Dual-temporal remote sensing image change detection method, device, equipment and medium

Abstract

The application belongs to the technical field of remote sensing image processing, and discloses a dual-time-phase remote sensing image change detection method, device, equipment and medium, which comprises the following steps: inputting remote sensing data set into a change detection network, and training basic model parameters through a minimum cross-entropy loss function. In the trained model, a perceiver network is embedded before each scale regularization layer of the encoder, the correlation and difference of dual-time-phase features are modeled, and weight scores are output. The dual-time-phase weight scores learned by the perceiver are used as the reference proportion of the reconstruction of the regularization layer, and the dynamic construction of the learnable affine parameter is guided. In the test stage, the perceiver network dynamically models the temporal feature distribution under the interference condition, generates reconstruction parameters for each phase to update the affine parameter of the regularization layer, and then introduces a loss function based on the minimum prediction entropy of the pixel level to update the perceiver network parameters. The application dynamically adapts to the interference, accurately models the time correlation, improves the change detection precision in complex scenes, and provides an effective technical solution for remote sensing monitoring.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a method, apparatus, device and medium for detecting changes in dual-temporal remote sensing images. Background Technology

[0002] Remote sensing change detection plays a crucial role in monitoring and understanding dynamic changes on the Earth's surface by analyzing dual-temporal remote sensing images. As a core technology in the field of Earth observation, it supports many important applications, including surface change analysis, post-disaster damage assessment, sustainable resource management, and urban sprawl monitoring. With continuous technological advancements, an increasing number of change detection methods have been proposed, significantly enhancing the breadth and effectiveness of remote sensing applications.

[0003] Due to the complexity of real-world conditions, dual-temporal remote sensing images are often affected by various interference factors, including natural variations and sensor noise introduced by the imaging and modulation processes, thus posing numerous challenges in practical applications. In recent years, some test-time adaptive methods have dynamically updated their models during the inference phase through self-training mechanisms to enhance their robustness under unknown interference conditions. These test-time adaptive methods typically introduce a teacher-student model, allowing the student model to mimic the teacher's stable predictions to achieve self-adaptation; other studies have introduced active sample selection mechanisms, choosing reliable and representative samples from the test data for model updates to reduce the risk of negative transfer caused by noise interference. Although the aforementioned test-time adaptive methods have achieved some success in natural image tasks, their direct application to remote sensing change detection tasks still has significant limitations.

[0004] The shortcomings and deficiencies of the existing technology are as follows:

[0005] 1. Most existing test-time adaptive methods lack modeling of the dynamic evolution relationship between two-phase images, resulting in an inability to effectively capture temporal clues of change. In change detection tasks, features of images at different time points exhibit strong correlations, and this temporal dependency is crucial for distinguishing between real and pseudo-changes. Ignoring temporal information may not only weaken the model's ability to respond to real changes but also easily lead to misjudgments of natural seasonality or illumination changes.

[0006] 2. In complex remote sensing scenarios, imaging interference introduces significant temporal distribution differences, causing shifts in the characteristic representation of the same ground feature at different times. This shift may not reflect actual surface changes but can mislead models into making incorrect predictions. Existing methods lack targeted mechanisms for modeling temporal differences, making it difficult to effectively identify interfered areas, especially in noise-dominated scenarios where false detections or missed detections are prone to occur. Summary of the Invention

[0007] To address the aforementioned issues, this application provides a method, apparatus, device, and medium for detecting changes in dual-temporal remote sensing images, which can significantly improve the accuracy of change detection under interference environments. Specifically, this application enhances the perception of change information by modeling the complex relationship between the two temporal phases; it introduces a regularized parameter reconstruction strategy to mitigate errors caused by image interference, addressing various types of noise present in real-world scenes; and it combines a robust test-time optimization mechanism to significantly improve the stability and robustness of model updates.

[0008] According to the first technical solution of this application, a method for detecting changes in dual-temporal remote sensing images is provided, the method comprising:

[0009] A remote sensing dataset is input into a change detection network comprising an encoder and a decoder. The change detection network is trained by minimizing the cross-entropy loss function to obtain a base model. The remote sensing dataset includes dual-temporal remote sensing data. The encoder is used to extract deep features from the dual-temporal remote sensing data, and the decoder is used to map the deep features into change detection results.

[0010] Based on the aforementioned basic model, a perceptron network is embedded before the regularization layer at each scale of the encoder; the perceptron network takes the bi-temporal features extracted by the encoder as input, learns the correlation and difference between the two temporal features, and outputs the corresponding weight scores.

[0011] The bi-temporal weight score output by the perceptron network is used as the reference weight for the reconstruction of the regularization layer, guiding the dynamic construction of learnable affine parameters and generating a set of dynamic regularization parameters for each temporal phase.

[0012] During the testing phase, the interfering dual-temporal remote sensing data is input into the base model in batches; the perceptron network dynamically models the temporal feature distribution under interference conditions, and generates corresponding reconstruction parameters for each temporal phase to update the affine parameters in the regularization layer; the dual-temporal features at different scales output by the encoder are processed by the decoder to generate the final change detection results;

[0013] A loss function based on minimizing pixel-level prediction entropy is introduced to encourage the model to make deterministic change predictions at each spatial location and update the parameters of the perceptron network.

[0014] Furthermore, in the process of inputting the remote sensing dataset into a change detection network comprising an encoder and a decoder, and training the change detection network by minimizing the cross-entropy loss function to obtain the base model:

[0015] The encoder responds to the input bi-temporal remote sensing data, converting it into a sequence representation and inputting it into a multi-layer encoder module. The multi-layer encoder module then obtains bi-temporal features at different scales. Each layer encoder module calculates the corresponding scale's bi-temporal features using the following formula:

[0016]

[0017] In the formula, and Block represents the first and second phase features of the l-th layer, respectively. l This represents the encoder module of layer l. and These represent the first and second phase characteristics of the (l-1)th layer, respectively.

[0018] The decoder decodes the difference between two temporal features to obtain the change features at the corresponding scale. It then unifies and fuses the change features at each scale to the same scale through an upsampling operation. After upsampling by two-dimensional transpose convolution, the result is fed into a linear mapping to generate the final change mask map.

[0019] Furthermore, the perceptron network takes the bi-temporal features extracted by the encoder as input, learns the correlation and difference between the two temporal features, and outputs the corresponding weight scores, including:

[0020] The absolute difference between two temporal features is calculated as the difference feature using the following formula:

[0021]

[0022] In the formula, Features of difference;

[0023] Based on two temporal features, the context-enhanced features are calculated using the following formula:

[0024]

[0025] In the formula, For the context enhancement features of layer l, ⊙ represents the element-wise multiplication operation. For the i-th phase feature of the l-th layer, The j-th phase feature of the l-th layer;

[0026] The two temporal features, context-enhanced features, and differential features are concatenated along the channel dimension using the following formula, and used as the input features for the perceptron network:

[0027]

[0028] In the formula, Let C represent the input features of the perceptron network, and C be the feature concatenation operation. These are context-enhanced features constructed based on two time phases, t1 and t2.

[0029] The sensor network responds to the input. The corresponding weighted score is output using the following formula:

[0030]

[0031] h2 = Sigmoid(h1W2 + b2)

[0032]

[0033] In the formula, h1 represents the output feature of the first layer linear transformation of the perceptron network after Sigmoid activation, W1 represents the learnable weight matrix of the first layer of the perceptron network, b1 represents the bias parameters of the first layer of the perceptron network, h2 represents the output feature of the second layer linear transformation of the perceptron network after Sigmoid activation, W2 represents the learnable weight matrix of the second layer of the perceptron network, and b2 represents the bias parameters of the second layer of the perceptron network. and These are the first weight score and the second weight score, respectively, and Split is the splitting function.

[0034] Furthermore, the bi-temporal weight scores output by the perceptron network are used as reference weights for the reconstruction of the regularization layer, guiding the dynamic construction of learnable affine parameters in the following ways:

[0035] The tensor of the i-th phase feature is input into the regularization layer, and the normalized feature is obtained by the normalization process of the regularization layer.

[0036] Using the biphase weight scores and auxiliary learning parameters output by the perceptron network, the temporal affine parameters are calculated; the temporal affine parameters are used to perform affine transformation on the normalized features.

[0037] Furthermore, using the biphase weight scores and auxiliary learning parameters output by the perceptron network, the temporal affine parameters are calculated using the following formula:

[0038]

[0039]

[0040] In the formula, γ is the reconstruction parameter of the i-th phase feature. l and β l These are shared scaling and offset parameters. and These are auxiliary learning parameters. and These are the first weight score and the second weight score, respectively.

[0041] Furthermore, the loss function based on minimizing pixel-level prediction entropy is expressed as:

[0042]

[0043] In the formula, Let B be the batch size, H be the image height, W be the image width, b be the batch index, i be the pixel index in the image height direction, and j be the pixel index in the image width direction. Let be the predicted entropy of pixel (i,j).

[0044] Furthermore, the formula for calculating the prediction entropy of pixel (i,j) is as follows:

[0045]

[0046] In the formula, P i,j This represents the probability that pixel (i,j) belongs to the change category.

[0047] Furthermore, the formula for calculating the prediction entropy of pixel (i,j) is as follows:

[0048]

[0049] In the formula, P i,j This represents the probability that pixel (i,j) belongs to the change category.

[0050] According to the second technical solution of this application, a dual-temporal remote sensing image change detection device is provided, the device comprising:

[0051] The basic model building module is configured to input a remote sensing dataset into a change detection network including an encoder and a decoder, and train the change detection network by minimizing the cross-entropy loss function to obtain a basic model; wherein, the remote sensing dataset includes dual-temporal remote sensing data, the encoder is used to extract deep features from the dual-temporal remote sensing data, and the decoder is used to map the deep features into change detection results;

[0052] The perceptron embedding module is configured to embed a perceptron network before the regularization layer at each scale of the encoder based on the base model; the perceptron network takes the bi-temporal features extracted by the encoder as input, learns the correlation and difference between the two temporal features, and outputs the corresponding weight scores.

[0053] The dynamic parameter construction module is configured to use the bi-temporal weight scores output by the perceptron network as the reference weights for the reconstruction of the regularization layer, guide the dynamic construction of learnable affine parameters, and generate a set of dynamic regularization parameters for each temporal phase.

[0054] The change detection inference module is configured to input interfering dual-temporal remote sensing data into the base model in batches during the testing phase; the perceptron network dynamically models the temporal feature distribution under interference conditions and generates corresponding reconstruction parameters for each temporal phase to update the affine parameters in the regularization layer; the dual-temporal features at different scales output by the encoder are processed by the decoder to generate the final change detection result;

[0055] The perceptron optimization module is configured to introduce a loss function based on minimizing pixel-level prediction entropy to encourage the model to make deterministic change predictions at each spatial location and update the parameters of the perceptron network.

[0056] According to the third technical solution of this application, an electronic device is provided, the electronic device comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the method described above.

[0057] According to the fourth technical solution of this application, a non-transitory computer-readable storage medium storing instructions is provided, which, when executed by a processor, performs the method described above.

[0058] The dual-temporal remote sensing image change detection methods, apparatuses, equipment, and media according to the various schemes of this application have at least the following technical effects:

[0059] 1) The basic model is adapted for temporal feature extraction. After embedding the perceptron network, the dynamic evolution relationship of the dual-temporal features is further learned. Unlike traditional methods that ignore temporal dependence, this application guides the construction of affine parameters through the output weights of the perceptron, enabling the model to accurately capture the dynamic evolution clues of the temporal phases, providing a temporal basis for judging real changes, thereby solving the defect of "misjudging natural changes".

[0060] 2) By utilizing the temporal weights generated by the perceptron, affine parameters are dynamically constructed to adapt to the interference scene. Addressing the issue of temporal distribution differences caused by imaging interference, this application customizes regularization parameters for each temporal phase, making feature normalization and transformation more closely match the temporal characteristics under interference (such as cloud cover and noise pollution). This overcomes the bottleneck of traditional static parameters being difficult to adapt to complex scenes, enabling effective identification of interference areas (such as shadow-covered areas).

[0061] 3) By forcing the model to output deterministic predictions through pixel-level entropy loss and combining the ability of dynamic parameters to adapt to interference, optimizations are made in terms of prediction confidence and scene adaptability. Compared with the frequent false detection / false detection problems in noisy scenes of existing technologies, this application significantly improves the clarity of boundaries and the accuracy of interference areas, providing more reliable technical support for remote sensing change detection and covering practical application scenarios such as land monitoring and disaster assessment.

[0062] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0063] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0064] Figure 1 A flowchart illustrating a dual-temporal remote sensing image change detection method provided in this application embodiment;

[0065] Figure 2 A flowchart illustrating the network architecture of the dual-temporal remote sensing image change detection method provided in this application embodiment;

[0066] Figure 3 This is a structural diagram of a dual-temporal remote sensing image change detection device provided in an embodiment of this application. Detailed Implementation

[0067] To enable those skilled in the art to better understand the technical solution of this application, the application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] One aspect of this application provides a method for detecting changes in dual-temporal remote sensing images. For example... Figure 1 As shown, Figure 1 This is a flowchart illustrating a method for detecting changes in dual-temporal remote sensing images, as provided in an embodiment of this application. The method includes the following steps S10-S50.

[0069] S10: Input the remote sensing dataset into a change detection network that includes an encoder and a decoder. Train the change detection network by minimizing the cross-entropy loss function to obtain the base model. The remote sensing dataset includes dual-temporal remote sensing data. The encoder is used to extract deep features from the dual-temporal remote sensing data, and the decoder is used to map the deep features to change detection results.

[0070] In this embodiment, the encoder can be a Transformer encoder, and the decoder can be an MLP decoder. In step S10, the remote sensing dataset is input into a change detection network consisting of a Transformer encoder and an MLP decoder. The model is trained by minimizing the cross-entropy loss function to obtain basic model parameters. The Transformer encoder is used to extract deep features from the dual-temporal remote sensing data. The MLP decoder first decodes the difference between the two temporal features to obtain the change features at the corresponding scale, and then unifies the change features at each scale to the same scale through upsampling. After feature fusion, preliminary change detection results are generated through two-dimensional transposed convolution upsampling and linear mapping, providing basic model support for subsequent processing.

[0071] In some embodiments, the encoder employs a Transformer as a feature extractor, using a multi-layer attention mechanism to model the features of the images at both temporal phases, sharing the same set of attention parameters at each layer to unify the modeling of bi-temporal information. After the images are converted into sequence representations, they are input into a Transformer containing a four-layer encoder module. Through feature extraction at each block, four levels of bi-temporal feature representations are obtained sequentially:

[0072]

[0073] In the formula, and Block represents the first and second phase features of the l-th layer, respectively. l This represents the encoder module of layer l. and These represent the first and second phase features of the (l-1)th layer, respectively. In the encoder, l = 1, 2, 3, 4, indicating a four-layer encoder module.

[0074] In some embodiments, the decoder first decodes the difference between two temporal features to obtain the change features at the corresponding scale; then, the change features at each scale are unified to the same scale through an upsampling operation, and feature fusion is performed; finally, the fused change features are fed into the Softmax function to generate the final change mask map as the change detection result.

[0075] Specifically, the encoder uses bilinear interpolation to convert the feature difference map at each scale. Upsampling to a uniform spatial scale H / 4×W / 4:

[0076]

[0077] In the formula, These are feature maps of differences at different scales; Upsample is the upsampling operation.

[0078] All upsampled difference feature maps are concatenated along the channel dimension and fused using a linear mapping (MLP) layer to obtain a unified feature representation F:

[0079]

[0080] In the formula, Concat represents the concatenation operation. These represent feature difference maps at the first, second, third, and fourth layer scales, respectively.

[0081] Finally, the upsampled features are fed into a linear mapping (MLP) layer to predict the final change mask map:

[0082]

[0083] In the formula, LinearMASK is a linear mapping layer, and this embodiment uses a single linear mapping layer (MLP).

[0084] The final output size of the transformed mask is H×W×N cls , where N cls =2 indicates the number of categories, which include both variable and non-variable categories.

[0085] S20: Based on the base model, a perceptron network is embedded before the regularization layer at each scale of the encoder; the perceptron network takes the bi-temporal features extracted by the encoder as input, learns the correlation and difference between the two temporal features, and outputs the corresponding weight scores.

[0086] In this embodiment, step S20 is based on the basic model trained in step S10, and a perceptron network is embedded before the regularization layer at each scale of the encoder. The network takes the two temporal features extracted by the Transformer encoder in step S10 as input, calculates the absolute difference between the two temporal features to obtain the difference features, calculates the cosine similarity and weights the current features to obtain the context enhancement features, and concatenates the relevant features and learns the correlation and difference between the two temporal features through linear mapping and nonlinear activation, and outputs the corresponding weight scores to provide a basis for parameter adjustment.

[0087] To more accurately model the differences and correlations between two temporal images, in some embodiments, a perceptron network is introduced into each layer of the base model, placed before the LayerNorm operation in each layer. First, the differences are modeled using the absolute difference between the two temporal phases as the difference feature:

[0088]

[0089] In the formula, These are characteristics of difference.

[0090] Calculate the cosine similarity between the global context vector and another temporal local feature, and weight the current feature to form a context-enhanced feature:

[0091]

[0092] In the formula, For the context enhancement features of layer l, ⊙ represents the element-wise multiplication operation. For the i-th phase feature of the l-th layer, This represents the j-th phase characteristic of the l-th layer.

[0093] Concatenate all features along the channel dimension to form the input of the perceptron:

[0094]

[0095] In the formula, Let C represent the input features of the perceptron network, and C be the feature concatenation operation. These are context-enhanced features constructed based on two time phases, t1 and t2.

[0096] The perceptron employs two layers of linear mapping and two layers of Sigmoid nonlinear activation functions, and its structure is as follows:

[0097]

[0098] h2 = Sigmoid(h1W2 + b2)

[0099] The output is split into corresponding weight scores based on the channel partitioning method:

[0100]

[0101] In the formula, h1 represents the output feature of the first layer linear transformation of the perceptron network after Sigmoid activation, W1 represents the learnable weight matrix of the first layer of the perceptron network, b1 represents the bias parameters of the first layer of the perceptron network, h2 represents the output feature of the second layer linear transformation of the perceptron network after Sigmoid activation, W2 represents the learnable weight matrix of the second layer of the perceptron network, and b2 represents the bias parameters of the second layer of the perceptron network. and These are the first weight score and the second weight score, respectively, and Split is the splitting function.

[0102] The aforementioned perceptron network integrates cross-temporal differences and contextual dependencies, and its output temporal-aware weights can be used to adjust scaling and offset parameters in LayerNorm, such as γ. l ,βl This provides an important basis for achieving more robust two-phase normalization.

[0103] S30: The bi-temporal weight scores output by the perceptron network are used as the reference weights for the reconstruction of the regularization layer, guiding the dynamic construction of learnable affine parameters and generating a set of dynamic regularization parameters for each temporal phase.

[0104] In this embodiment, step S30 uses the bi-temporal weight score output by the perceptron network in step S20 as the reference weight for regularization layer reconstruction. Combined with the auxiliary learning parameters to guide the dynamic construction of learnable affine parameters, a set of dynamic regularization parameters is generated for each temporal phase to achieve adaptation to the disturbed features.

[0105] In some embodiments, the two learnable affine parameters (weight parameter γ and bias parameter β) in the regularization layer are adjusted in step S30. First, the working mechanism of the regularization layer LayerNorm is as follows, assuming the input of the l-th layer LayerNorm is the feature tensor of the i-th time phase:

[0106]

[0107] In the formula, Let B be the characteristic tensor of the i-th time phase; and let C be the characteristic tensor of the i-th time phase. dimensional features; It is the field of real numbers.

[0108] LayerNorm calculates the mean and variance along the channel dimension for each sample and location, which are then used for normalization.

[0109]

[0110] In the formula, Let be the mean and standard deviation matrices of the l-th layer, respectively, and let ∈>0 be a small constant to prevent numerical instability. The input features of the l-th layer, This represents the normalized feature of the l-th layer.

[0111] After normalization, a shared affine transformation is performed:

[0112]

[0113] In the formula, To share the features after affine transformation, These are shared scaling and offset parameters.

[0114] Using the biphase weighted fraction of the perceptron output and auxiliary learning parameters Calculate the phase affine parameters:

[0115]

[0116] in, It is the reconstruction parameter of the i-th phase feature.

[0117] When sharing affine transformations, one can use... Replace γ l ,β l To achieve shared affine transformation, the adjusted parameters are used. The method of performing phase-specific normalized affine transformations improves the model's adaptability to differences between two phases.

[0118] S40: During the testing phase, the interfering dual-temporal remote sensing data is input into the base model in batches; the perceptron network dynamically models the temporal feature distribution under interference conditions, generates corresponding reconstruction parameters for each temporal phase, and uses them to update the affine parameters in the regularization layer; the dual-temporal features at different scales output by the encoder are processed by the decoder to generate the final change detection results.

[0119] In this embodiment, step S40 is the testing phase, in which the interfering dual-temporal remote sensing data is input into the network optimized by steps S10-S30; in step S20, the embedded perceptron network dynamically models the temporal feature distribution under interference conditions, generates corresponding reconstruction parameters, and constructs logic to update the affine parameters of the regularization layer based on the parameters of step S30; the dual-temporal features of different scales output by the Transformer encoder after parameter update are sent to the MLP decoder, and the final change detection result is generated through the same feature difference decoding, upsampling to a unified scale, fusion and mapping processing as in step S10.

[0120] S50: Introduces a loss function based on minimizing pixel-level prediction entropy to encourage the model to make deterministic change predictions at each spatial location and update the parameters of the perceptron network.

[0121] In this embodiment, step S50 introduces a loss function based on minimizing pixel-level prediction entropy. This function is based on the pixel-level prediction results output in step S40. By minimizing the prediction entropy, it encourages the model to make deterministic predictions at each spatial location. The loss value is then backpropagated to update the parameters of the perceptron network in step S20, thereby improving its feature modeling ability in interference scenarios and enhancing the detection accuracy of step S40.

[0122] In some embodiments, a loss function design strategy based on minimizing prediction entropy is employed. This strategy is specifically designed for adaptive scenarios during unsupervised testing, aiming to improve the model's generalization ability to noisy data. In the testing phase lacking labeled supervision, this strategy, by minimizing the prediction entropy of the model's output, prompts the network to make more deterministic judgments under the current input, thereby enhancing the model's adaptability to the test domain.

[0123] Specifically, suppose the input image is processed by the network and outputs a probability map of changes. Where B represents the batch size, and H and W are the spatial dimensions of the image, respectively. The predicted value P for each pixel... i,j Let represent the probability that pixel (i,j) belongs to the change category, with a value range of [0,1]. Then, the prediction entropy of this pixel can be defined as:

[0124]

[0125] Therefore, the average entropy loss function for the entire graph is defined as:

[0126]

[0127] In some embodiments, such as Figure 2 The diagram shown is a flowchart of the network architecture of the dual-temporal remote sensing image change detection method provided in this application embodiment. Figure 2 The core process of a dual-temporal remote sensing image change detection method for complex interference is presented. Through the process of feature encoding → temporal feature modeling → perceptron parameter adjustment → decoding prediction → loss optimization, accurate change detection under interference scenarios is achieved. This corresponds to the implementation process of the basic model construction, perceptron embedding, dynamic parameter adjustment, result output, and loss optimization of the dual-temporal remote sensing image change detection method (corresponding to steps S10-S50 respectively).

[0128] Specifically, the encoder receives dual-temporal remote sensing images, extracts their deep features, and outputs feature representations x for both temporal phases. t1 and x t2 Dual-temporal characteristics x t1 and x t2 These represent the encoded features of remote sensing images taken at different times, and serve as the basic input for subsequent modeling of temporal relationships.

[0129] Dual-phase characteristics x t1 and x t2 The input consists of the cosine branch and the diff branch. The cosine branch is used for correlation modeling, relating x... t1 and x t2 The cosine similarity (cosine operation) is calculated to capture the correlation between features from two different time periods, such as the feature matching degree of the same feature in different time periods, and outputs the correlation features to characterize the similarity patterns of features from different time periods. The diff branch is used for difference modeling, calculating x through diff operations (such as absolute difference, Euclidean distance, etc.). t1 and x t2 The differences are explicitly extracted to reveal the changes in dual-temporal features, such as the differences in features caused by the addition, disappearance, or changes in attributes of ground features. The differential features are output, and the change patterns of temporal features are considered.

[0130] After being processed by the cos and diff branches, the features are then processed by the C (concatenation) operation, which concatenates and fuses the relevant and dissimilar features to obtain composite features containing both correlation and dissimilar information, which serve as the input basis for the perceptron network.

[0131] Perceptor network ( Figure 2 Chinese f A The module receives the concatenated composite features, learns the dynamic correlation and difference patterns of the two-temporal features through multi-level linear transformation and nonlinear activation, and outputs a set of weight scores. and This set of weights reflects the reliability and correlation of the dual-temporal features under the current interference scenario, providing a basis for subsequent dynamic adjustment of regularization parameters.

[0132] The features output by the encoder are normalized to stabilize the feature distribution, while the perceptron output is utilized. and By combining learnable auxiliary parameters, phase-specific affine parameters are dynamically constructed to achieve the goal of customizing regularization parameters for each phase, allowing the regularization process to adapt to the phase characteristic differences under disturbance.

[0133] The decoder receives the dynamically regularized dual-temporal features, performs multi-layer sensing, and outputs the change detection results. Figure 2 The black-and-white binary image on the right shows white regions representing changes and black regions representing unchanged regions. The entropy minimization loss L... CD Based on the pixel-level prediction results output by the decoder, the prediction entropy is calculated and minimized to constrain the model output to make more certain change predictions, thereby improving the accuracy of the detection results. At the same time, backpropagation optimizes the parameters of the perceptron, encoder, and decoder.

[0134] Another aspect of this application embodiment provides a dual-temporal remote sensing image change detection device, such as... Figure 3 The diagram shown is a structural diagram of a dual-temporal remote sensing image change detection device provided in an embodiment of this application. The dual-temporal remote sensing image change detection device includes:

[0135] The basic model building module 301 is configured to input a remote sensing dataset into a change detection network including an encoder and a decoder, and train the change detection network by minimizing the cross-entropy loss function to obtain a basic model; wherein, the remote sensing dataset includes dual-temporal remote sensing data, the encoder is used to extract deep features from the dual-temporal remote sensing data, and the decoder is used to map the deep features into change detection results;

[0136] The perceptron embedding module 302 is configured to embed a perceptron network before the regularization layer at each scale of the encoder based on the base model; the perceptron network takes the bi-temporal features extracted by the encoder as input, learns the correlation and difference between the two temporal features, and outputs the corresponding weight scores.

[0137] The dynamic parameter construction module 303 is configured to use the bi-temporal weight score output by the perceptron network as the reference weight for the reconstruction of the regularization layer, guide the dynamic construction of learnable affine parameters, and generate a set of dynamic regularization parameters for each temporal phase.

[0138] The change detection inference module 304 is configured to input interfering dual-temporal remote sensing data into the base model in batches during the testing phase; the perceptron network dynamically models the temporal feature distribution under interference conditions and generates corresponding reconstruction parameters for each temporal phase to update the affine parameters in the regularization layer; the dual-temporal features at different scales output by the encoder are processed by the decoder to generate the final change detection result;

[0139] The perception optimization module 305 is configured to introduce a loss function based on minimizing pixel-level prediction entropy to encourage the model to make deterministic change predictions at each spatial location and update the parameters of the perceptron network.

[0140] It should be noted that the dual-temporal remote sensing image change detection device provided in the above embodiments and the dual-temporal remote sensing image change detection method provided in the aforementioned embodiments belong to the same concept. The specific way in which each module and unit performs operations has been described in detail in the method embodiments, and will not be repeated here.

[0141] Another aspect of this application provides an electronic device, including: a controller; and a memory for storing one or more programs, which, when executed by the controller, perform the methods described in the various embodiments above.

[0142] Another aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above. This computer-readable storage medium may be included in the electronic device described in the above embodiments, or it may exist independently and not assembled into the electronic device.

[0143] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. The transmitted data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0144] Another aspect of this application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various embodiments described above.

[0145] According to one aspect of the embodiments of this application, a computer system is also provided, including a Central Processing Unit (CPU), which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) or a program loaded from storage into random access memory (RAM), such as performing the methods described above. Various programs and data required for system operation are also stored in the RAM. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0146] For example, a computer system includes a Central Processing Unit (CPU), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or loaded from storage into random access memory (RAM), such as executing the methods described in the above embodiments. The RAM also stores various programs and data required for system operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0147] The following components are connected to the I / O interface: input components including keyboards, mice, etc.; output components including cathode ray tubes (CRTs), liquid crystal displays (LCDs), and speakers; storage components including hard drives; and communication components including network interface cards such as LAN (Local Area Network) cards and modems. The communication components perform communication processing via networks such as the Internet. Drives are also connected to the I / O interface as needed. Removable media, such as disks, optical discs, magneto-optical discs, semiconductor memories, etc., are installed on the drive as needed so that computer programs read from them can be installed into the storage components as required.

[0148] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs various functions defined in the system of this application.

[0149] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0150] The module units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0151] The above embodiments are only used to illustrate this application and are not intended to limit this application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of this application. Therefore, all equivalent technical solutions also fall within the scope of this application, and the patent protection scope of this application should be defined by the claims.

Claims

1. A dual temporal remote sensing image change detection method, characterized in that, The method includes: A remote sensing dataset is input into a change detection network comprising an encoder and a decoder. The change detection network is trained by minimizing the cross-entropy loss function to obtain a base model. The remote sensing dataset includes dual-temporal remote sensing data. The encoder is used to extract deep features from the dual-temporal remote sensing data, and the decoder is used to map the deep features into change detection results. Based on the aforementioned basic model, a perceptron network is embedded before the regularization layer at each scale of the encoder; the perceptron network takes the bi-temporal features extracted by the encoder as input, learns the correlation and difference between the two temporal features, and outputs the corresponding weight scores. The bi-temporal weight score output by the perceptron network is used as the reference weight for the reconstruction of the regularization layer, guiding the dynamic construction of learnable affine parameters and generating a set of reconstruction parameters for each temporal phase. During the testing phase, the interfering dual-temporal remote sensing data is input into the base model in batches; the perceptron network dynamically models the temporal feature distribution under interference conditions, and generates corresponding reconstruction parameters for each temporal phase to update the affine parameters in the regularization layer; the dual-temporal features at different scales output by the encoder are processed by the decoder to generate the final change detection results; A loss function based on minimizing pixel-level prediction entropy is introduced to encourage the model to make deterministic change predictions at each spatial location and update the parameters of the perceptron network.

2. The method of claim 1, wherein, In the process of inputting remote sensing datasets into a change detection network that includes an encoder and a decoder, and training the change detection network by minimizing the cross-entropy loss function to obtain the base model: The encoder responds to the input bi-temporal remote sensing data, converting it into a sequence representation and inputting it into a multi-layer encoder module. The multi-layer encoder module then obtains bi-temporal features at different scales. Each layer encoder module calculates the corresponding scale's bi-temporal features using the following formula: wherein and respectively represent a first temporal feature and a second temporal feature of the layer, respectively represent a first temporal feature and a second temporal feature of the layer, and respectively represent a first temporal feature and a second temporal feature of the layer; The decoder decodes the difference between two temporal features to obtain the change features at the corresponding scale. It then unifies and fuses the change features at each scale to the same scale through an upsampling operation. After upsampling by two-dimensional transpose convolution, the result is fed into a linear mapping to generate the final change mask map.

3. The method of claim 2, wherein, The perceptron network takes the bi-temporal features extracted by the encoder as input, learns the correlation and difference between the two temporal features, and outputs the corresponding weight scores, including: The absolute difference between two temporal features is calculated as the difference feature using the following formula: In the formula, is a differential feature; Based on two temporal features, the context-enhanced features are calculated using the following formula: wherein is a first contextual enhancement feature of the layer, is an element-wise multiplication operation, is a first layer, i temporal feature, is a first layer, j temporal feature; The two temporal features, context-enhanced features, and differential features are concatenated along the channel dimension using the following formula, and used as the input features for the perceptron network: wherein is an input feature of the sensor network, is a feature concatenation operation, is a context enhanced feature constructed based on two time phases t1 and t2; The network of perceptrons outputs a corresponding weight score in response to an input , by the following equation: wherein is the output feature of the first layer of the perceptron network after a sigmoid activation of a linear transformation, is a learnable weight matrix of the first layer of the perceptron network, is a bias parameter of the first layer of the perceptron network, is the output feature of the second layer of the perceptron network after a sigmoid activation of a linear transformation, is a learnable weight matrix of the second layer of the perceptron network, is a bias parameter of the second layer of the perceptron network, and are a first weight fraction and a second weight fraction, respectively, is a splitting function.

4. The method of claim 1, wherein, Using the bi-temporal weight scores output by the perceptron network as reference weights for the reconstruction of the regularization layer, the dynamic construction of learnable affine parameters can be guided by the following methods: The first i The tensor of the temporal features is input into the regularization layer, and normalized features are obtained through the regularization layer. Using the biphase weight scores and auxiliary learning parameters output by the perceptron network, the temporal affine parameters are calculated; the temporal affine parameters are used to perform affine transformation on the normalized features.

5. The method of claim 4, wherein, Using the biphase weight scores and auxiliary learning parameters output by the perceptron network, the temporal affine parameters are calculated using the following formula: wherein is a first i reconstruction parameter of a temporal feature, and are shared scaling and offset parameters, and are auxiliary learning parameters, and are a first weight fraction and a second weight fraction, respectively.

6. The method of claim 1, wherein, The loss function based on minimizing pixel-level prediction entropy is expressed as: wherein, is a loss function based on pixel-level prediction entropy minimization, is a batch size, is a height of an image, is a width of an image, is an index of a batch, is a pixel index in the image height direction, is a pixel index in the image width direction, is a pixel prediction entropy.

7. The method of claim 6, wherein, The pixel The formula for calculating the prediction entropy is: wherein representing a pixel is the probability of the change category.

8. A dual temporal remote sensing image change detection apparatus, characterized in that, The device includes: The basic model building module is configured to input a remote sensing dataset into a change detection network including an encoder and a decoder, and train the change detection network by minimizing the cross-entropy loss function to obtain a basic model; wherein, the remote sensing dataset includes dual-temporal remote sensing data, the encoder is used to extract deep features from the dual-temporal remote sensing data, and the decoder is used to map the deep features into change detection results; The perceptron embedding module is configured to embed a perceptron network before the regularization layer at each scale of the encoder based on the base model; the perceptron network takes the bi-temporal features extracted by the encoder as input, learns the correlation and difference between the two temporal features, and outputs the corresponding weight scores. The dynamic parameter construction module is configured to use the bi-temporal weight scores output by the perceptron network as the reference weights for the reconstruction of the regularization layer, guide the dynamic construction of learnable affine parameters, and generate a set of reconstruction parameters for each temporal phase. The change detection inference module is configured to input interfering dual-temporal remote sensing data into the base model in batches during the testing phase; the perceptron network dynamically models the temporal feature distribution under interference conditions and generates corresponding reconstruction parameters for each temporal phase to update the affine parameters in the regularization layer; the dual-temporal features at different scales output by the encoder are processed by the decoder to generate the final change detection result; The perceptron optimization module is configured to introduce a loss function based on minimizing pixel-level prediction entropy to encourage the model to make deterministic change predictions at each spatial location and update the parameters of the perceptron network.

9. An electronic device, comprising: The electronic device includes: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 7. 10.A non-transitory computer-readable storage medium storing instructions, wherein, When the instructions are executed by the processor, the method according to any one of claims 1 to 7 is performed.

Images