A phase confidence guided semi-supervised remote sensing image change detection method

Abstract

The present application relates to the technical field of remote sensing image processing and computer vision, and particularly relates to a stage confidence guided semi-supervised remote sensing image change detection method, comprising: S1: obtaining a labeled sample set and an unlabeled sample set; S2: constructing a student network and a teacher network; S3: performing supervised preheating training on the student network through the labeled sample set; S4: initializing the parameters of the teacher network based on the parameters of the preheated student network; S5: inputting the unlabeled sample into the current teacher network to generate a reliable pixel mask of the unlabeled sample; S6: inputting the unlabeled sample into the preheated student network, and jointly updating the student network parameters through an unlabeled consistency constraint loss and supervised constraint; S7: updating the parameters of the teacher network; S8: repeating S5-S7 to perform semi-supervised joint training; and S9: inputting the dual-phase remote sensing image into the change detection model to output a change detection map. The present application can improve the accuracy and stability of remote sensing image change detection.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer vision technology, specifically to a semi-supervised remote sensing image change detection method guided by stage confidence. Background Technology

[0002] Remote sensing image change detection aims to automatically identify areas of change such as land cover, buildings, roads, water bodies, vegetation, or disaster damage from remote sensing images of the same geographic area at different time phases. This provides crucial data support for land resource monitoring, disaster assessment, urban expansion analysis, and ecological environment governance. As the resolution of remote sensing sensors continues to improve, changing targets exhibit characteristics such as small scale, sparse distribution, complex backgrounds, and significant differences in appearance across time phases. Traditional manual rule-based methods and fully supervised deep learning methods typically rely on large amounts of dense pixel-level annotations, resulting in high annotation costs and insufficient transferability.

[0003] Existing patent CN109993727A discloses a remote sensing image change detection method based on deep learning. This method acquires remote sensing images of the same area at different times, performs preprocessing, image transformation, and difference feature extraction, and then uses a deep learning model to complete change detection. While this approach can reduce manual identification workload by utilizing difference maps, in scenarios where registration errors, seasonal lighting differences, and sparse variations in small targets coexist, the difference maps are prone to containing many false changes, and it is insufficient for utilizing large amounts of unlabeled imagery.

[0004] Existing patent CN112991257B discloses a rapid change detection method for heterogeneous remote sensing images based on semi-supervised Siamese networks. This method constructs a semi-supervised Siamese network using a pre-trained feature network and generates a change difference map after training high-level features with a small number of samples from unchanging regions. While this approach is suitable for rapid detection of heterogeneous remote sensing images, the pseudo-supervisory information mainly relies on local samples and the difference map, lacking a dynamic filtering mechanism for reliable regions that changes with the training stage. This makes it prone to pseudo-label drift under low-labeling conditions.

[0005] Existing patent CN116343033A discloses a semi-supervised method and apparatus for building change detection guided by consistency loss. It improves building change detection performance by generating pseudo-labels through a lightweight encoder-decoder, multi-decoder consistency, and confidence thresholds. This scheme primarily focuses on building change scenarios, and there is still room for improvement in teacher model stability, threshold stage scheduling, joint constraints between global prior confidence and change prediction confidence, and complementary modeling of change response and steady-state response in bi-temporal features.

[0006] In summary, existing technologies still suffer from problems such as insufficient pseudo-supervised pixel selection, mismatch between teacher update strategies and training phases, easy confusion between real changes and background drift, missed detection of multi-scale small target changes, and broken change boundaries under limited annotation conditions. Therefore, there is a need for a remote sensing image change detection method that can utilize a small number of labeled samples and a large number of unlabeled samples, while simultaneously enhancing the capabilities of dual-temporal feature separation, reliable pixel selection, and multi-scale prediction refinement. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a stage-confidence-guided semi-supervised remote sensing image change detection method that can improve the problems of pseudo-supervised noise accumulation, fixed teacher updates, under-constraint of dual-temporal fusion, and missed detection of small target changes in the existing semi-supervised remote sensing change detection methods, thereby improving the accuracy and stability of remote sensing image change detection.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A stage-confidence-guided semi-supervised remote sensing image change detection method includes:

[0010] S1: Obtain a labeled sample set containing dual-temporal remote sensing images and their actual change masks, as well as an unlabeled sample set containing only dual-temporal remote sensing images; the dual-temporal remote sensing images include two remote sensing images, one from the first time phase and one from the second time phase.

[0011] S2: Construct student and teacher networks with identical network structures and independently maintained parameters; the parameters of the student network are updated through gradient backpropagation, while the parameters of the teacher network are updated based on the parameters of the student network using an exponential moving average method;

[0012] S3: Supervised warm-up training of the student network is performed using only the labeled sample set to obtain the warmed-up student network;

[0013] S4: Copy the preheated student network parameters to the teacher network to complete the parameter initialization of the teacher network;

[0014] S5: Input unlabeled samples into the current teacher network, and generate a reliable pixel mask of the unlabeled samples as a pseudo-supervised region based on the reliability judgment;

[0015] S6: Input labeled samples into the student network to calculate the supervised loss, and simultaneously input unlabeled samples into the student network. Calculate the unlabeled consistency constraint loss between the student network output and the teacher network output only at the pixel positions selected by the trusted pixel mask. Update the parameters of the student network based on the supervised loss and the unlabeled consistency constraint loss.

[0016] S7: Update the parameters of the teacher network based on the updated parameters of the student network;

[0017] S8: Repeat steps S5 to S7 to perform semi-supervised joint training, iteratively update the parameters of the student network and the teacher network until the preset number of iterations is reached or the training loss tends to stabilize. Finally, based on the change detection accuracy on the reserved validation set, select one of the trained teacher network and student network as the change detection model.

[0018] S9: Input the dual-temporal remote sensing image to be detected into the trained change detection model, and output the corresponding change detection map.

[0019] Preferably, in step S2, both the student network and the teacher network include a temporal feature extraction module for extracting multi-scale features, a differential-stability collaborative integration module for modeling the change response and steady-state response of multi-scale features, and a prior-constrained multi-scale decoding module for outputting a global prior probability map and a change prediction probability map; wherein, the teacher-student semi-supervised training framework also includes a stage-based trusted mutual training module used only in the training phase.

[0020] The stage-based trusted mutual training module includes a stage scheduling submodule, a trusted pixel selection submodule, and a consistency constraint submodule. The stage scheduling submodule calculates the smoothing coefficient and trusted threshold of the teacher network based on the current iteration number. The trusted pixel selection submodule generates a trusted pixel mask based on the global prior probability map and change prediction probability map of the teacher network output. The consistency constraint submodule calculates the unlabeled consistency constraint loss between the student network output and the teacher network output only at the pixel positions selected by the trusted pixel mask.

[0021] Preferably, the temporal feature extraction module includes a first temporal feature extraction submodule and a second temporal feature extraction submodule; the first temporal feature extraction submodule is used to extract multi-scale first temporal features of the first temporal remote sensing image, and the second temporal feature extraction submodule is used to extract multi-scale second temporal features of the second temporal remote sensing image; the multi-scale first temporal features and the second temporal features undergo lightweight channel spatial recalibration processing before entering the differential stability collaborative integration module.

[0022] Preferably, the differential-stability collaborative integration module is used to model the change response and steady-state response of the first-phase features and the second-phase features at the same scale, including: generating the change response based on the absolute value of the difference between the first-phase features and the second-phase features; generating the steady-state response based on the sum of the first-phase features and the second-phase features; performing lightweight convolution calibration on the change response and the steady-state response respectively; and fusing the calibrated change response and the steady-state response element-wise according to the scale map adaptively generated by the calibrated change response and the steady-state response to obtain the multi-scale differential-stability integrated features.

[0023] The formula for calculating the multi-scale differential stability integration characteristics is as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] In the formula: and They represent the first The first and second phase characteristics of the scale; Indicates the first Response to scale changes; Indicates the first Steady-state response at scale; , They represent the first Scale-calibrated change response and steady-state response; , and Represents a lightweight mapping function; This represents the Sigmoid activation function; This represents the adaptive fusion scaling diagram; This indicates element-wise multiplication; Indicates the first The differential stability integration characteristics of scale.

[0031] Preferably, the prior-constrained multi-scale decoding module includes a global prior generation submodule, a prior-guided fusion submodule, a local boundary enhancement submodule, and a context prediction submodule. The global prior generation submodule generates a global prior branch logit and a global change prior map from the mid-to-deep differentially stable integrated features. The prior-guided fusion submodule uses the global change prior map to perform gated fusion on the shallow-to-mid-layer differentially stable integrated features. The local boundary enhancement submodule performs residual enhancement on the gated fused features to obtain the enhanced decoded features. The context prediction submodule uses a multi-scale convolutional receptive field to output a change probability map.

[0032] The formula for calculating the global prior generation submodule is:

[0033] ;

[0034] ;

[0035] In the formula: This represents the global prior fusion feature obtained by fusing the differential stability integration features from the middle to deep layers; This represents the global prior branch, logit. Represents the global prior probability graph; Represents the global prior mapping function;

[0036] The calculation formula for the local boundary enhancement submodule is as follows:

[0037] ;

[0038] ;

[0039] In the formula: This indicates the characteristics after gating fusion; Indicates the boundary enhancement mapping function; Represents a local boundary confidence plot; This represents the enhanced decoding features;

[0040] The context prediction submodule first performs channel compression on the enhanced decoded features, then extracts context features through multiple parallel convolutional branches with different receptive fields, and concatenates the results of the parallel branches with the compressed features and performs 1×1 convolution to obtain context fusion features. Finally, the prediction convolution outputs the change prediction branch logit and change probability map based on the context fusion features.

[0041] The formula for calculating the probability diagram of change is:

[0042] ;

[0043] ;

[0044] In the formula: Represents context-fused features; The change prediction branch is represented by logit; Represents a probability diagram of change; This represents the context prediction mapping function.

[0045] Preferably, step S3 specifically includes the following processing steps:

[0046] S301: Obtain the labeled samples in the labeled sample set and their actual change masks;

[0047] S302: Input the labeled samples into the student network, and after processing by the temporal feature extraction module, the differential stability collaborative integration module and the prior constraint multi-scale decoding module, output the global prior branch logit and the change prediction branch logit;

[0048] S303: Supervised loss is calculated based on the global prior branch logit and the change prediction branch logit, respectively, and the corresponding real change mask.

[0049] The formula for calculating supervised loss is:

[0050] ;

[0051] In the formula: This indicates a loss due to oversight; The global prior branch logit represents the student network output; The logit branch represents the prediction branch for changes in student network output; A mask representing the true changes in the labeled samples; Represents the binary cross-entropy based on logit;

[0052] S304: Update the parameters of the student network using gradient backpropagation based on supervised loss;

[0053] S305: Repeat steps S301 to S304 until the preset number of supervised warm-up iterations is reached or the supervised loss tends to stabilize, and the supervised warm-up student network is obtained.

[0054] Preferably, in step S7, updating the parameters of the teacher network based on the parameters of the student network means: calculating the smoothing coefficient and trust threshold of the teacher network based on the current iteration number through the stage scheduling submodule of the stage trustworthy mutual training module; and performing an exponential moving average on the parameters of the student network based on the smoothing coefficient to update the parameters of the teacher network.

[0055] The calculation formula for the stage scheduling submodule is as follows:

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] In the formula: Indicates the first The stage scheduling value for the next iteration; and For scheduling parameters; This represents the total number of iterations. and These represent the lower and upper bounds of the smoothing coefficient, respectively. and These represent the lower and upper bounds of the confidence threshold, respectively. , They represent the first The smoothing coefficient and confidence threshold for each iteration; and They represent the first The parameters of the student network and the teacher network in the next iteration; Indicates the first The parameters of the teacher network in the next iteration.

[0061] Preferably, step S5 specifically includes the following processing steps:

[0062] S501: Obtain unlabeled samples from the unlabeled sample set;

[0063] S502: Input unlabeled samples into the teacher network, and after processing by the temporal feature extraction module, the differential stability collaborative integration module and the prior constraint multi-scale decoding module, output the global prior probability map and change prediction probability map corresponding to the unlabeled samples;

[0064] S503: Through the trusted pixel filtering submodule of the stage trusted mutual training module, the global prior probability map and the change prediction probability map are judged on the basis of the binary classification uncertainty center. When both the global prior probability map and the change prediction probability map meet the trusted threshold requirements and are located on the same side of the binary classification uncertainty center, the corresponding pixel is determined as a trusted pixel. Then, a trusted pixel mask is generated based on all trusted pixels.

[0065] The formula for calculating the reliable pixel mask is:

[0066] ;

[0067] In the formula: Indicates the first The next iteration's reliable pixel mask; Indicates the first The next iteration of the teacher network on unlabeled samples Output global prior probability map; Indicates the first The confidence threshold for the next iteration; Indicates the first The next iteration of the teacher network on unlabeled samples Probability plot of output change prediction.

[0068] Preferably, step S6 specifically includes the following processing steps:

[0069] S601: Obtain labeled samples and their true change masks from the labeled sample set, as well as unlabeled samples and their corresponding reliable pixel masks;

[0070] S602: Input the labeled samples into the student network. After processing by the temporal feature extraction module, the differential-stability collaborative integration module, and the prior constraint multi-scale decoding module, output the global prior branch logit and the change prediction branch logit. Based on the global prior branch logit and the change prediction branch logit, respectively, calculate the supervised loss with the corresponding real change mask.

[0071] S603: Input unlabeled samples into the student network. After processing by the temporal feature extraction module, the differential-stability collaborative integration module, and the prior constraint multi-scale decoding module, the global prior branch logit and the change prediction branch logit of the student network are output. The corresponding global prior probability map and change prediction probability map are obtained by activating them with Sigmoid respectively.

[0072] S604: Through the consistency constraint submodule of the stage trusted mutual training module, the global prior probability map and change prediction probability map output by the teacher network are used as soft supervision targets. Only at the pixel position selected by the trusted pixel mask, the unlabeled consistency constraint loss is calculated based on the global prior branch logit and change prediction branch logit output by the student network, respectively.

[0073] The formula for calculating the loss due to unlabeled consistency constraints is as follows:

[0074] ;

[0075] In the formula: This represents the loss due to unlabeled consistency constraints; Indicates the first The next iteration of the student network at the pixel The global prior branch logit is output at this point; Indicates the first The next iteration of the student network at the pixel The output change prediction branch logit; Indicates the first Next iteration pixel A reliable pixel mask at the location; Indicates the first The next iteration of the teacher network at the pixel The global prior probability map output at the location; Indicates the first The next iteration of the teacher network at the pixel The probability of change in the output at that point; Represents the binary cross-entropy based on logit; This indicates a positive number that prevents the denominator from being zero.

[0076] S605: Calculate the total training loss based on supervised loss and unlabeled consistency constraint loss;

[0077] The formula for calculating the total training loss is:

[0078] ;

[0079] In the formula: Indicates the total training loss; Indicates the weight of unlabeled consistency constraints;

[0080] S606: Optimize student network parameters through gradient backpropagation based on total training loss.

[0081] Preferably, in step S9, the first and second temporal remote sensing images of the dual-temporal remote sensing image to be detected are input into the trained change detection model. After processing by the temporal feature extraction module, the differential stability collaborative integration module, and the prior constraint multi-scale decoding module, a change probability map is obtained. The change probability map is thresholded to obtain an initial binary change map, and morphological filtering and connected component screening are selectively applied to the initial binary change map to obtain the final binary change detection map.

[0082] Compared with existing technologies, the stage-confidence-guided semi-supervised remote sensing image change detection method of this invention has the following advantages:

[0083] This invention synchronously adjusts the teacher smoothing coefficient and the confidence threshold through a stage scheduling function, enabling the teacher network to quickly absorb effective information from the student network and expand the coverage of unlabeled pixels in the early stages of training, and enabling the teacher network to have a smoother parameter trajectory and provide more reliable pseudo-supervision signals in the later stages of training.

[0084] This invention employs a joint screening of a global prior probability map and a change prediction probability map, applying consistency constraints only to pixels where both are far from the uncertainty center and share the same direction. This approach can suppress pseudo-change supervision caused by registration errors, seasonal changes in illumination, and complex background textures.

[0085] This invention models both the change response and the steady-state response simultaneously through a differential-steady-state collaborative integration module, and uses an adaptive scaling diagram for element-by-element fusion. This can highlight the real changes while retaining the prior stable structure, and reduce over-detection caused by single differential modeling and under-detection caused by single common modeling.

[0086] This invention integrates global change prior, local boundary enhancement, and multi-scale context prediction into a prior-constrained multi-scale decoding module, which can improve the detection quality of small target changes, sparse changes, and boundary regions, while avoiding the dispersion of technical solutions caused by too many independent English module names.

[0087] The various modules and their functional designs in the stage-confidence-guided change detection model of this invention have a clear synergistic relationship: the differential-stability collaborative integration module is used to separate the true change response and the cross-temporal steady-state structure at the feature level, reducing pseudo-change responses caused by registration errors, illumination differences, and seasonal variations; the prior-constrained multi-scale decoding module is used at the decoding level to jointly use global change priors, local boundary enhancements, and multi-scale contextual information for change region prediction, improving the integrity of small target changes and boundary regions; the stage-confidence mutual training module is used at the training level to synchronously adjust the smoothness of teacher parameters and the strictness of confidence pixel selection according to the training stage, and introduces unlabeled consistency constraints only at pixels where both the global prior probability map and the change prediction probability map are mutually reliable and have the same direction. Through the mutual cooperation of the various modules in the stage-confidence-guided change detection model, the model can simultaneously reduce pseudo-supervision noise and improve the utilization efficiency of unlabeled samples under a small number of pixel-level annotations, thereby improving the accuracy and stability of remote sensing image change detection. Attached Figure Description

[0088] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0089] Figure 1 and Figure 2 The flowchart and logic diagram are for a phase confidence-guided semi-supervised remote sensing image change detection method.

[0090] Figure 3 Network structure diagram of the stage confidence-guided change detection model (student network and teacher network).

[0091] Figure 4 This is a schematic diagram of differential-stable collaborative feature integration and prior decoding.

[0092] Figure 5 This is a flowchart for phased trusted mutual training. Detailed Implementation

[0093] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0094] The following detailed explanation illustrates the specific implementation methods:

[0095] Example:

[0096] This embodiment discloses a semi-supervised remote sensing image change detection method guided by stage confidence.

[0097] like Figure 1 and Figure 2 As shown, a semi-supervised remote sensing image change detection method guided by stage confidence includes:

[0098] S1: Obtain a labeled sample set containing dual-temporal remote sensing images and their actual change masks, as well as an unlabeled sample set containing only dual-temporal remote sensing images; the dual-temporal remote sensing images include two remote sensing images, one from the first time phase and one from the second time phase.

[0099] S2: Construct student and teacher networks with identical network structures and independently maintained parameters; the parameters of the student network are updated through gradient backpropagation, while the parameters of the teacher network are updated based on the parameters of the student network using an exponential moving average method;

[0100] S3: Supervised warm-up training of the student network is performed using only the labeled sample set to obtain the warmed-up student network;

[0101] S4: Copy the preheated student network parameters to the teacher network to complete the parameter initialization of the teacher network;

[0102] S5: Input unlabeled samples into the current teacher network, and generate a reliable pixel mask of the unlabeled samples as a pseudo-supervised region based on the reliability judgment;

[0103] S6: Input labeled samples into the student network to calculate the supervised loss, and simultaneously input unlabeled samples into the student network. Calculate the unlabeled consistency constraint loss between the student network output and the teacher network output only at the pixel positions selected by the trusted pixel mask. Update the parameters of the student network based on the supervised loss and the unlabeled consistency constraint loss.

[0104] S7: Update the parameters of the teacher network based on the updated parameters of the student network;

[0105] S8: Repeat steps S5 to S7 to perform semi-supervised joint training, iteratively update the parameters of the student network and the teacher network until the preset number of iterations is reached or the training loss tends to stabilize. Finally, based on the change detection accuracy on the reserved validation set, select one of the trained teacher network and student network as the change detection model.

[0106] S9: Input the dual-temporal remote sensing image to be detected into the trained change detection model, and output the corresponding change detection map.

[0107] To better illustrate the technical solution of the present invention, this embodiment will be described in more detail through the following parts.

[0108] I. Training Dataset

[0109] In the specific implementation process, the steps for constructing labeled and unlabeled sample sets include:

[0110] (1) Acquire two remote sensing images of the same geographic area in the first and second time phases (i.e., different acquisition times); the first time phase remote sensing image and the second time phase remote sensing image cover the same geographic area but are acquired at different times.

[0111] (2) Perform geometric registration and radiometric correction on the two remote sensing images. Geometric registration ensures that the same ground feature has a consistent spatial location in the two images. Perform radiometric correction or color normalization to reduce the impact of sensor differences, illumination differences and seasonal differences on change detection.

[0112] (3) The registered image is cropped into fixed-size image blocks and the pixel values ​​are normalized.

[0113] (4) Use image patches with pixel-level variation masks as annotation samples to construct an annotation sample set;

[0114] (5) Use image blocks without pixel-level variation masks as unlabeled samples to construct an unlabeled sample set;

[0115] (6) Perform flipping, rotation, scale perturbation, color perturbation or strength enhancement on labeled and unlabeled sample sets.

[0116] During training, mixed batches are drawn from the two classes of samples, and data augmentation operations such as random flipping, rotation, cropping, color perturbation, strong enhancement or weak enhancement are performed on the input images to improve the model's adaptability to changes in real-world scenes.

[0117] II. Stage Confidence-Guided Change Detection Model

[0118] like Figure 3 As shown, both the student network and the teacher network (i.e., the stage-confidence-guided change detection model) include a temporal feature extraction module for extracting multi-scale features, a differential-stability collaborative integration module for modeling the change response and steady-state response of multi-scale features, and a prior-constrained multi-scale decoding module for outputting the global prior probability map and the change prediction probability map; among them, the teacher-student semi-supervised training framework also includes a stage-confidence mutual training module used only in the training phase.

[0119] The stage-based trusted mutual training module includes a stage scheduling submodule, a trusted pixel selection submodule, and a consistency constraint submodule. The stage scheduling submodule calculates the smoothing coefficient and trusted threshold of the teacher network based on the current iteration number. The trusted pixel selection submodule generates a trusted pixel mask based on the global prior probability map and change prediction probability map of the teacher network output. The consistency constraint submodule calculates the unlabeled consistency constraint loss between the student network output and the teacher network output only at the pixel positions selected by the trusted pixel mask.

[0120] This embodiment is divided into four functional modules according to technical effects: feature extraction, differential stability integration, prior decoding, and stage mutual training, so that a clear hierarchical relationship is formed between the model structure and the training mechanism.

[0121] III. Temporal Feature Extraction Module

[0122] In the specific implementation process, the temporal feature extraction module includes a first temporal feature extraction submodule and a second temporal feature extraction submodule. The first temporal feature extraction submodule is used to extract multi-scale first temporal features of the first temporal remote sensing image, and the second temporal feature extraction submodule is used to extract multi-scale second temporal features of the second temporal remote sensing image. The multi-scale first and second temporal features undergo lightweight channel spatial recalibration processing before entering the differential stability collaborative integration module to suppress background redundancy and highlight change-related areas.

[0123] IV. Differential Stability Collaborative Integration Module

[0124] like Figure 4 As shown, the differential-stability collaborative integration module is used to model the change response and steady-state response of the first-phase features and the second-phase features at the same scale. It includes: generating the change response based on the absolute value of the difference between the first-phase features and the second-phase features; generating the steady-state response based on the sum of the first-phase features and the second-phase features; performing lightweight convolution calibration on the change response and the steady-state response respectively; and fusing the calibrated change response and the steady-state response element-wise according to the scale map adaptively generated by the calibrated change response and the steady-state response to obtain the multi-scale differential-stability integrated features.

[0125] For the first temporal feature of the i-th scale Second phase characteristics The differential stability collaborative integration module first generates the change response. and steady-state response Change response emphasizes the change response between two images, which helps to highlight the true changes; steady-state response emphasizes the stable structure between two images, which helps to suppress false changes caused by registration bias, seasonal changes in illumination, and high-frequency texture drift.

[0126] Specifically, the calculation formula for multi-scale differential stability integration characteristics is as follows:

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] In the formula: and They represent the first The first and second phase characteristics of the scale; Indicates the first Response to scale changes; Indicates the first Steady-state response at scale; , They represent the first Scale-calibrated change response and steady-state response; , and Represents a lightweight mapping function; This represents the Sigmoid activation function; This represents the adaptive fusion scaling diagram; This indicates element-wise multiplication; Indicates the first The differential stability integration characteristics of scale.

[0134] V. Prior Constraint Multi-Scale Decoding Module

[0135] In the specific implementation process, the prior constraint multi-scale decoding module includes a global prior generation submodule, a prior guided fusion submodule, a local boundary enhancement submodule, and a context prediction submodule.

[0136] The global prior generation submodule generates a global prior branch logit and a global change prior map P from the mid-to-deep differential stability integration features. The global change prior map P is used to highlight areas of significant change, serving as an auxiliary branch in supervised training and as a source of prior credibility in the stage-based credible mutual training module.

[0137] The formula for calculating the global prior generation submodule is:

[0138] ;

[0139] ;

[0140] In the formula: This represents the global prior fusion feature obtained by fusing the differential stability integration features from the middle to deep layers; This represents the global prior branch, logit. Represents the global prior probability graph; This represents the global prior mapping function.

[0141] The prior-guided fusion submodule receives shallow to mid-layer differentially stable integration features and performs gated fusion under the guidance of the global change prior graph P to obtain fused features T.

[0142] The local boundary enhancement submodule further generates a local boundary confidence map E based on the fused feature T, and enhances the boundary integrity and the significance of weak changes through residuals to obtain the enhanced decoding features.

[0143] The calculation formula for the local boundary enhancement submodule is as follows:

[0144] ;

[0145] ;

[0146] In the formula: This indicates the characteristics after gating fusion; Indicates the boundary enhancement mapping function; Represents a local boundary confidence plot; This represents the enhanced decoding features;

[0147] The context prediction submodule first performs channel compression on the enhanced decoded features, then extracts context features through multiple parallel convolutional branches with different receptive fields, and concatenates the results of the parallel branches with the compressed features and performs 1×1 convolution fusion. Finally, it outputs the change prediction branch logit and change probability map through the prediction convolution.

[0148] ;

[0149] ;

[0150] ;

[0151] ;

[0152] In the formula: Represents the set of receptive fields; Represents the contextual features of the r-th receptive field branch; Represents context-fused features; The change prediction branch is represented by logit; Represents a probability diagram of change; This represents the context prediction mapping function.

[0153] VI. Preliminary Training in Student Networking

[0154] In the specific implementation process, the processing steps for the initial online training of students include:

[0155] S301: Obtain the labeled samples (dual-temporal remote sensing images) in the labeled sample set and their actual change masks;

[0156] S302: Input the labeled samples into the student network, and after processing by the temporal feature extraction module, the differential stability collaborative integration module and the prior constraint multi-scale decoding module, output the global prior branch logit and the change prediction branch logit;

[0157] S303: Supervised loss is calculated based on the global prior branch logit and the change prediction branch logit, respectively, and the corresponding real change mask.

[0158] The formula for calculating supervised loss is:

[0159] ;

[0160] In the formula: This indicates a loss due to oversight; The global prior branch logit represents the student network output; The logit branch represents the prediction branch for changes in student network output; A mask representing the true changes in the labeled samples; Represents the binary cross-entropy based on logit;

[0161] S304: Update the parameters of the student network using gradient backpropagation based on supervised loss;

[0162] S305: Repeat steps S301 to S304 until the preset number of supervised warm-up iterations is reached or the supervised loss tends to stabilize, and the supervised warm-up student network is obtained.

[0163] VII. Update the teacher network

[0164] like Figure 5 As shown, the phased trusted mutual training module is trained based on a teacher-student framework. The student network parameters are denoted as... Teacher network parameters are recorded as , This indicates the current training iteration number. This represents the total number of training iterations. In the supervised warm-up phase, the student network is updated using only the labeled sample set to obtain a stable initial representation. After the supervised warm-up training is completed, the teacher network parameters are initialized with the warmed-up student network parameters, and then the semi-supervised joint training phase begins.

[0165] During the semi-supervised training phase, define the phase scheduling value. Based on the scheduling values ​​of this stage, the smoothing coefficient of the teacher parameters is calculated synchronously. and confidence threshold In the early stages of training, a lower smoothing coefficient and a lower confidence threshold are used to enable the teacher network to follow the student network more quickly and increase the availability of unlabeled pixels. In the later stages of training, a higher smoothing coefficient and a higher confidence threshold are used to make the teacher network trajectory smoother and improve the reliability of pseudo-supervision.

[0166] In summary, updating the teacher network parameters based on the student network parameters refers to: calculating the smoothing coefficient and confidence threshold of the teacher network based on the current iteration number through the stage scheduling submodule of the stage trusted mutual training module; performing an exponential moving average on the student network parameters based on the smoothing coefficient to update the teacher network parameters; wherein the confidence threshold is used for the confidence pixel screening in step S5.

[0167] The formula is expressed as:

[0168] ;

[0169] ;

[0170] ;

[0171] ;

[0172] In the formula: Indicates the first The stage scheduling value for the next iteration; and For scheduling parameters; This represents the total number of iterations. and These represent the lower and upper bounds of the smoothing coefficient, respectively. and These represent the lower and upper bounds of the confidence threshold, respectively. , They represent the first The smoothing coefficient and confidence threshold for each iteration; and They represent the first The parameters of the student network and the teacher network in the next iteration; Indicates the first The parameters of the teacher network in the next iteration.

[0173] 8. Trusted Pixel Mask

[0174] For unlabeled samples The teacher network outputs a global prior probability graph. and probability of change And based on the confidence threshold Constructing a trustworthy pixel mask A trustworthy pixel mask requires that both the global prior probability map and the probability of change be far from the center of uncertainty 1 / 2, and that they be oriented in the same direction, so that only the pixel regions that are jointly trustworthy are retained.

[0175] In the specific implementation process, the steps for generating a reliable pixel mask include:

[0176] S501: Obtain unlabeled samples from the unlabeled sample set (dual-temporal remote sensing imagery).

[0177] S502: Input unlabeled samples into the teacher network, and after processing by the temporal feature extraction module, the differential stability collaborative integration module and the prior constraint multi-scale decoding module, output the global prior probability map and change prediction probability map corresponding to the unlabeled samples;

[0178] S503: Through the trusted pixel filtering submodule of the stage trusted mutual training module, the global prior probability map and the change prediction probability map are judged on the basis of the binary classification uncertainty center. When both the global prior probability map and the change prediction probability map meet the trusted threshold requirements and are located on the same side of the binary classification uncertainty center, the corresponding pixel is determined as a trusted pixel. Then, a trusted pixel mask is generated based on all trusted pixels.

[0179] The formula for calculating the reliable pixel mask is:

[0180] ;

[0181] In the formula: Indicates the first The next iteration's reliable pixel mask; Indicates the first The next iteration of the teacher network on unlabeled samples Output global prior probability map; Indicates the first The confidence threshold for the next iteration; Indicates the first The next iteration of the teacher network on unlabeled samples Probability plot of output change prediction.

[0182] IX. Student Network Updates

[0183] like Figure 5As shown, the steps for updating the student network include:

[0184] S601: Obtain labeled samples and their true change masks from the labeled sample set, as well as unlabeled samples and their corresponding reliable pixel masks;

[0185] S602: Input the labeled samples into the student network. After processing by the temporal feature extraction module, the differential-stability collaborative integration module, and the prior constraint multi-scale decoding module, output the global prior branch logit and the change prediction branch logit. Calculate the supervised loss based on the global prior branch logit and the change prediction branch logit and the corresponding real change mask. The calculation logic is the same as step S303.

[0186] S603: Input unlabeled samples into the student network. After processing by the temporal feature extraction module, the differential-stability collaborative integration module, and the prior constraint multi-scale decoding module, the global prior branch logit and the change prediction branch logit of the student network are output. The corresponding global prior probability map and change prediction probability map are obtained by activating them with Sigmoid respectively.

[0187] S604: Through the consistency constraint submodule of the stage trusted mutual training module, the global prior probability map and change prediction probability map output by the teacher network are used as soft supervision targets. Only at the pixel position selected by the trusted pixel mask, the unlabeled consistency constraint loss is calculated based on the global prior branch logit and change prediction branch logit output by the student network, respectively.

[0188] The formula for calculating the loss due to unlabeled consistency constraints is as follows:

[0189] ;

[0190] In the formula: This represents the loss due to unlabeled consistency constraints; Indicates the first The next iteration of the student network at the pixel The global prior branch logit is output at this point; Indicates the first The next iteration of the student network at the pixel The output change prediction branch logit; Indicates the first Next iteration pixel A reliable pixel mask at the location; Indicates the first The next iteration of the teacher network at the pixel The global prior probability map output at the location; Indicates the first The next iteration of the teacher network at the pixel The probability of change in the output at that point; Represents the binary cross-entropy based on logit; This indicates a positive number that prevents the denominator from being zero.

[0191] S605: Calculate the total training loss based on supervised loss and unlabeled consistency constraint loss;

[0192] The formula for calculating the total training loss is:

[0193] ;

[0194] In the formula: Indicates the total training loss; Indicates the weight of unlabeled consistency constraints;

[0195] S606: Optimize student network parameters through gradient backpropagation based on total training loss.

[0196] 10. Reasoning Stage

[0197] In the specific implementation process, during the inference stage, the first and second phase remote sensing images of the dual-temporal remote sensing images to be detected are input into the trained change detection model. After processing by the temporal feature extraction module, the differential stability collaborative integration module, and the prior constraint multi-scale decoding module, a change probability map is obtained. The change probability map is then thresholded to obtain an initial binary change map. Morphological filtering and connected component screening are then selectively applied to the initial binary change map to obtain the final binary change detection map.

[0198] In one optional effect verification method, the method of this embodiment is compared with a semi-supervised change detection method that only uses a fixed confidence threshold, a method that only uses a single change response for dual-temporal fusion, and a change detection method that does not include local boundary enhancement. The comparison results show that this embodiment reduces the interference of unreliable pseudo-supervision on the student network in the early stages of training through the stage-based reliable mutual training module, and improves the reliability of pseudo-supervision regions in the later stages of training; it reduces false detections caused by registration bias, illumination changes, and background texture drift through the differential-stability collaborative integration module; and it enhances the response of sparse small targets and boundary regions through the prior-constrained multi-scale decoding module. Therefore, this embodiment can obtain more complete change regions, fewer pseudo-change patches, and smoother change boundaries.

[0199] In the feature fusion part, the present invention uses an adaptive scale map to dynamically integrate the change response and the steady-state response. In the decoding part, global prior, boundary enhancement and context prediction are used as a whole decoding module. In the training part, a stage-based trusted mutual training mechanism is used to uniformly describe teacher smoothing, trusted threshold and trusted pixel selection, thereby improving the change detection performance of the model in semi-supervised scenarios.

[0200] In summary, the present invention has the following beneficial effects:

[0201] This invention synchronously adjusts the teacher smoothing coefficient and the confidence threshold through a stage scheduling function, enabling the teacher network to quickly absorb effective information from the student network and expand the coverage of unlabeled pixels in the early stages of training, and enabling the teacher network to have a smoother parameter trajectory and provide more reliable pseudo-supervision signals in the later stages of training.

[0202] This invention employs a joint screening of a global prior probability map and a change prediction probability map, applying consistency constraints only to pixels where both are far from the uncertainty center and share the same direction. This approach can suppress pseudo-change supervision caused by registration errors, seasonal changes in illumination, and complex background textures.

[0203] This invention models both the change response and the steady-state response simultaneously through a differential-steady-state collaborative integration module, and uses an adaptive scaling diagram for element-by-element fusion. This can highlight the real changes while retaining the prior stable structure, and reduce over-detection caused by single differential modeling and under-detection caused by single common modeling.

[0204] This invention integrates global change prior, local boundary enhancement, and multi-scale context prediction into a prior-constrained multi-scale decoding module, which can improve the detection quality of small target changes, sparse changes, and boundary regions, and enhance the synergy of global prior, local boundary information, and multi-scale context information in the decoding process.

[0205] The various modules and their functional designs in the stage-confidence-guided change detection model of this invention have a clear synergistic relationship: the differential-stability collaborative integration module is used to separate the true change response and the cross-temporal steady-state structure at the feature level, reducing pseudo-change responses caused by registration errors, illumination differences, and seasonal variations; the prior-constrained multi-scale decoding module is used at the decoding level to jointly use global change priors, local boundary enhancements, and multi-scale contextual information for change region prediction, improving the integrity of small target changes and boundary regions; the stage-confidence mutual training module is used at the training level to synchronously adjust the smoothness of teacher parameters and the strictness of confidence pixel selection according to the training stage, and introduces unlabeled consistency constraints only at pixels where both the global prior probability map and the change prediction probability map are mutually reliable and have the same direction. Through the mutual cooperation of the various modules in the stage-confidence-guided change detection model, the model can simultaneously reduce pseudo-supervision noise and improve the utilization efficiency of unlabeled samples under a small number of pixel-level annotations, thereby improving the accuracy and stability of remote sensing image change detection.

[0206] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A semi-supervised remote sensing image change detection method guided by stage confidence, characterized in that, include: S1: Obtain a labeled sample set containing dual-temporal remote sensing images and their actual change masks, as well as an unlabeled sample set containing only dual-temporal remote sensing images; the dual-temporal remote sensing images include two remote sensing images, one from the first time phase and one from the second time phase. S2: Construct student and teacher networks with identical network structures and independently maintained parameters; the parameters of the student network are updated through gradient backpropagation, while the parameters of the teacher network are updated based on the parameters of the student network using an exponential moving average method; S3: Supervised warm-up training of the student network is performed using only the labeled sample set to obtain the warmed-up student network; S4: Copy the preheated student network parameters to the teacher network to complete the parameter initialization of the teacher network; S5: Input unlabeled samples into the current teacher network, and generate a reliable pixel mask of the unlabeled samples as a pseudo-supervised region based on the reliability judgment; S6: Input labeled samples into the student network to calculate the supervised loss, and simultaneously input unlabeled samples into the student network. Calculate the unlabeled consistency constraint loss between the student network output and the teacher network output only at the pixel positions selected by the trusted pixel mask. Update the parameters of the student network based on the supervised loss and the unlabeled consistency constraint loss. S7: Update the parameters of the teacher network based on the updated parameters of the student network; S8: Repeat steps S5 to S7 to perform semi-supervised joint training, iteratively update the parameters of the student network and the teacher network until the preset number of iterations is reached or the training loss tends to stabilize. Finally, based on the change detection accuracy on the reserved validation set, select one of the trained teacher network and student network as the change detection model. S9: Input the dual-temporal remote sensing image to be detected into the trained change detection model, and output the corresponding change detection map.

2. The semi-supervised remote sensing image change detection method guided by stage confidence as described in claim 1, characterized in that: In step S2, both the student network and the teacher network include a temporal feature extraction module for extracting multi-scale features, a differential-stability collaborative integration module for modeling the change response and steady-state response of multi-scale features, and a prior-constrained multi-scale decoding module for outputting a global prior probability map and a change prediction probability map; among them, the teacher-student semi-supervised training framework also includes a stage-based trusted mutual training module used only in the training phase. The stage-based trusted mutual training module includes a stage scheduling submodule, a trusted pixel selection submodule, and a consistency constraint submodule. The stage scheduling submodule calculates the smoothing coefficient and trusted threshold of the teacher network based on the current iteration number. The trusted pixel selection submodule generates a trusted pixel mask based on the global prior probability map and change prediction probability map of the teacher network output. The consistency constraint submodule calculates the unlabeled consistency constraint loss between the student network output and the teacher network output only at the pixel positions selected by the trusted pixel mask.

3. The semi-supervised remote sensing image change detection method guided by stage confidence as described in claim 2, characterized in that: The temporal feature extraction module includes a first temporal feature extraction submodule and a second temporal feature extraction submodule; The first temporal feature extraction submodule is used to extract multi-scale first temporal features from the first temporal remote sensing image, and the second temporal feature extraction submodule is used to extract multi-scale second temporal features from the second temporal remote sensing image. The multi-scale first-phase and second-phase features undergo lightweight channel spatial recalibration processing before entering the differential stability collaborative integration module.

4. The semi-supervised remote sensing image change detection method guided by stage confidence as described in claim 3, characterized in that: The differential-stability co-integration module is used to model the change response and steady-state response of the first-phase and second-phase features at the same scale. It includes: generating the change response based on the absolute value of the difference between the first-phase and second-phase features; generating the steady-state response based on the sum of the first-phase and second-phase features; performing lightweight convolution calibration on the change response and steady-state response respectively; and fusing the calibrated change response and steady-state response element-wise according to the scale map adaptively generated by the calibrated change response and steady-state response to obtain the multi-scale differential-stability integrated features. The formula for calculating the multi-scale differential stability integration characteristics is as follows: ； ； ； ； ； ； In the formula: and They represent the first The first and second phase characteristics of the scale; Indicates the first Response to scale changes; Indicates the first Steady-state response at scale; , They represent the first Scale-calibrated change response and steady-state response; , and Represents a lightweight mapping function; This represents the Sigmoid activation function; This represents the adaptive fusion scaling diagram; This indicates element-wise multiplication; Indicates the first The differential stability integration characteristics of scale.

5. The semi-supervised remote sensing image change detection method guided by stage confidence as described in claim 4, characterized in that: The prior-constrained multi-scale decoding module includes a global prior generation submodule, a prior-guided fusion submodule, a local boundary enhancement submodule, and a context prediction submodule. The global prior generation submodule generates a global prior branch logit and a global change prior map from the mid-to-deep differentially stable integrated features. The prior-guided fusion submodule uses the global change prior map to perform gated fusion on the shallow-to-mid-layer differentially stable integrated features. The local boundary enhancement submodule performs residual enhancement on the gated fused features to obtain the enhanced decoded features. The context prediction submodule uses a multi-scale convolutional receptive field to output a change probability map. The formula for calculating the global prior generation submodule is: ； ； In the formula: This represents the global prior fusion feature obtained by fusing the differential stability integration features from the middle to deep layers; This represents the global prior branch, logit. Represents the global prior probability graph; Represents the global prior mapping function; The calculation formula for the local boundary enhancement submodule is as follows: ； ； In the formula: This indicates the characteristics after gating fusion; Indicates the boundary enhancement mapping function; Represents a local boundary confidence plot; This represents the enhanced decoding features; The context prediction submodule first performs channel compression on the enhanced decoded features, then extracts context features through multiple parallel convolutional branches with different receptive fields, and concatenates the results of the parallel branches with the compressed features and performs 1×1 convolution to obtain context fusion features. Finally, the prediction convolution outputs the change prediction branch logit and change probability map based on the context fusion features. The formula for calculating the probability diagram of change is: ； ； In the formula: Represents context-fused features; The change prediction branch is represented by logit; Represents a probability diagram of change; This represents the context prediction mapping function.

6. The semi-supervised remote sensing image change detection method guided by stage confidence as described in claim 5, characterized in that: Step S3 specifically includes the following processing steps: S301: Obtain the labeled samples in the labeled sample set and their actual change masks; S302: Input the labeled samples into the student network, and after processing by the temporal feature extraction module, the differential stability collaborative integration module and the prior constraint multi-scale decoding module, output the global prior branch logit and the change prediction branch logit; S303: Supervised loss is calculated based on the global prior branch logit and the change prediction branch logit, respectively, and the corresponding real change mask. The formula for calculating supervised loss is: ； In the formula: This indicates a loss due to oversight; The global prior branch logit represents the student network output; The logit branch represents the prediction branch for changes in student network output; A mask representing the true changes in the labeled samples; Represents the binary cross-entropy based on logit; S304: Update the parameters of the student network using gradient backpropagation based on supervised loss; S305: Repeat steps S301 to S304 until the preset number of supervised warm-up iterations is reached or the supervised loss tends to stabilize, and the supervised warm-up student network is obtained.

7. The semi-supervised remote sensing image change detection method guided by stage confidence as described in claim 2, characterized in that: In step S7, updating the parameters of the teacher network based on the parameters of the student network means: calculating the smoothing coefficient and trust threshold of the teacher network based on the current iteration number through the stage scheduling submodule of the stage trustworthy mutual training module; and performing an exponential moving average on the parameters of the student network based on the smoothing coefficient to update the parameters of the teacher network. The calculation formula for the stage scheduling submodule is as follows: ； ； ； ； In the formula: Indicates the first The stage scheduling value for the next iteration; and For scheduling parameters; This represents the total number of iterations. and These represent the lower and upper bounds of the smoothing coefficient, respectively. and These represent the lower and upper bounds of the confidence threshold, respectively. , They represent the first The smoothing coefficient and confidence threshold for each iteration; and They represent the first The parameters of the student network and the teacher network in the next iteration; Indicates the first The parameters of the teacher network in the next iteration.

8. The semi-supervised remote sensing image change detection method guided by stage confidence as described in claim 5, characterized in that: Step S5 specifically includes the following processing steps: S501: Obtain unlabeled samples from the unlabeled sample set; S502: Input unlabeled samples into the teacher network, and after processing by the temporal feature extraction module, the differential stability collaborative integration module and the prior constraint multi-scale decoding module, output the global prior probability map and change prediction probability map corresponding to the unlabeled samples; S503: Through the trusted pixel filtering submodule of the stage trusted mutual training module, the global prior probability map and the change prediction probability map are judged on the basis of the binary classification uncertainty center. When both the global prior probability map and the change prediction probability map meet the trusted threshold requirements and are located on the same side of the binary classification uncertainty center, the corresponding pixel is determined as a trusted pixel. Then, a trusted pixel mask is generated based on all trusted pixels. The formula for calculating the reliable pixel mask is: ； In the formula: Indicates the first The next iteration's reliable pixel mask; Indicates the first The next iteration of the teacher network on unlabeled samples Output of the global prior probability map; Indicates the first The confidence threshold for the next iteration; Indicates the first The next iteration of the teacher network on unlabeled samples Probability plot of output change prediction.

9. The semi-supervised remote sensing image change detection method guided by stage confidence as described in claim 2, characterized in that: Step S6 specifically includes the following processing steps: S601: Obtain labeled samples and their true change masks from the labeled sample set, as well as unlabeled samples and their corresponding reliable pixel masks; S602: Input the labeled samples into the student network. After processing by the temporal feature extraction module, the differential-stability collaborative integration module, and the prior constraint multi-scale decoding module, output the global prior branch logit and the change prediction branch logit. Based on the global prior branch logit and the change prediction branch logit, respectively, calculate the supervised loss with the corresponding real change mask. S603: Input unlabeled samples into the student network. After processing by the temporal feature extraction module, the differential-stability collaborative integration module, and the prior constraint multi-scale decoding module, the global prior branch logit and the change prediction branch logit of the student network are output. The corresponding global prior probability map and change prediction probability map are obtained by activating them with Sigmoid respectively. S604: Through the consistency constraint submodule of the stage trusted mutual training module, the global prior probability map and change prediction probability map output by the teacher network are used as soft supervision targets. Only at the pixel position selected by the trusted pixel mask, the unlabeled consistency constraint loss is calculated based on the global prior branch logit and change prediction branch logit output by the student network, respectively. The formula for calculating the loss due to unlabeled consistency constraints is as follows: ； In the formula: This represents the loss due to unlabeled consistency constraints; Indicates the first The next iteration of the student network at the pixel The global prior branch logit is output at this point; Indicates the first The next iteration of the student network at the pixel The output change prediction branch logit; Indicates the first Next iteration pixel A reliable pixel mask at the location; Indicates the first The next iteration of the teacher network at the pixel The global prior probability map output at the location; Indicates the first The next iteration of the teacher network at the pixel The probability of change in the output at that point; Represents the binary cross-entropy based on logit; This indicates a positive number that prevents the denominator from being zero. S605: Calculate the total training loss based on supervised loss and unlabeled consistency constraint loss; The formula for calculating the total training loss is: ； In the formula: Indicates the total training loss; Indicates the weight of unlabeled consistency constraints; S606: Optimize student network parameters through gradient backpropagation based on total training loss.

10. The semi-supervised remote sensing image change detection method guided by stage confidence as described in claim 2, characterized in that: In step S9, the first and second phase remote sensing images of the dual-temporal remote sensing imagery to be detected are input into the trained change detection model. After processing by the temporal feature extraction module, the differential stability collaborative integration module, and the prior constraint multi-scale decoding module, a change probability map is obtained. The change probability map is thresholded to obtain an initial binary change map. Then, morphological filtering and connected component screening are selectively applied to the initial binary change map to obtain the final binary change detection map.

Images