layer are interconnected to obtain the training results, which are then saved.

[0081] The training result refers to the output image of land cover changes.

[0082] In this embodiment, if the loss value calculated by the loss function tends to be stable, the trained model can be used for land use change detection. If it does not tend to be stable, the model parameters need to be readjusted and the next round of training needs to be carried out.

[0083] In this embodiment, the training strategy of the remote sensing land cover change detection model is as follows: if the final model loss function does not decrease for three consecutive times, the most recently saved training result is returned, the learning rate is halved, and the Adam optimizer is used as the loss function convergence position search scheme.

[0084] S125. Input the test set into the trained model, extract features at each level, fuse the segmentation results corresponding to the features extracted at each layer during the downsampling process to obtain the output result image, and compare it with the actual result corresponding to the test set. The model that meets the requirements can be used as a remote sensing land cover change detection model.

[0085] S130. Output the detection result image.

[0086] Please see Figures 5 to 7 This paper proposes an algorithm to improve the efficiency of land use change patch updates by introducing deep learning neural networks. The Unet++ neural network model, a direct detection algorithm for land use change areas based on deep learning neural networks, forms a more practical and adaptable land use update working mode. The Unet++ model's prediction results have less noise interference at the detection boundaries and better match the actual boundaries. It can detect the boundary edges of buildings and, to a certain extent, maintain the continuity and uniformity of slender, uniform features such as roads. It has stronger detection capabilities for small-area patches and better connectivity for relatively slender patches such as roads. Some changed patches that cannot be detected in U-shaped networks can be detected in the Unet++ network. Overall, the Unet++ network has a stronger ability to detect changes, and its recall rate is significantly improved compared to the U-shaped neural network.

[0087] The aforementioned remote sensing land cover change detection method involves inputting the image to be detected into a remote sensing land cover change detection model for detection, thereby obtaining the detection result image. The remote sensing land cover change detection model is obtained by training a Unet++ neural network model. The Unet++ model is applied to the remote sensing land cover change detection process to improve the efficiency of land cover patch updating, increase the accuracy, precision, recall, and correctness of land cover change detection, and has strong practicality.

[0088] Figure 8 This is a schematic block diagram of a remote sensing land cover change detection device 300 provided in an embodiment of the present invention. Figure 8 As shown, corresponding to the above-described remote sensing land cover change detection method, the present invention also provides a remote sensing land cover change detection device 300. This remote sensing land cover change detection device 300 includes a unit for performing the above-described remote sensing land cover change detection method, and the device can be configured in a server. Specifically, please refer to... Figure 8 The remote sensing land cover change detection device 300 includes an image acquisition unit 301, a detection unit 302, and an output unit 303.

[0089] Image acquisition unit 301 is used to acquire an image to be detected; detection unit 302 is used to input the image to be detected into a remote sensing land cover change detection model for detection to obtain a detection result image; output unit 303 is used to output the detection result image.

[0090] In one embodiment, the remote sensing land use change detection device 300 further includes a model generation unit, which is used to train a Unet++ neural network model using the annual land use update results as a sample library to obtain a remote sensing land use change detection model.

[0091] In one embodiment, the model generation unit includes a sample library creation subunit, a dataset creation subunit, a construction subunit, a training subunit, a loss function construction subunit, a storage subunit, an optimization subunit, and a testing subunit.

[0092] The sample library creation subunit is used to create a land use change sample library required for deep learning based on the annual land use update results; the dataset creation subunit is used to create training, validation, and test sets based on the land use change sample library; the construction subunit is used to construct the structure of the Unet++ neural network model; specifically, the construction subunit is used to set a loss function that conforms to the usage scenario, set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model, set the initial learning rate and optimization strategy to complete the optimization of the Unet++ model; the training subunit is used to input the training set into the Unet++ neural network model for training, and save the training results with the set loss function being no longer optimizable as an indicator;

[0093] The training subunit is used to input the training set into the Unet++ neural network model for training, and save the training results with the loss function being no longer optimizable as the indicator.

[0094] The test subunit is used to input the test set into the trained model, extract features at each level, fuse the segmentation results corresponding to the features extracted at each layer during the downsampling process to obtain the output result image, and compare it with the actual result corresponding to the test set. The model that meets the requirements can be used as a remote sensing land cover change detection model.

[0095] In one embodiment, a sub-unit is constructed to set a loss function that conforms to the usage scenario, and to set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model, and to set the initial learning rate and optimization strategy; the downsampling channels of the four U-shaped networks are merged, and the four U-shaped models are integrated into one model to obtain the Unet++ neural network model;

[0096] The Unet++ neural network model consists of a deep U-shaped network as its backbone, with branches consisting of U-shaped networks of different depths. Each branch is derived from a corresponding downsampling layer. The loss function for each branch is selected based on the characteristics of the branch model and adjusted by weights. The loss functions of each branch are then combined to form the final model loss function.

[0097] In one embodiment, the training subunit is used to input the training set into the Unet++ neural network model for training, and to connect the feature extraction results of the same layer to obtain training results and save the training results.

[0098] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the aforementioned remote sensing land cover change detection device 300 and its various units can be found in the corresponding descriptions in the foregoing method embodiments. For the sake of convenience and brevity, these details will not be repeated here.

[0099] The aforementioned remote sensing land cover change detection device 300 can be implemented as a computer program, which can, for example... Figure 9 It runs on the computer device shown.

[0100] Please see Figure 9 , Figure 9 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0101] See Figure 9 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0102] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a remote sensing land cover change detection method.

[0103] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0104] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a remote sensing land cover change detection method.

[0105] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0106] The processor 502 is used to run a computer program 5032 stored in the memory to perform the following steps:

[0107] Acquire the image to be detected; input the image to be detected into the remote sensing land cover change detection model for detection to obtain the detection result image; output the detection result image;

[0108] The remote sensing land use change detection model refers to the model obtained by training a Unet++ neural network model using annual land use update results as a sample database.

[0109] In one embodiment, when the processor 502 implements the step of the remote sensing land use change detection model being obtained by training a Unet++ neural network model using annual land use update results as a sample database, the specific steps are as follows:

[0110] A sample library of land use changes required for deep learning was created based on the annual land use update results.

[0111] Training, validation, and test sets were created based on the aforementioned land use change sample database.

[0112] Constructing the structure of the Unet++ neural network model;

[0113] The training set is input into the Unet++ neural network model for training. The training results are saved with the loss function being deemed no longer optimizable as the criterion.

[0114] The test set is input into the trained model, features at each level are extracted, and the segmentation results corresponding to the features extracted at each layer during the downsampling process are fused to obtain the output result image. The result image is then compared with the actual result corresponding to the test set. Models that meet the requirements can be used as remote sensing land cover change detection models.

[0115] The structure for building the Unet++ neural network model includes:

[0116] Set a loss function that suits the use case, and set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model. Set the initial learning rate and optimization strategy to complete the optimization of the Unet++ model.

[0117] In one embodiment, when the processor 502 implements the steps of setting a loss function that conforms to the usage scenario, setting the weights of U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model, setting the initial learning rate and optimization strategy, and completing the optimization steps of the Unet++ model, the specific implementation is as follows:

[0118] Set a loss function that suits the usage scenario, and set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model. Set the initial learning rate and optimization strategy; merge the downsampling channels of the four U-shaped networks, and integrate the four U-shaped models into one model to obtain the Unet++ neural network model.

[0119] The Unet++ neural network model consists of a deep U-shaped network as its backbone, with branches consisting of U-shaped networks of different depths. Each branch is derived from a corresponding downsampling layer. The loss function for each branch is selected based on the characteristics of the branch model and adjusted by weights. The loss functions of each branch are then combined to form the final model loss function.

[0120] In one embodiment, when the processor 502 implements the step of inputting the training set into the Unet++ neural network model for training, and saving the training results with the set loss function being no longer optimizable as an indicator, the specific implementation is as follows:

[0121] The training set is input into the Unet++ neural network model for training, and the feature extraction results of the same layer are interconnected to obtain the training results, which are then saved.

[0122] The training strategy of the remote sensing land cover change detection model is as follows: if the final model loss function does not decrease for three consecutive times, the most recently saved training result is returned, the learning rate is halved, and the Adam optimizer is used as the scheme for finding the convergence position of the loss function.

[0123] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0124] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0125] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform the following steps:

[0126] Acquire the image to be detected; input the image to be detected into the remote sensing land cover change detection model for detection to obtain the detection result image; output the detection result image;

[0127] The remote sensing land use change detection model refers to the model obtained by training a Unet++ neural network model using annual land use update results as a sample database.

[0128] In one embodiment, when the processor executes the computer program to implement the remote sensing land use change detection model, which is obtained by training a Unet++ neural network model using annual land use update results as a sample database, the processor specifically implements the following steps:

[0129] A sample library of land use changes required for deep learning was created based on the annual land use update results.

[0130] Training, validation, and test sets were created based on the aforementioned land use change sample database.

[0131] Constructing the structure of the Unet++ neural network model;

[0132] The training set is input into the Unet++ neural network model for training. The training results are saved with the loss function being deemed no longer optimizable as the criterion.

[0133] The test set is input into the trained model, features at each level are extracted, and the segmentation results corresponding to the features extracted at each layer during the downsampling process are fused to obtain the output result image. The result image is then compared with the actual result corresponding to the test set. Models that meet the requirements can be used as remote sensing land cover change detection models.

[0134] The structure for building the Unet++ neural network model includes:

[0135] Set a loss function that suits the use case, and set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model. Set the initial learning rate and optimization strategy to complete the optimization of the Unet++ model.

[0136] In one embodiment, when the processor executes the computer program to implement the loss function that conforms to the usage scenario, sets the weights of U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model, sets the initial learning rate and optimization strategy, and completes the optimization steps of the Unet++ model, the specific implementation is as follows:

[0137] Set a loss function that suits the usage scenario, and set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model. Set the initial learning rate and optimization strategy; merge the downsampling channels of the four U-shaped networks, and integrate the four U-shaped models into one model to obtain the Unet++ neural network model.

[0138] The Unet++ neural network model consists of a deep U-shaped network as its backbone, with branches consisting of U-shaped networks of different depths. Each branch is derived from a corresponding downsampling layer. The loss function for each branch is selected based on the characteristics of the branch model and adjusted by weights. The loss functions of each branch are then combined to form the final model loss function.

[0139] In one embodiment, when the processor executes the computer program to implement the step of inputting the training set into the Unet++ neural network model for training, saving the training results with the set loss function being no longer optimizable as an indicator, the specific implementation is as follows:

[0140] The training set is input into the Unet++ neural network model for training, and the feature extraction results of the same layer are interconnected to obtain the training results, which are then saved.

[0141] The training strategy of the remote sensing land cover change detection model is as follows: if the final model loss function does not decrease for three consecutive times, the most recently saved training result is returned, the learning rate is halved, and the Adam optimizer is used as the scheme for finding the convergence position of the loss function.

[0142] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0143] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0144] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0145] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0146] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0147] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A remote sensing method for detecting land cover change based on optimized Unet++ neural network deep learning, characterized in that, include: Acquire the image to be detected; The image to be detected is input into the remote sensing land cover change detection model for detection to obtain the detection result image; Output the detection result image; The remote sensing land use change detection model refers to the model obtained by training a Unet++ neural network model using annual land use update results as a sample database. The remote sensing land use change detection model refers to the model obtained by training a Unet++ neural network model using annual land use update results as a sample database, including: Based on the annual land use update results, a sample library of land use changes required for deep learning was created. Training, validation, and test sets were created based on the aforementioned land use change sample database. Constructing the structure of the Unet++ neural network model; The training set is input into the Unet++ neural network model for training. The training results are saved with the loss function being deemed no longer optimizable as the criterion. The test set is input into the trained model, features at each level are extracted, and the segmentation results corresponding to the features extracted at each layer during the downsampling process are fused to obtain the output result image. The result image is then compared with the actual result corresponding to the test set. Models that meet the requirements can be used as remote sensing land cover change detection models. The structure for building the Unet++ neural network model includes: Set a loss function that suits the use case, and set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model. Set the initial learning rate and optimization strategy to complete the optimization of the Unet++ model. The process involves setting a loss function suitable for the usage scenario, setting the weights of different depth U-shaped network branches in the comprehensive loss function of the Unet++ neural network model, setting the initial learning rate and optimization strategy, and optimizing the Unet++ model, including: Set a loss function that suits the usage scenario, and set the weights of the U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model. Set the initial learning rate and optimization strategy; merge the downsampling channels of the four U-shaped networks, and integrate the four U-shaped models into one model to obtain the Unet++ neural network model. The Unet++ neural network model consists of a deep U-shaped network as its backbone, with branches consisting of U-shaped networks of different depths. Each branch is derived from a corresponding downsampling layer. The loss function for each branch is selected based on the characteristics of the branch model and adjusted by weights. The loss functions of each branch are then combined to form the final model loss function.

2. The remote sensing land cover change detection method based on optimized Unet++ neural network deep learning according to claim 1, characterized in that, The step of inputting the training set into the Unet++ neural network model for training, using the inability of the set loss function to be further optimized as an indicator, and saving the training results includes: The training set is input into the Unet++ neural network model for training, and the feature extraction results of the same layer are interconnected to obtain the training results, which are then saved.

3. The remote sensing land cover change detection method based on optimized Unet++ neural network deep learning according to claim 2, characterized in that, The training strategy of the remote sensing land cover change detection model is as follows: if the final model loss function does not decrease for three consecutive times, the most recently saved training result is returned, the learning rate is halved, and the Adam optimizer is used as the scheme for finding the convergence position of the loss function.

4. A remote sensing device for detecting changes in land cover, characterized in that, include: The image acquisition unit is used to acquire the image to be detected; The detection unit is used to input the image to be detected into the remote sensing land cover change detection model for detection, so as to obtain the detection result image; The output unit is used to output the detection result image; It also includes a model generation unit, which is used to train the Unet++ neural network model using the annual land use update results as a sample library to obtain a remote sensing land cover change detection model; The model generation unit includes: The sample library creation sub-unit is used to create a land use change sample library required for deep learning based on the annual land use update results. The dataset creation subunit is used to create training, validation, and test sets based on the land use change sample library. Construct sub-units to build the structure of the Unet++ neural network model; Sub-units are constructed to set loss functions suitable for the usage scenario, and to set the weights of U-shaped network branches of different depths in the comprehensive loss function of the Unet++ neural network model, as well as the initial learning rate and optimization strategy, to optimize the Unet++ model. The downsampling channels of the four U-shaped networks are merged, and the four U-shaped models are integrated into one model to obtain the Unet++ neural network model. The backbone of the Unet++ neural network model is the deepest U-shaped network, and the branches are U-shaped networks of different depths, derived from corresponding downsampling layers. The loss function of each branch is selected based on the characteristics of the branch model and adjusted through weights. The loss functions of each branch are then combined into the final model loss function. The training subunit is used to input the training set into the Unet++ neural network model for training, and save the training results with the loss function being no longer optimizable as the indicator. The test subunit is used to input the test set into the trained model, extract features at each level, fuse the segmentation results corresponding to the features extracted at each layer during the downsampling process to obtain the output result image, and compare it with the actual result corresponding to the test set. The model that meets the requirements can be used as a remote sensing land cover change detection model.

5. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1 to 3.

6. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 3.

Images