Remote sensing image change detection method, system and device and readable storage medium

Abstract

The invention belongs to the technical field of remote sensing change detection, and discloses a remote sensing image change detection method, system and device and a readable storage medium. The remote sensing image change detection method comprises the following steps: acquiring a remote sensing image pair to be subjected to change detection; based on the obtained remote sensing image pair to be subjected to change detection, the trained change detection network is used for detection, and a remote sensing change detection result is obtained; wherein the change detection network comprises a twin network encoder, a single decoder and a post-processing module. According to the technical scheme disclosed by the invention, the problem of insufficient utilization of feature information in the existing technical scheme can be solved, and the accuracy of change detection is improved; in addition, by relieving the overfitting problem existing in the prior art, the generalization ability of the model is effectively enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing change detection, relates to the field of remote sensing image change detection based on deep learning, and in particular to a remote sensing image change detection method, system, device and readable storage medium. Background Art

[0002] With the rapid development of remote sensing technology and earth observation technology, modern satellite systems have the ability to conduct long-term and continuous monitoring of the earth's surface. By analyzing high-resolution image pairs of the same location taken at different time points, remote sensing image change detection technology can reveal changes in surface targets such as land use, vegetation cover, and urban buildings. This information is of immeasurable value to fields such as urban planning, environmental protection, and disaster monitoring and response.

[0003] In recent years, remote sensing change detection methods based on deep learning have attracted widespread attention due to their excellent feature extraction capabilities and learning performance. Compared with traditional methods, deep learning models can automatically learn complex features in images and better adapt to the needs of modern change detection. However, existing deep learning models for change detection often fail to fully utilize the large amount of potential feature information in the network, resulting in insufficient prediction accuracy for the changed area; further, in order to effectively utilize these rich feature information, it is necessary to design more sophisticated feature fusion modules to further improve the accuracy of change detection; in summary, the existing deep learning model network for remote sensing change detection still has a lot of room for improvement in feature fusion utilization. In addition to the problem of feature fusion, the existing technical solutions generally have overfitting in the training process, especially in the training set, which limits the generalization ability of the model; therefore, solving the overfitting problem and improving the generalization performance of the model are also one of the problems that need to be solved in the current research on deep learning change detection networks. Summary of the invention

[0004] The purpose of the present invention is to provide a remote sensing image change detection method, system, device and readable storage medium to solve one or more of the above-mentioned technical problems. In the technical solution disclosed in the present invention, a dense feature connection structure is adopted, which can solve the problem of insufficient utilization of feature information in the prior art solution and improve the accuracy of change detection; in addition, a drop_channel regularized classification layer is connected to the output part of the change detection network, which can effectively enhance the generalization ability of the model by alleviating the overfitting problem existing in the prior art solution.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a remote sensing image change detection method, comprising the following steps: Obtaining a pair of remote sensing images to be changed; Based on the acquired remote sensing image pairs to be detected, the trained change detection network is used to perform detection to obtain remote sensing change detection results; Wherein, the change detection network includes: a twin network encoder, a single decoder and a post-processing module; The twin network encoder is used to input a remote sensing image pair and encode it to obtain encoding features; wherein, in the twin network encoder, each network encoder includes a plurality of serially connected feature extraction modules, and a plurality of feature fusion modules are arranged between two adjacent feature extraction modules; each feature extraction module adopts a dense feature connection structure; The single decoder is used to input the encoded features and map them to obtain a change prediction probability matrix; wherein the single decoder includes a plurality of serially connected feature extraction modules and a drop_channel regular classification layer arranged at the output; the input of the first feature extraction module in the single decoder includes the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder and the features of the minimum size output by the twin network encoder; the input of each feature extraction module in the single decoder except the first feature extraction module includes the output of the feature extraction module of the previous layer and the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder; the drop_channel regular classification layer is used to input the features output by the last feature extraction module in the single decoder and adjust the number of channels to obtain a change prediction probability matrix; The post-processing module is used to input the change prediction probability matrix and perform argmax processing to obtain remote sensing change detection results; Among them, the number of feature extraction modules in each network encoder in the twin network encoder is the same as the number of feature extraction modules in a single decoder and the sizes are corresponding.

[0006] A further improvement of the present invention is that The remote sensing change detection result is a binary image that identifies the changes in the remote sensing image pair.

[0007] A further improvement of the present invention is that The dense feature connection structure is composed of multiple layers of convolutional units connected in series, and the output of each layer of convolutional units is used as the input of the multi-feature fusion module.

[0008] A further improvement of the present invention is that The fusion calculation expression in the multi-feature fusion module is: Y=CBR1(x)+CBR3(cat(CBR1(x),dCBR3(CBR3(CBR1(x))),dCBR3(CBR5(CBR1(x))),dCBR3(CBR7(CBR1(x))))); In the formula, Y represents the fusion result; CBR1 represents the convolution block structure with input channel number c1, output channel number c2, and convolution kernel size 1×1; x represents the input feature of the multi-feature fusion module; c1 represents the number of channels of input feature x; c2 represents the number of channels of the fusion result Y; CBR3 represents the convolution block structure with input channel number 4*c2, output channel number c2, and convolution kernel size 3×3; cat represents splicing along the channel direction; CBR3, CBR5, and CBR7 represent the convolution block structures with input and output both c2 and convolution kernel size 3×3, 5×5, and 7×7 respectively; dCBR3 represents depth-wise convolution with convolution kernel size 3×3.

[0009] A further improvement of the present invention is that The training steps of the change detection network include: Acquire a change detection image dataset; wherein each sample in the change detection image dataset includes a remote sensing image sample pair and a label; Based on the change detection image dataset, the change detection network is trained using the Adam optimizer to minimize the constraint loss function. After reaching the preset convergence condition, the trained change detection network is obtained.

[0010] A further improvement of the present invention is that In the step of obtaining the change detection image dataset, the collected initial remote sensing change detection dataset is first oversampled, and then data enhancement technology is used to obtain the final change detection image dataset.

[0011] A further improvement of the present invention is that In the step of using the Adam optimizer to constrain the loss function to minimize the change detection network, the loss function adopts the cross entropy loss.

[0012] A second aspect of the present invention provides a remote sensing image change detection system, comprising: A data acquisition module, used to acquire remote sensing image pairs to be changed; A detection module is used to perform detection based on the acquired remote sensing image pairs to be detected by using the trained change detection network to obtain remote sensing change detection results; Wherein, the change detection network includes: a twin network encoder, a single decoder and a post-processing module; The twin network encoder is used to input a remote sensing image pair and encode it to obtain encoding features; wherein, in the twin network encoder, each network encoder includes a plurality of serially connected feature extraction modules, and a plurality of feature fusion modules are arranged between two adjacent feature extraction modules; each feature extraction module adopts a dense feature connection structure; The single decoder is used to input the encoded features and map them to obtain a change prediction probability matrix; wherein the single decoder includes a plurality of serially connected feature extraction modules and a drop_channel regular classification layer arranged at the output; the input of the first feature extraction module in the single decoder includes the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder and the features of the minimum size output by the twin network encoder; the input of each feature extraction module in the single decoder except the first feature extraction module includes the output of the feature extraction module of the previous layer and the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder; the drop_channel regular classification layer is used to input the features output by the last feature extraction module in the single decoder and adjust the number of channels to obtain a change prediction probability matrix; The post-processing module is used to input the change prediction probability matrix and perform argmax processing to obtain remote sensing change detection results; Among them, the number of feature extraction modules in each network encoder in the twin network encoder is the same as the number of feature extraction modules in a single decoder and the sizes are corresponding.

[0013] According to a third aspect of the present invention, there is provided an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the remote sensing image change detection method as described in any one of the first aspect of the present invention is implemented.

[0014] According to a fourth aspect of the present invention, a non-transitory computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the remote sensing image change detection method as described in any one of the first aspects of the present invention is implemented.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention discloses a new remote sensing image change detection method, specifically a remote sensing image change detection based on dense multi-feature fusion and regular classification, which adopts a dense feature connection structure, can solve the problem of insufficient utilization of feature information in the existing technical solution, and can improve the accuracy of change detection; in addition, a drop_channel regular classification layer is connected to the output part of the change detection network, and the drop_channel regular classification layer can effectively enhance the generalization ability of the model by alleviating the overfitting problem existing in the existing technical solution. Specifically explanatory, in the twin network encoder, the present invention adopts a dense feature connection structure, the core of the design of this structure is that it can comprehensively and systematically save a large number of feature tensors extracted from the original data, which is different from the mode in which feature information is transmitted layer by layer in the traditional network and may gradually decay. The dense connection ensures that the features of each layer can be directly accessed and used by the subsequent layers. This preservation mechanism effectively solves the problem of insufficient utilization of feature information in the existing method, so that the network can more comprehensively capture the subtle differences and complex patterns in the data, thereby improving the richness and accuracy of feature representation. In addition to preserving feature tensors, the dense feature connection structure also realizes multi-level and multi-dimensional fusion of features through a multi-feature fusion module. A large number of features extracted from each layer will be further integrated and processed by this module to ensure the maximum retention of effective information. In the output part of the change detection network, the present invention connects a drop_channel regular classification layer. The function of this layer is to introduce a regularization mechanism in the model training process, which simulates different network configurations and input changes by randomly discarding some channels (i.e., subsets of feature maps). This random discarding strategy forces the network to learn to rely more on diversified feature representations during training, rather than just relying on certain specific or overly powerful features, which helps the network to show stronger adaptability and generalization capabilities when facing new data or unseen scenarios. The introduction of the Drop_channel regular classification layer also effectively improves the robustness of the model. By simulating different channel loss situations, the network gradually learns during training how to make accurate predictions even when features are incomplete or damaged. This improvement in robustness is particularly important for change detection tasks in practical applications. In actual scenarios, the quality of input data is often affected to a certain extent due to factors such as lighting changes, occlusion, and noise. Robust models can better cope with these challenges and provide more reliable and accurate detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention 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 some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 is a flow chart of a remote sensing image change detection method in an embodiment of the present invention; Figure 2 is a schematic diagram of the overall framework of a change detection network in an embodiment of the present invention; Figure 3 is a schematic diagram of a dense feature connection structure in an embodiment of the present invention; Figure 4 is a structural diagram of a multi-feature fusion module in an embodiment of the present invention; Figure 5 is a schematic diagram of an upsampling and channel adjustment structure in an embodiment of the present invention; Figure 6 is a schematic diagram of a drop_channel (random channel dropout) regular classification layer in an embodiment of the present invention; Figure 7 is a schematic diagram of comparison of parameter quantity-F1 Score on the LEVIR-CD dataset in an embodiment of the present invention; Figure 8 It is a schematic diagram of a remote sensing image change detection system in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments and technical solutions are only part of the embodiments of the present invention, not all of the embodiments.

[0019] All other embodiments obtained by those of ordinary skill in the art without creative work based on the technical solutions disclosed in the embodiments of the present invention belong to the scope of protection of the present invention. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device including a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0020] See also Figure 1 and Figure 2A remote sensing image change detection method provided by an embodiment of the present invention is specifically a remote sensing image change detection method based on dense multi-feature fusion and regular classification, comprising the following steps: Step 1, obtaining a pair of remote sensing images to be detected for changes; explanatory, the pair of remote sensing images is two remote sensing images taken at different times in the same area; Step 2, based on the remote sensing image pair obtained in step 1, using the trained change detection network to perform detection to obtain a remote sensing change detection result; in an exemplary technical solution, the remote sensing change detection result is a binary image that identifies the change of the remote sensing image pair; Wherein, the change detection network includes: a twin network encoder, a single decoder and a post-processing module; The twin network encoder is used to input a remote sensing image pair and encode it to obtain encoding features; wherein, in the twin network encoder, each network encoder includes a plurality of serially connected feature extraction modules, a plurality of feature fusion modules are arranged between two adjacent feature extraction modules, and each feature extraction module adopts a dense feature connection structure; The single decoder is used to input the encoded features and map them to obtain a change prediction probability matrix; wherein the single decoder includes a plurality of serially connected feature extraction modules and a drop_channel regular classification layer; the input of the first feature extraction module includes the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder and the features of the minimum size output by the twin network encoder; the input of each feature extraction module except the first feature extraction module includes the output of the feature extraction module of the previous layer and the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder; the drop_channel regular classification layer is used to input the features output by the last feature extraction module and adjust the number of channels to obtain a change prediction probability matrix; The post-processing module is used to input the change prediction probability matrix and perform argmax processing to obtain remote sensing change detection results; Among them, the number of feature extraction modules in each network encoder in the twin network encoder is the same as the number of feature extraction modules in a single decoder and the sizes are corresponding.

[0021] In the technical solution provided by the embodiment of the present invention, the dense feature connection structure used in the twin network encoder will save a large number of feature tensors extracted by the deep learning network from the original data, which solves the problem of insufficient utilization of feature information in the existing methods. At the same time, a large number of features in each layer will be further fused through the multi-feature fusion module to achieve the purpose of preserving effective information as much as possible; in addition, the drop_channel regular classification layer designed by connecting the output part of the change detection network can effectively enhance the generalization ability of the model.

[0022] See also Figure 3 ,In one embodiment of the present invention, the dense feature connection structure is composed of multiple layers of convolutional units in series, and the output of each layer of convolutional units is used as the input of the multi-feature fusion module; For example, taking a dense feature connection structure composed of three layers of convolutional units connected in series as an example, its expression is: Y=DownSample(Fuse(cat(Conv3(Conv2(Conv1(x)),Conv2(Conv1(x)),Conv1(x))))) Where DownSample() is the downsampling operator; Fuse is the proposed multi-feature fusion module; Conv1, Conv2, and Conv3 are of the same structure, and their calculation process is: Conv(x)=ReLU(Bn(Conv(x))); In the formula, Conv, Bn, and ReLU are the common convolution, normalization, and activation structures in volume networks, respectively.

[0023] See also Figure 4 In one embodiment of the present invention, the change detection network includes a multi-feature fusion module and is used to fuse a large number of features obtained from the dense feature connection structure network. The calculation process is as follows: The features obtained by the densely connected network are input into the multi-feature fusion module, which performs information fusion through a series of calculations. The fusion calculation expression is: Y=CBR1(x)+CBR3(cat(CBR1(x),dCBR3(CBR3(CBR1(x))),dCBR3(CBR5(CBR1(x))),dCBR3(CBR7(CBR1(x))))); In the formula, Y represents the fusion result; x represents the input feature of the multi-feature fusion module; CBR1 represents the convolution block structure with the input channel number c1, the output channel number c2, and the convolution kernel size of 1×1 (i.e., Conv, Bn, and ReLU are combined); CBR3 represents the convolution block structure with the input channel number 4*c2, the output channel number c2, and the convolution kernel size of 3×3; CBR3, CBR5, and CBR7 represent the convolution block structures with the input and output both c2 and the convolution kernel sizes of 3×3, 5×5, and 7×7, respectively; dCBR3 represents the depth-wise convolution with a convolution kernel size of 3×3; cat represents splicing along the channel direction; c1 represents the number of channels of the input feature x, and c2 represents the number of channels of the fusion result Y.

[0024] In one embodiment of the present invention, a change detection data set may be first obtained, and then the change detection network may be trained using sample data in the change detection data set, and the network weights may be updated under the constraints of the loss function, and finally the network with updated weights may be used for change detection; Among them, a change detection image dataset taken in the same area at different times is obtained, and oversampling is used to increase the data volume in the dataset; the change detection network is trained by minimizing the loss function using the Adam optimizer.

[0025] The specific exemplary technical solution also includes: Preprocessing, the preprocessing steps include: oversampling the collected remote sensing change detection data set; wherein, for each pair of remote sensing images used for change detection and their corresponding labels, the method of the embodiment of the present invention cuts them into sample blocks of size 256×256 pixels; further, in order to enhance the continuity and coverage of the samples, each sample block is designed with an overlapping area of ​​128 pixels with other sample blocks to ensure that the model has the opportunity to see more and more diverse data.

[0026] In addition, in order to improve the generalization ability of the model, an embodiment of the present invention adopts data enhancement technology; wherein, for each pair of dual-phase change detection images, the method of the present invention performs data enhancement by random horizontal flipping, random vertical flipping and random rotation (90, 180, 270 degrees); by adjusting the probability parameters, it is ensured that the probability of each transformation occurring is equal, including the identity transformation, so that the model can learn the characteristics of remote sensing images from multiple angles and directions.

[0027] During the training process, the method of the embodiment of the present invention uses the Adam optimizer for parameter optimization, where the momentum parameter is set to 5e-3; in addition, the training cycle (epoch) is set to 200, the batch size is set to 8, and the learning rate is set to 0.0025. The settings of these hyperparameters are intended to ensure that the model can converge efficiently and stably.

[0028] In a specific embodiment of the present invention, the calculation process of the change detection network of the remote sensing image includes the following steps: Step 1: feature information extraction; wherein, The twin network encoder and single decoder are both implemented using pure convolutional network structures; the twin network encoder adopts a dense feature connection structure design, such as Figure 3 As shown in the figure, each network encoder contains four stages of convolutional layers (i.e., feature extraction modules), and the number of convolutional blocks used in each stage is [2,2,3,3]; each convolutional block is composed of a basic CBR (Convolution-Batch Normalization-ReLU) module, i.e., a 3×3 convolutional layer followed by BatchNorm and ReLU. According to the requirements of the dense feature connection structure, except for the last stage, the feature tensors of each stage are saved, and the number of feature tensors saved in each stage is [2,2,3]. These feature tensors are input into the multi-feature fusion module (MFF, Multi Feature Fusion), and the fusion result is used as the first feature input of the next stage. The specific structure of the multi-feature fusion module is shown in the figure. Figure 4 As shown, the calculation steps are as follows: the multiple features obtained are spliced ​​along the channel direction and copied 5 times; each splicing result first passes through a 1×1 CBR structure to ensure that the number of output bands is aligned with the number of bands required by the fusion module. For example, if the MFF structure is used to fuse the features of the second stage with a band number of C2, the output band number of the 1×1 convolution will be equal to the band number C3 of the third stage feature. 1×3+3×1 refers to the concatenation of CBR structures with convolution kernel sizes of 1×3 and 3×1. The 3×3 structure in the green box refers to a depth-wise convolution with a convolution kernel size of 3×3, and the convolution is also followed by BatchNorm and ReLU structures. In addition to being input into the MFF module, the feature maps obtained by each CBR module of the encoder must also be retained as intermediate variables. For a dual-stream encoder (i.e., a twin network encoder), each pair of corresponding feature maps is subtracted to obtain a differential feature map, corresponding to Figure 2 Medium grey cube.

[0029] Step 2: high-resolution feature map acquisition; wherein, The lowest resolution feature map in a single decoder is equal to the lowest resolution differential feature map; the input feature map of each other stage also adopts the design of dense connection structure, such as Figure 5 As shown in the figure, the feature map of each stage takes the feature map of the previous stage and all the differential features of the encoder at the corresponding scale as input; the feature map of the previous stage is upsampled after channel adjustment through the CBR module, and then spliced ​​with all the feature maps of the same scale in the differential feature map along the channel direction. This part of the calculation corresponds to Figure 2 The U module in the decoder is then passed through a specified number of CBR modules to obtain the output of this stage. The number of CBR modules contained in each stage of the decoder is [3,3,2,2] respectively.

[0030] Step 3: Output the change graph; where: The specific structure of the drop_channel regular classification layer is as follows Figure 6 As shown in the figure, the feature map obtained by the decoder first passes through a CBR module to adjust the number of channels to four times the original number of channels, and the subsequent forward propagation distinguishes between the training process and the inference process; among them, Reasoning process: A 3×3 convolution is directly followed to adjust the number of channels to 2, and then the change prediction is obtained through argmax mapping.

[0031] Training process: Generate a random B×C×1×1 tensor, where B is the batch size and C is the number of channels after expansion, and each element is 0 with a probability of 0.5 and 1 with a probability of 0.5. The random tensor is multiplied by the tensor after the expansion channel and then multiplied by 2, followed by a 3×3 convolution to adjust the number of channels to 2, and then mapped by argmax to obtain the change prediction.

[0032] Loss function part: The loss function uses cross entropy loss, which is used to calculate the change detection error of the network on the training set during network training, quantified as a loss term, and then back-propagated to update the network parameters.

[0033] In summary, in order to address the problem of insufficient utilization of feature information in existing methods, the present invention uses a dense feature connection structure to transform the deep learning network. This design will save a large number of feature tensors extracted from the original data by the deep learning network to preserve effective information as much as possible; and a multi-feature fusion module is used to fuse a large number of features obtained from the dense feature connection scheme, so that the change detection network can learn effective features and information as much as possible. In order to address the overfitting problem of existing methods on the training set, the present invention designs a drop_channel classification layer, which can effectively alleviate the problem of overfitting of the model on the training set.

[0034] In a specific embodiment of the present invention, an ablation experiment was performed on the LEVIR-CD dataset to illustrate the effect of the improved technical means of the embodiment of the present invention, as shown in Table 1 and Table 2.

[0035] Table 1. Dense multi-feature fusion ablation experimental results

[0036] Table 2. Drop channel regular classification layer ablation experimental results

[0037] From the ablation experiment results in Table 1 and Table 2, it can be seen that the dense multi-feature fusion module and the Drop channel regular classification module improve the F1score index of the Baseline model on the LEVIR-CD dataset by 0.93% and 0.31%, respectively, which fully demonstrates the effectiveness and significant progress of the improved technical means of the present invention.

[0038] Comparative experiments were conducted on the LEVIR-CD and WHU-CD datasets commonly used in remote sensing change detection with other change detection networks to illustrate the superiority of the present invention. The comparative experimental results are shown in Tables 3 and 4.

[0039] Table 3. Experimental results of LEVIR-CD dataset

[0040] Table 4. Experimental results of WHU-CD dataset

[0041] Through comparative experiments with other mainstream methods on the LEVIR-CD and WHU-CD change detection datasets, it can be seen that the method proposed in the present invention makes the F1Score exceed the second place by 0.65% and 0.86% on LEVIR-CD and WHU-CD respectively, which illustrates the superiority of the change detection method proposed in the present invention.

[0042] The accuracy and model parameter quantity of the present invention are compared with other methods on LEVIR-CD to illustrate the superiority of the present invention. The comparison results are shown in Table 5.

[0043] Table 5. Comparison of network parameters and accuracy (LEVIR-CD)

[0044] Based on Table 5 and Figure 7By comparing the network parameter amount and change detection accuracy, it can be seen that the method proposed in the embodiment of the present invention achieves better change detection accuracy while maintaining a smaller number of parameters, which fully demonstrates the superiority of the change detection method proposed in the present invention.

[0045] In summary, the embodiment of the present invention discloses a remote sensing change detection method based on dense multi-feature fusion and regular classification, which innovatively improves the classic twin convolutional network architecture to meet the needs of modern remote sensing change detection. The method of the present invention effectively solves the problem of insufficient utilization of dual-phase feature information in the prior art by designing a dense feature connection scheme and using a fusion module to fuse a large amount of information obtained by the dense feature connection scheme, thereby improving the accuracy of change detection. In addition, the present invention also designs a drop_channel regular classification layer, which strengthens the predictive ability of the model by alleviating the overfitting problem existing in the existing method. Furthermore, the overall network structure of the present invention adopts a pure convolutional network design, which limits the amount of network parameters while ensuring the powerful learning ability of the network, increasing the application prospects and practicality of the design.

[0046] The following are device embodiments of the present invention, which can be used to implement the method embodiments of the present invention. For details not disclosed in the device embodiments, please refer to the method embodiments of the present invention.

[0047] See also Figure 8 In an embodiment of the present invention, a remote sensing image change detection system is provided, comprising: A data acquisition module, used to acquire remote sensing image pairs to be changed; A detection module is used to perform detection based on the acquired remote sensing image pairs to be detected by using the trained change detection network to obtain remote sensing change detection results; Wherein, the change detection network includes: a twin network encoder, a single decoder and a post-processing module; The twin network encoder is used to input a remote sensing image pair and encode it to obtain encoding features; wherein, in the twin network encoder, each network encoder includes a plurality of serially connected feature extraction modules, and a plurality of feature fusion modules are arranged between two adjacent feature extraction modules; each feature extraction module adopts a dense feature connection structure; The single decoder is used to input the encoded features and map them to obtain a change prediction probability matrix; wherein the single decoder includes a plurality of serially connected feature extraction modules and a drop_channel regular classification layer arranged at the output; the input of the first feature extraction module in the single decoder includes the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder and the features of the minimum size output by the twin network encoder; the input of each feature extraction module in the single decoder except the first feature extraction module includes the output of the feature extraction module of the previous layer and the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder; the drop_channel regular classification layer is used to input the features output by the last feature extraction module in the single decoder and adjust the number of channels to obtain a change prediction probability matrix; The post-processing module is used to input the change prediction probability matrix and perform argmax processing to obtain remote sensing change detection results; Among them, the number of feature extraction modules in each network encoder in the twin network encoder is the same as the number of feature extraction modules in a single decoder and the sizes are corresponding.

[0048] In one embodiment of the present invention, a computer device is provided, the computer device includes a processor and a memory, the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used to perform the operation of the remote sensing image change detection method.

[0049] In one embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM (Random Access Memory) memory, or a non-volatile memory (non-volatile memory), such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the remote sensing image change detection method in the above embodiment.

[0050] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) containing computer-usable program codes.

[0051] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0052] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0053] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A remote sensing image change detection method, characterized in that: The following steps are involved: Obtaining a pair of remote sensing images to be changed; Based on the acquired remote sensing image pairs to be detected, the trained change detection network is used to perform detection to obtain remote sensing change detection results; Wherein, the change detection network includes: a twin network encoder, a single decoder and a post-processing module; The twin network encoder is used to input a remote sensing image pair and encode it to obtain encoding features; wherein, in the twin network encoder, each network encoder includes a plurality of serially connected feature extraction modules, and a plurality of feature fusion modules are arranged between two adjacent feature extraction modules; each feature extraction module adopts a dense feature connection structure; The single decoder is used to input the encoded features and map them to obtain a change prediction probability matrix; wherein the single decoder includes a plurality of serially connected feature extraction modules and a drop_channel regular classification layer arranged at the output; the input of the first feature extraction module in the single decoder includes the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder and the features of the minimum size output by the twin network encoder; the input of each feature extraction module in the single decoder except the first feature extraction module includes the output of the feature extraction module of the previous layer and the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder; the drop_channel regular classification layer is used to input the features output by the last feature extraction module in the single decoder and adjust the number of channels to obtain a change prediction probability matrix; The post-processing module is used to input the change prediction probability matrix and perform argmax processing to obtain remote sensing change detection results; Among them, the number of feature extraction modules in each network encoder in the twin network encoder is the same as the number of feature extraction modules in a single decoder and the sizes are corresponding.

2. A remote sensing image change detection method according to claim 1, characterized in that: The remote sensing change detection result is a binary image that identifies the changes in the remote sensing image pair.

3. The remote sensing image change detection method according to claim 1, characterized in that: The dense feature connection structure is composed of multiple layers of convolutional units connected in series, and the output of each layer of convolutional units is used as the input of the multi-feature fusion module.

4. A remote sensing image change detection method according to claim 3, characterized in that: The fusion calculation expression in the multi-feature fusion module is: Y=CBR1(x)+CBR3(cat(CBR1(x),dCBR3(CBR3(CBR1(x))),dCBR3(CBR5(CBR1(x))),dCBR3(CBR7(CBR1(x))))); In the formula, Y represents the fusion result; CBR1 represents the convolution block structure with input channel number c1, output channel number c2, and convolution kernel size 1×1; x represents the input feature of the multi-feature fusion module; c1 represents the number of channels of input feature x; c2 represents the number of channels of the fusion result Y; CBR3 represents the convolution block structure with input channel number 4*c2, output channel number c2, and convolution kernel size 3×3; cat represents splicing along the channel direction; CBR3, CBR5, and CBR7 represent the convolution block structures with input and output both c2 and convolution kernel size 3×3, 5×5, and 7×7 respectively; dCBR3 represents depth-wise convolution with convolution kernel size 3×3.

5. The remote sensing image change detection method according to claim 1, characterized in that: The training steps of the change detection network include: Acquire a change detection image dataset; wherein each sample in the change detection image dataset includes a remote sensing image sample pair and a label; Based on the change detection image dataset, the change detection network is trained using the Adam optimizer to minimize the constraint loss function. After reaching the preset convergence condition, the trained change detection network is obtained.

6. A remote sensing image change detection method according to claim 5, characterized in that: In the step of obtaining the change detection image dataset, the collected initial remote sensing change detection dataset is first oversampled, and then data enhancement technology is used to obtain the final change detection image dataset.

7. A remote sensing image change detection method according to claim 5, characterized in that: In the step of using the Adam optimizer to constrain the loss function to minimize the change detection network, the loss function adopts the cross entropy loss.

8. A remote sensing image change detection system, characterized in that: include: A data acquisition module, used to acquire remote sensing image pairs to be changed; A detection module is used to perform detection based on the acquired remote sensing image pairs to be detected by using the trained change detection network to obtain remote sensing change detection results; Wherein, the change detection network includes: a twin network encoder, a single decoder and a post-processing module; The twin network encoder is used to input a remote sensing image pair and encode it to obtain encoding features; wherein, in the twin network encoder, each network encoder includes a plurality of serially connected feature extraction modules, and a plurality of feature fusion modules are arranged between two adjacent feature extraction modules; each feature extraction module adopts a dense feature connection structure; The single decoder is used to input the encoded features and map them to obtain a change prediction probability matrix; wherein the single decoder includes a plurality of serially connected feature extraction modules and a drop_channel regular classification layer arranged at the output; the input of the first feature extraction module in the single decoder includes the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder and the features of the minimum size output by the twin network encoder; the input of each feature extraction module in the single decoder except the first feature extraction module includes the output of the feature extraction module of the previous layer and the differential features obtained by subtracting two feature extraction modules of the same size in the twin network encoder; the drop_channel regular classification layer is used to input the features output by the last feature extraction module in the single decoder and adjust the number of channels to obtain a change prediction probability matrix; The post-processing module is used to input the change prediction probability matrix and perform argmax processing to obtain remote sensing change detection results; Among them, the number of feature extraction modules in each network encoder in the twin network encoder is the same as the number of feature extraction modules in a single decoder and the sizes are corresponding.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the remote sensing image change detection method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the remote sensing image change detection method according to any one of claims 1 to 7 is implemented.

Images