Deep network for remote sensing change detection based on ddm-like brain decision mechanism

Abstract

This invention relates to the field of remote sensing image change detection, specifically a deep network for remote sensing change detection based on the brain-like decision-making mechanism of Distortion Decision Model (DDM). This invention designs a phased, multi-scale inference unit, distinct from one-time modeling, which continuously refines prior decision features through multi-scale guiding factors to simulate the iterative evidence generation process of DDM. Secondly, a stochastic representation unit is introduced, employing reparameterized sampling technology to impart randomness to features, accurately modeling the stochastic diffusion characteristics in the brain-like decision-making process. Then, a decision aggregation unit is constructed to deeply mine the temporal correlations of the multi-stage decision graph, simulating the core evidence accumulation process of DDM. Finally, a loss-guiding unit is built, driving the model to gradually focus on the real change region through multi-loss joint constraints, simulating the directional drift process of brain-like decision-making, thereby accurately locating the change region. Experiments show that this model achieves excellent results on both qualitative and quantitative metrics across multiple datasets and their perturbation datasets.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image change detection, specifically a remote sensing change detection deep network based on the DDM brain-like decision-making mechanism. Background Technology

[0002] With the continuous development of remote sensing technology and the constant improvement of image acquisition methods, change detection using remote sensing images as a data source has become a research hotspot. Remote sensing image change detection identifies changed areas by comparing images of the same region at different times, and it has wide applications in urban planning, resource management, and environmental protection.

[0003] Because dual-temporal remote sensing images are acquired at different times (or even in different seasons), the controllability of their imaging conditions is far lower than that of ordinary natural scene change detection tasks. Currently, the mainstream remote sensing change detection methods can be divided into two main categories: feature extraction and domain adaptation.

[0004] Feature extraction-based methods can be further divided into two categories: traditional methods and deep learning methods. Traditional methods (covering image arithmetic operations, image transformation, post-classification, etc.) generally rely on manually designed features, which not only struggle to overcome performance bottlenecks but also have significant shortcomings in generalization ability. Furthermore, these methods often involve key hyperparameters such as segmentation thresholds, the optimal values ​​of which are highly dependent on professional knowledge and practical experience. In contrast, deep learning, with its powerful feature representation and nonlinear fitting capabilities, has become the mainstream technology in remote sensing change detection. Typical deep network architectures include Convolutional Neural Networks (CNN), Transformer, Mamba, Diffusion Probability Model (DDPM), and their hybrid models.

[0005] Convolutional neural networks (CNNs) extract features by sliding convolutional kernels across local spatial domains. They excel at capturing local spatial information such as image edges and textures, and are highly efficient at processing data with significant local features. Typical models include fully convolutional networks (FCNs), U-Net, and their variants. However, limited by a fixed local receptive field, CNNs struggle to effectively model long-range dependencies. To address this limitation, methods based on the Transformer multi-head self-attention mechanism have been extensively studied. This mechanism enables global context modeling in both spatial and temporal dimensions, significantly enhancing the global perception capability for change detection in high-resolution remote sensing images.

[0006] Furthermore, the Mamba model innovatively introduces an omnidirectional selective scanning module, enabling adaptive scanning of images along multiple directions to effectively capture large-scale spatial features from different orientations. Its application in remote sensing change detection tasks fully demonstrates its superior spatiotemporal relationship modeling capabilities and global context capture efficiency. The core idea of ​​the diffusion model is to gradually erode the original structure of the data through a forward diffusion process, and then learn the ability to recover true data from noise through a reverse denoising process, thereby achieving accurate modeling and generation of data distribution. Current research typically utilizes multi-scale architectures to extract hierarchical features from different diffusion time steps to improve the model's robustness in complex scenarios.

[0007] While feature extraction-based deep models have made significant progress in remote sensing change detection, these methods generally face the challenge of insufficient generalization ability. When applied to cross-domain scenarios, the difference in data distribution between the source and target domains often leads to a substantial drop in model performance. To address this issue, domain adaptation methods have emerged. These methods treat bi-temporal images as source and target domain data respectively, with the core objective being to enable models trained on the source domain to effectively adapt to the target domain through specific strategies. For example, some studies decouple the features of bi-temporal images into style features and content features, and by mining change patterns in the aligned content feature space, the interference of style factors such as illumination differences on the detection results can be reduced. However, the generalization ability of existing domain adaptation methods still has limitations, and they cannot fully meet the needs of practical applications when faced with complex and variable remote sensing data. It is particularly important to note that the illumination conditions of bi-temporal remote sensing images are usually dynamically changing. When illumination differences further widen the data distribution difference between the source and target domains, the model's detection performance in the target domain may decline sharply.

[0008] Illumination conditions have a dynamic and uncertain impact on remote sensing images. Based on this characteristic, illumination perturbations in remote sensing images can be viewed as adversarial attacks in change detection tasks. From an adversarial defense perspective, improving model robustness is one effective way to address the problem of illumination perturbations in remote sensing images. In terms of improving model robustness, adversarial training augments clean data with adversarial data during training and automatically adapts deep network models to adversarial examples through learning. However, current adversarial training either focuses on filtering out adversarial perturbations or on creating fuzzy gradients, but these approaches leave room for more complex attacks, resulting in most defenses targeting only one or a few types of attacks. It is well known that the human brain is currently the only truly robust general-purpose intelligent system in nature. Clearly, drawing inspiration from the brain's information processing mechanisms to build more powerful, general-purpose, and robust machine intelligence is one of the important future development trends. Summary of the Invention

[0009] Remote sensing image change detection aims to identify regions of change in dual-time images. Traditional methods typically model and locate change regions directly in a single step, which is difficult to handle performance degradation caused by differences in imaging time (such as illumination perturbation and seasonal changes). This invention treats illumination perturbation and change detection as an "adversarial attack and defense" problem. Drawing on cognitive neuroscience and the Drift-Diffusion Model (DDM), it proposes a deep network for remote sensing change detection based on the brain-like decision-making mechanism of DDM.

[0010] This invention is achieved using the following technical solution: a remote sensing change detection deep network based on the brain-like decision-making mechanism of DDM. First, a staged multi-scale inference unit (MRU) is designed to continuously correct the prior features of decision-making through multi-scale guiding factors, simulating the evidence iterative generation process of DDM. Second, a stochastic representation unit (SRU) is introduced, which uses reparameterized sampling technology to impart randomness to features, accurately modeling the random diffusion characteristics in the brain-like decision-making process. Then, a decision aggregation unit (DAU) is constructed to deeply mine the temporal correlation of the multi-stage decision graph, realizing the simulation of the core evidence accumulation process of DDM. Finally, a loss guidance unit (LGU) is built, which drives the model to gradually focus on the real change area through multi-loss joint constraints, simulating the directional drift process of brain-like decision-making, thereby accurately locating the change.

[0011] The aforementioned remote sensing change detection deep network based on the DDM brain-like decision-making mechanism uses U-Net to map bi-temporal image features to feature spaces of different dimensions in the staged multi-scale inference unit (MRU). The second largest-scale feature output by U-Net is defined as the initial decision prior feature. C represents the channel, H represents the image height, W represents the image width, and the remaining smaller-scale features constitute the guide factor set. , As a guiding factor;

[0012] The phased multi-scale reasoning unit (MRU) in The process is executed in consecutive stages, and each stage is further divided into a guidance update module and a decision feature generation module.

[0013] The boot update module is designed to use specific boot factors As a reference, attention mechanisms can be used to analyze the prior features of decisions made in the previous stage. Perform iterative updates to generate new prior features for decision-making. The implementation process is as follows: , ,in, Indicates attention weights. , and It is a learnable projection layer that represents matrix multiplication;

[0014] The decision feature generation module aims to update the prior decision features at this stage. By incorporating dual-temporal features and enhancing feature representation through a guided attention mechanism, change-discriminating information is explicitly separated and represented, ultimately generating decision features for the current stage. The implementation process is as follows:

[0015] First, the prior characteristics of decision-making... With the dual-temporal characteristics of this stage Interaction, {1,2}, respectively generate state update features for propagation. Projective characteristics of decision reasoning ; , , , , and It is a set of parameter matrices;

[0016] Finally, a cross-reverse attention mechanism is introduced to capture the projection features of this stage. , The differences between them are used to generate the final decision features. The formula is: In the formula, , and It is a set of parameter matrices W, , and It is another set of parameter matrices W.

[0017] The aforementioned remote sensing change detection deep network based on the DDM brain-like decision-making mechanism, in the stochastic representation unit (SRU), firstly, dual-time images... Differences Figure X Through a weight Parameterized prior networks to generate prior distributions Then, the dual-time image And the label Y is input by a weight A parameterized posterior network is used to learn the posterior distribution. Meanwhile, in order to train the prior network through the posterior network, a loss function is used. To penalize the difference between the posterior and prior distributions: ;

[0018] Then, using reparameterized sampling techniques, multiple samplings are performed from the prior distribution to obtain the sampling features. where n is the number of samples;

[0019] Finally, the sampled features will be... With decision characteristics They were fed into the change detection head respectively. To obtain the decision results, Where © represents the feature concatenation operation, This represents the time step state.

[0020] The aforementioned remote sensing change detection deep network based on the DDM brain-like decision-making mechanism, in the decision aggregation unit (DAU), will store the previous hidden state. and the current time step state Input the timing memory module (TMM) to obtain the activation values ​​of the reset and update gates. , , ,in, It's a door reset. It is the update gate, and σ(·) is the Sigmoid activation function;

[0021] Based on the filtering results of the reset door, the hidden state of the previous stage is... Adjust the weights and then compare them with the current time step state. After concatenation, reconfigured features are generated through convolution and activation functions. , Where Tanh is the activation function;

[0022] By updating the weight adjustment of the gates, the hidden states accumulated in the previous stage are restored. The hidden state at the current stage is obtained by dynamically fusing it with the currently generated reset features. That is, to complete one accumulation of evidence: Through the aforementioned time-series iterative calculations, the model can gradually integrate the effective information from the multi-stage decision graph, completing a brain-like evidence accumulation process; ultimately, the hidden state output by the Temporal Memory Module (TMM) in the final stage is used as the basis for this process. This serves as the final result of the change detection.

[0023] The aforementioned remote sensing change detection deep network based on the DDM brain-like decision-making mechanism uses binary cross-entropy (BCE) loss to supervise the prediction map in the loss guidance unit (LGU). Considering the multi-stage stepwise localization characteristic of this network, in order to fully utilize the features of intermediate layers and avoid gradient vanishing or overfitting during training, the final hidden state of each stage is... With constraints applied, the BCE loss for stage i is defined as follows: ,in, This represents the pixel value of the label at position (h, w). This represents the probability that the pixel at position (h,w) in the model prediction at stage i is a changed pixel; and to integrate the constraints of multiple losses, a joint loss function is constructed. , is defined as: .

[0024] To address the performance degradation caused by illumination perturbations, this invention breaks through the limitations of traditional single-step prediction models, transforming illumination perturbation and change detection into a modeling perspective of "adversarial attack and defense." It deeply integrates human neuroscience cognitive patterns with the core characteristics of the DDM (Digital Dependent Model) to propose a deep remote sensing change detection network based on the brain-like decision-making mechanism of the DDM. First, a phased multi-scale inference unit is designed, which integrates an attention mechanism to generate decision features while iteratively updating prior features. Second, a stochastic representation unit is proposed, which learns the posterior distribution, trains the prior distribution with KL divergence as a constraint, and then obtains diverse features through reparameterized sampling. These features are then fused with the decision features to form a decision graph that combines randomness and discriminativity. Next, a decision aggregation unit is built to accumulate evidence from the multi-stage decision graph. Finally, through the directional constraints of the loss-guided unit, the model is driven to focus on the real change region, achieving modeling of directional drift. Experiments show that the network of this invention significantly outperforms existing mainstream methods in terms of comprehensive indicators and qualitative performance on multiple datasets and their perturbation datasets, fully verifying the effectiveness, robustness and strong resistance to complex lighting interference of the model, and providing a new paradigm with both accuracy and generalization ability for remote sensing change detection tasks. Attached Figure Description

[0025] Figure 1 The diagram shows the overall architecture of a remote sensing change detection deep network based on the DDM brain-like decision-making mechanism. In the diagram, a is a staged multi-scale reasoning unit: simulating the iterative evidence generation process of the brain-like decision-making mechanism; b is a random representation unit: simulating the random diffusion process; c is a decision aggregation unit: simulating the evidence accumulation process; and d is a loss guidance unit: simulating the directional drift process.

[0026] Figure 2 This is a stage architecture diagram of a multi-scale reasoning unit, where a is the guidance update module and b is the decision feature generation module.

[0027] Figure 3 This is a structural diagram of the decision aggregation unit.

[0028] Figure 4 Images showing the detection results of LEVIR-CD using different methods: (a) T1 image, (b) T2 image, (c) Ground Truth, (d) DTCDSCN, (e) BIT, (f) Change Former, (g) ICIFNet, (h) DMINet, (i) USSFC-Net, (j) Change Mamba, (k) our.

[0029] Figure 5The images show the detection results of SECOND using different methods: (a) T1 image, (b) T2 image, (c) Ground Truth, (d) DTCDSCN, (e) BIT, (f) Change Former, (g) ICIFNet, (h) DMINet, (i) USSFC-Net, (j) Change Mamba, and (k) our.

[0030] Figure 6 The images show the detection results of SYSUCD using different methods: (a) T1 image, (b) T2 image, (c) Ground Truth, (d) DTCDSCN, (e) BIT, (f) Change Former, (g) ICIFNet, (h) DMINet, (i) USSFC-Net, (j) Change Mamba, and (k) our.

[0031] Figure 7 This diagram illustrates normal and extreme lighting attacks.

[0032] Figure 8 The figure shows the quantitative experimental results of each method on three datasets with normal illumination perturbation.

[0033] Figure 9 The images show the detection results of LEVIR-CD under normal illumination attack using different methods: (a) T1 image, (b) T2 image, (c) Ground Truth, (d) DTCDSCN, (e) BIT, (f) Change Former, (g) ICIFNet, (h) DMINet, (i) USSFC-Net, (j) Change Mamba, and (k) our.

[0034] Figure 10 The images show the detection results of SECOND under normal illumination attack using different methods: (a) T1 image, (b) T2 image, (c) Ground Truth, (d) DTCDSCN, (e) BIT, (f) Change Former, (g) ICIFNet, (h) DMINet, (i) USSFC-Net, (j) Change Mamba, and (k) our.

[0035] Figure 11The images show the detection results of SYSUCD under normal illumination attack using different methods: (a) T1 image, (b) T2 image, (c) Ground Truth, (d) DTCDSCN, (e) BIT, (f) Change Former, (g) ICIFNet, (h) DMINet, (i) USSFC-Net, (j) Change Mamba, and (k) our.

[0036] Figure 12 The figure shows the quantitative experimental results of each method on three datasets with extreme lighting perturbations.

[0037] Figure 13 The images show the detection results of LEVIR-CD under extreme lighting attacks using different methods: (a) T1 image, (b) T2 image, (c) Ground Truth, (d) DTCDSCN, (e) BIT, (f) Change Former, (g) ICIFNet, (h) DMINet, (i) USSFC-Net, (j) Change Mamba, and (k) our.

[0038] Figure 14 The images show the detection results of SECOND under extreme lighting attacks using different methods: (a) T1 image, (b) T2 image, (c) Ground Truth, (d) DTCDSCN, (e) BIT, (f) Change Former, (g) ICIFNet, (h) DMINet, (i) USSFC-Net, (j) Change Mamba, and (k) our.

[0039] Figure 15 The images show the detection results of SYSUCD under extreme lighting attacks using different methods: (a) T1 image, (b) T2 image, (c) Ground Truth, (d) DTCDSCN, (e) BIT, (f) Change Former, (g) ICIFNet, (h) DMINet, (i) USSFC-Net, (j) Change Mamba, and (k) our.

[0040] Figure 16 The T-SNE results for each model are shown in the figures: (a) BIT, (b) Change Former, (c) ICIFNet, and (d) our.

[0041] Figure 17 This is a visualization of the key features of the network model, (a) a two-time image. , (b) Image Characteristics corresponding to each stage (c) Image Characteristics corresponding to each stage (d) Initial decision prior features Q and update results at each stage (e) Guiding Factor (f) Heat map of the change plot and the ground truth (GT). Detailed Implementation

[0042] A remote sensing change detection deep network based on the DDM brain-like decision-making mechanism enhances region recognition by simulating human neurocognitive patterns. Figure 1 The overall architecture of the network is shown. First, a phased multi-scale inference unit (MRU), distinct from one-time modeling, is designed to continuously refine prior decision features through multi-scale guiding factors, simulating the iterative evidence generation process of the decision-making process (DDM). Second, a stochastic representation unit (SRU) is introduced, which uses reparameterized sampling technology to impart randomness to features, accurately modeling the stochastic diffusion characteristics in the brain-like decision-making process. Then, a decision aggregation unit (DAU) is constructed to deeply mine the temporal correlations of the multi-stage decision graph, simulating the core evidence accumulation process of the DDM. Finally, a loss guidance unit (LGU) is built, which drives the model to gradually focus on the real change region through multi-loss joint constraints, simulating the directional drift process of brain-like decision-making, thereby accurately locating changes.

[0043] Multi-scale reasoning unit (MRU)

[0044] In DDM, evidence is generated and dynamically updated through multi-timestep iterations. Inspired by the human visual system and drift-diffusion models, this invention proposes a staged multi-scale reasoning unit (MRU). This unit aims to continuously refine decision-making prior features through a series of small-scale guiding factors across multiple stages, thereby simulating the iterative evidence generation process in drift-diffusion models.

[0045] This invention utilizes the powerful feature extraction capabilities of U-Net to map bi-temporal image features to feature spaces of different dimensions. The second largest-scale feature output by U-Net is defined as the initial decision prior feature. C represents the channel, H represents the image height, W represents the image width, and the remaining smaller-scale features constitute the guide factor set. ,in =3, these guiding factors carry local discriminative information at different sizes, which are used to drive the iterative update of decision prior features.

[0046] The phased multi-scale reasoning unit (MRU) in The process is executed in consecutive stages, each stage of which can be further divided into a guidance update module and a decision feature generation module, such as... Figure 1 As shown.

[0047] Bootstrapping Update Module: This module is designed to use specific boot factors. As a reference, attention mechanisms can be used to analyze the prior features of decisions made in the previous stage. Iterative updates are performed to continuously integrate multi-scale discriminative information, thereby generating new prior features for decision-making. The implementation process is as follows:

[0048] ,

[0049] ,

[0050] in, Indicates attention weights. , and It is a learnable projection layer that represents matrix multiplication.

[0051] Decision Feature Generation Module: This module aims to update the prior decision features at this stage. By incorporating dual-temporal features and enhancing feature representation through a guided attention mechanism, change-discriminating information is explicitly separated and represented, ultimately generating decision features for the current stage. ,like Figure 2 As shown.

[0052] First, a reverse softmax method is proposed, which can suppress the weights of similar regions and amplify the responses of different regions. Its calculation formula is as follows:

[0053] ,

[0054] Secondly, a two-way interaction mechanism is used to balance feature propagation with current decision-making needs. Prior decision features are then... With the dual-temporal characteristics of this stage ( The {1,2}) interaction generates state update features for transmission. Projective characteristics of decision reasoning The specific process is as follows:

[0055] ,

[0056] ,

[0057] ,

[0058] ,

[0059] It should be noted that in the above-mentioned bidirectional interaction mechanism, forward attention and reverse attention share the same set of parameter matrices W (including...). and ).

[0060] Finally, a cross-reverse attention mechanism is introduced to capture the projection features of this stage. , The differences between them are used to generate the final decision features. The formula is:

[0061]

[0062] ;

[0063] In the formula, , and It is a set of parameter matrices W, , and It is another set of parameter matrices W.

[0064] Randomized Representation Unit (SRU)

[0065] In Directed Decision Making (DDM), random diffusion refers to the noise continuously injected during the evidence accumulation process. This noise generates random walk trajectories, leading to variability in decision outcomes. To simulate this mechanism, this invention proposes a Stochastic Representation Unit (SRU). Its fusion design with the Moderately Variable Unit (MRU) enables the model to capture the inherent uncertainty and volatility in the decision-making process.

[0066] First, using dual-time images alone. Differences Figure X Through a weight Parameterized prior networks to generate prior distributions Furthermore, this prior distribution is modeled as an axis-aligned multivariate Gaussian distribution, denoted as . The mean and variance are obtained as follows:

[0067] ,

[0068] in, For learnable prior distribution generators, such as Figure 1 As shown.

[0069] In addition, in order to generate the posterior distribution, the original two-time images are... And the label Y is input by a weight A parameterized posterior network is used to learn the posterior distribution. , denoted as The mean and variance are obtained as follows:

[0070] ,

[0071] in, For learnable posterior distribution generators, such as Figure 1 As shown. To train the prior network from the posterior network, the Kullback-Leibler divergence (KL) loss is used. To penalize the difference between the posterior and prior distributions:

[0072] ,

[0073] Then, using reparameterized sampling techniques, multiple samplings are performed from the prior distribution to obtain the sampling features. (n is the number of samplings, and N is the number of stages), which ensures the differentiability of the sampling process and introduces randomness into the model.

[0074] Finally, in order to simulate the characteristics of random diffusion, the sampling features ( ) and decision-making characteristics They were fed into the change detection head respectively. The process of obtaining the decision result is as follows:

[0075] ,

[0076] Where © represents the feature concatenation operation and the time step state. This lays the foundation for the accumulation of subsequent evidence.

[0077] Decision Aggregation Unit (DAU)

[0078] In DDM (Discretionary Decision Mapping), evidence accumulation is a dynamic process that continuously integrates task-related information until a decision threshold is reached and a final decision is output. In this invention, the multiple decision graph sequences generated at each stage exhibit significant temporal dependencies, prompting the introduction of a Decision Aggregation Unit (DAU) to explore the evolutionary patterns among multi-stage decision features. This design is highly compatible with the evidence accumulation process at the physical mechanism level.

[0079] like Figure 3 As shown, the previous hidden state and the current time step state Inputting the temporal memory module (TMM) yields the activation values ​​of the reset and update gates. The reset gate filters out redundant information irrelevant to the change detection task; the update gate determines the key, valid information that should be retained and passed to subsequent stages. As shown in the formula:

[0080] ,

[0081] ,

[0082] ,

[0083] in, It's a door reset. This is the update gate, and σ(·) is the Sigmoid activation function. Based on the filtering result of the reset gate, the hidden state from the previous stage is... Adjust the weights and then compare them with the current time step state. After concatenation, reconfigured features are generated through convolution and activation functions. This feature focuses on the integration of key decision-making information at the current stage with valid historical evidence:

[0084] ,

[0085] The Tanh activation function is used to map feature values ​​to the [-1,1] interval, ensuring the stability of network training.

[0086] By updating the weight adjustment of the gates, the hidden states accumulated in the previous stage are restored. The hidden state at the current stage is obtained by dynamically fusing it with the currently generated reset features. That is, to complete one accumulation of evidence:

[0087] ,

[0088] in, Through the above time-series iterative calculations, the model can gradually integrate the effective information of the multi-stage decision graph and complete the brain-like evidence accumulation process.

[0089] Loss Guidance Unit LGU

[0090] In Directional Drift (DDM), the core characteristic of directional drift is that the decision-making process is driven by a goal-oriented signal and gradually shifts towards the correct decision direction. To simulate this characteristic, this invention constructs a Loss Guidance Unit (LGU). This unit, through phased supervision and multi-loss co-optimization, guides the model to gradually focus on the real change region while balancing model training stability and final detection accuracy, thereby achieving a brain-like directional drift effect.

[0091] The predicted map is supervised using binary cross-entropy (BCE) loss. Considering the multi-stage stepwise localization characteristic of this model, to fully utilize the features of intermediate layers and avoid gradient vanishing or overfitting during training, the final hidden state of each stage is... Apply constraints.

[0092] The BCE loss for stage i is defined as follows:

[0093] ,

[0094] in, This represents the pixel value of the label at position (h, w). This represents the probability that the pixel at position (h,w) predicted by the model in stage i is a changed pixel.

[0095] To integrate the constraints of multiple losses, this invention constructs a joint loss function. , is defined as:

[0096] ,

[0097] This loss function, through multiple constraints, ensures the stability of training at each stage while clarifying the optimization direction of the model. It aligns perfectly with the directional drift characteristics of DDM, ultimately yielding the output of the final stage. This serves as the final result of the change detection.

[0098] Experimental setup

[0099] A. Dataset

[0100] To evaluate the effectiveness of the network, experiments were conducted using three remote sensing image change detection (CD) benchmark datasets (LEVIRCD, SECOND, and SYSUCD). Each dataset contains two remote sensing images of the same geographic area acquired at different times, along with corresponding change detection labels.

[0101] B. Experimental parameters

[0102] The network is implemented using the PyTorch framework and trained and tested on a single NVIDIA RTX4090 GPU. Model optimization utilizes the AdamW optimizer with a weight decay of 5e-4 and an initial learning rate of 0.001. The optimizer's learning rate is dynamically adjusted, decaying to 0.7 times its original value every 20 training epochs. Furthermore, the total number of training epochs for each dataset is set to 150, with a batch size of 16. Validation is performed after each training epoch, and the model with the best performance on the validation set is used to evaluate the model on the test set.

[0103] C. Evaluation Indicators

[0104] In terms of objective evaluation, this invention uses five evaluation indicators—precision (P), recall (Re), F1, overall accuracy (OA), and mean intersection-union ratio (mIoU)—to comprehensively evaluate the experimental results.

[0105] Experimental results

[0106] A. Comparison of conventional data

[0107] 1) Objective indicators

[0108] To verify the effectiveness of the network on bi-temporal remote sensing images, it was compared with several state-of-the-art CD methods, including DTCDSCN, BIT, Change Former, ICIFNet, USSFC-Net, DMINet, and Change Mamba.

[0109] Table 1 Performance comparison of different methods on different datasets

[0110]

[0111] Table 1 shows the quantitative metrics of different models on three clean, attack-free datasets (LEVIR-CD, SECOND, and SYSUCD). The network of this invention performs excellently on comprehensive metrics such as F1 score and mIou across multiple datasets. For example, it achieves an F1 score of 90.03 on the LEVRICD dataset. This may be attributed to the DDM brain-like decision-making mechanism, which is highly consistent with human cognition in detecting changes in remote sensing images, and it has gradually achieved accurate localization of changed regions.

[0112] 2) Visual comparison

[0113] The detection results of each method on the LEVIRCD, SECOND, and SYSUCD datasets are as follows: Figure 4-6 As shown, many algorithms, such as DMINet and ChangeMamba, are prone to false positives when detecting small targets. Models DTCDSCN and BIT exhibit varying degrees of false positives near edges when detecting multiple changing targets. For large target regions, USSFC-Net shows significant false positives. This invention designs a staged multi-scale inference unit, using a series of guiding factors to update the decision prior features, ultimately achieving a more ideal qualitative result.

[0114] B. Comparison under normal lighting attack

[0115] To test the model's robustness under normal lighting attacks, a lighting perturbation test set was created. First, the differences in the HSV color spaces (H represents hue, S represents saturation, L represents brightness) of the three real datasets, LEVIRCD, SECOND, and SYSUCD, were analyzed, with ranges of [-5.25, 12.57], [4.65, 30.82], and [-28.81, 29.95], respectively. These HSV color space differences, obtained from the original datasets, can effectively simulate lighting changes under some real-world conditions. Then, within these ranges, the color space values ​​of the pixels in each test set image were randomly changed to generate a lighting perturbation test set. For example... Figure 7 As shown, this test set can simulate weather conditions under various environments.

[0116] 1) Objective indicators

[0117] Figure 8 Quantitative metrics for different models on a test set with illumination perturbations are presented. Under normal illumination perturbations, our model still maintains the leading F1 score and mIoU across all datasets. This is attributed to the fact that the proposed stochastic representation unit effectively simulates random perturbations caused by noise such as illumination, thereby significantly improving the model's robustness to interference.

[0118] 2) Visual comparison

[0119] The detection results of each method on the ordinary illumination perturbation test sets of LEVIRCD, SECOND, and SYSUCD are as follows: Figure 9-11 Adding illumination perturbation attacks further impacts the detection performance of algorithms. Most methods, such as BIT, ICIFNet, DMINet, and Change Mamba, experience some degree of false positives and false negatives compared to before the perturbation. The model proposed in this invention demonstrates superior performance in terms of target integrity, false positives, and false negatives, likely due to the loss-guided unit. This unit simulates the directional drift process of the drift-diffusion model, focusing change-related information and suppressing irrelevant redundant information through multiple loss constraints, thereby improving the accuracy of change area localization and the reliability of detection results.

[0120] C. Comparison under extreme lighting attacks

[0121] In addition, to further verify the ability to resist interference, the interference amplitude was deliberately increased to induce visible light fluctuations in order to obtain a corresponding test set. For example... Figure 7 As shown, this test set can simulate weather conditions in extreme environments to evaluate the model's performance in harsh environments.

[0122] 1) Objective indicators

[0123] Figure 12 Quantitative metrics for different models on a test set with extreme lighting perturbations are presented. Under extreme lighting perturbations, most models (Change Mamba, USSFC-Net) show a significant decrease in F1 and mIoU, while our model consistently maintains the highest F1 and mIoU. This is attributed to the decision aggregation unit's sequential aggregation of multiple decision graphs, which improves model robustness while preserving details of the changed regions.

[0124] 2) Visual comparison

[0125] The detection results of each method on the extreme illumination perturbation test sets of LEVIRCD, SECOND, and SYSUCD are as follows: Figure 13-15Theoretically, while the data content remains unchanged, the pixel distribution undergoes significant changes. Many models, such as USSFC-Net and Change Mamba, exhibit numerous false negatives under extreme lighting attacks. Simultaneously, ICIFNet and Change Former also show a substantial increase in false positive rates. Although this model also experienced some minor false negatives and false positives, it generally demonstrates strong robustness against interference. This may be attributed to the invention's approach of treating lighting perturbation and change detection as an adversarial attack and defense problem, thereby enhancing the model's robustness by learning from human cognitive patterns.

[0126] C. Ablation test

[0127] 1) Ablation of SRU and DAU

[0128] Table 2 Module Ablation Experiment

[0129]

[0130] This invention validated the effectiveness of each module (Random Representation Unit (SRU) and Decision Aggregation Unit (DAU)) through ablation experiments, as shown in Table 2, where √ indicates the corresponding module was used. The random sampling operation of the SRU simulates a random diffusion mechanism, and the DAU introduces evidence accumulation functionality into the model, which are precisely the key elements of DDM. This demonstrates that DDM contributes to improving model performance.

[0131] 2) Ablation of distributed sample numbers

[0132] Table 3. Number of samples for ablation experiments

[0133]

[0134] To enhance feature diversity, this model integrates both convolutional features and distributed sampling features. Distributed sampling is achieved by selecting different numbers of features with a fixed number of iterations. Table 3 analyzes the impact of the number of sampled features on objective evaluation. It is evident that as the number of distributed samples increases, the model performance exhibits a continuous downward trend. This phenomenon may stem from the fact that excessively large feature samples exacerbate the redundancy in the feature space, thereby interfering with the model's learning of effective decision features and compromising the stability of subsequent decision sets.

[0135] 3) Ablation of distributed sampling times

[0136] Table 4. Distribution of sampling times in ablation experiments

[0137]

[0138] To introduce controllable randomness to simulate decision-making perturbations, this invention performs multiple independent random samplings of the distribution and fuses each sampling result with the decision features separately. Table 4 analyzes the impact of the number of samplings on objective evaluation. Based on the previous experiments that have confirmed the effectiveness of the random diffusion unit, the number of samplings in this experiment starts from 2. It is easy to see that when the number of samplings increases from 2 to 3, the model performance improves to some extent; however, when the number exceeds 3, the performance begins to decline slightly. Clearly, moderately increasing the number of samplings can effectively enrich the randomness of the model and improve its adaptability to complex disturbances; however, when the number of samplings is too high, multiple fusion processes introduce excessive redundant information, which not only increases the computational burden of the model but may also lead to overfitting of the model to local noise features, thus weakening the core decision-making ability.

[0139] 4) Stage-based ablation

[0140] Table 5 Stage-based ablation experiments

[0141]

[0142] Table 5 analyzes the impact of the number of iteration stages N. Since the previous experiments have confirmed the effectiveness of the evidence aggregation unit, the number of stages in this experiment started from 2. It is easy to see that performance improves from N=2 to N=3, but significantly decreases when N>3. This may be because as the number of stages increases, very small-scale features appear, making it difficult to learn more accurate regions of change.

[0143] D. Parameter Analysis

[0144] Table 6 Experimental Loss Function Coefficients

[0145]

[0146] To verify the effectiveness of each loss function and obtain its optimal parameter configuration, experimental tuning of the loss function regularization was performed on the LEVIRCD dataset. As shown in Table 6, The optimal parameter combination is (1,1). Then, this was extended to the remaining two datasets, and good performance metrics were still achieved. This indicates that the model has good generalization performance and is not overly sensitive to regularization parameters.

[0147] E. Visualization

[0148] 1) Decoupling Spatial Visualization

[0149] To more intuitively demonstrate the effectiveness of change detection, T-SNE dimensionality reduction technology was used to project changed and invariant features onto a two-dimensional plane for visualization analysis. For example... Figure 16As shown, yellow and blue represent changing and invariant pixel features, respectively. The visualization results demonstrate that this model's semantic boundary clarity is significantly superior to other models. Qualitatively, this model performs better on multiple metrics, which aligns with the visualization results. This demonstrates that by mimicking the cognitive patterns of human neuroscience, the model can more accurately distinguish between changing and invariant pixels, thereby effectively enhancing its ability to differentiate semantic boundaries.

[0150] 2) Characteristic heatmap

[0151] To provide a more intuitive understanding of the model, a representative set of samples was selected from the LEVIRCD dataset to visualize the feature maps generated by the model at different stages, such as... Figure 17 As shown. From Figure 17 (bd) It can be observed that as the number of stages increases, the decision prior features gradually achieve accurate differentiation between changing and invariant regions. Furthermore, the dual-time image features gradually focus on changing regions. In addition, the generated change heatmap [ Figure 17 (f)] The model exhibits highly accurate target localization and fine boundary, indicating that by incorporating the DDM brain-like decision-making mechanism, it can effectively capture changing targets and improve the localization accuracy of changing areas.

Claims

1. A remote sensing change detection deep network based on the DDM brain-like decision-making mechanism, characterized in that: First, a phased multi-scale reasoning unit (MRU) was designed to continuously modify the prior features of decision-making through multi-scale guiding factors, simulating the iterative evidence generation process of the decision-making process (DDM). Second, a stochastic representation unit (SRU) was introduced, which, relying on reparameterized sampling technology, endows features with randomness and accurately models the stochastic diffusion characteristics in the brain-like decision-making process. Then, a decision aggregation unit (DAU) was constructed to deeply mine the temporal correlation of the multi-stage decision graph and realize the simulation of the core evidence accumulation process of the DDM. Finally, a loss guidance unit (LGU) was built, which drives the model to gradually focus on the real change area through multi-loss joint constraints, simulating the directional drift process of brain-like decision-making, thereby accurately locating the change.

2. The remote sensing change detection deep network based on the DDM brain-like decision-making mechanism according to claim 1, characterized in that: In the staged multi-scale inference unit (MRU), U-Net is used to map bi-temporal image features to feature spaces of different dimensions; the second largest-scale feature output by U-Net is defined as the initial decision prior feature. C represents the channel, H represents the image height, W represents the image width, and the remaining smaller-scale features constitute the guide factor set. , As a guiding factor; The phased multi-scale reasoning unit (MRU) in The process is executed in consecutive stages, and each stage is further divided into a guidance update module and a decision feature generation module. The boot update module is designed to use specific boot factors As a reference, attention mechanisms can be used to analyze the prior features of decisions made in the previous stage. Perform iterative updates to generate new prior features for decision-making. The implementation process is as follows: , ,in, Indicates attention weights. , and It is a learnable projection layer that represents matrix multiplication; The decision feature generation module aims to update the prior decision features at this stage. By incorporating dual-temporal features and enhancing feature representation through a guided attention mechanism, change-discriminating information is explicitly separated and represented, ultimately generating decision features for the current stage. The implementation process is as follows: First, the prior characteristics of decision-making... With the dual-temporal characteristics of this stage Interaction, {1,2}, respectively generate state update features for propagation. Projective characteristics of decision reasoning ; , , , , and It is a set of parameter matrices; Finally, a cross-reverse attention mechanism is introduced to capture the projection features of this stage. , The differences between them are used to generate the final decision features. The formula is: In the formula, , and It is a set of parameter matrices W, , and It is another set of parameter matrices W.

3. The remote sensing change detection deep network based on the DDM brain-like decision-making mechanism according to claim 2, characterized in that: In the stochastic representation unit (SRU), firstly, the dual-time image... The difference graph X is obtained through a weighted graph. Parameterized prior networks to generate prior distributions Then, the dual-time image And the label Y is input by a weight A parameterized posterior network is used to learn the posterior distribution. Meanwhile, in order to train the prior network through the posterior network, a loss function is used. To penalize the difference between the posterior and prior distributions: ; Then, using reparameterized sampling techniques, multiple samplings are performed from the prior distribution to obtain the sampling features. where n is the number of samples; Finally, the sampled features will be... With decision characteristics They were fed into the change detection head respectively. To obtain the decision results, Where © represents the feature concatenation operation, This represents the time step state.

4. The remote sensing change detection deep network based on the DDM brain-like decision-making mechanism according to claim 3, characterized in that: In the decision aggregation unit (DAU), the previous hidden state is... and the current time step state Input the timing memory module (TMM) to obtain the activation values ​​of the reset and update gates. , , ,in, It's a door reset. It is the update gate, and σ(·) is the Sigmoid activation function; Based on the filtering results of the reset door, the hidden state of the previous stage is... Adjust the weights and then compare them with the current time step state. After concatenation, reconfigured features are generated through convolution and activation functions. , Where Tanh is the activation function; By updating the weight adjustment of the gates, the hidden states accumulated in the previous stage are restored. The hidden state at the current stage is obtained by dynamically fusing it with the currently generated reset features. That is, to complete one accumulation of evidence: Through the aforementioned time-series iterative calculations, the model can gradually integrate the effective information from the multi-stage decision graph, completing a brain-like evidence accumulation process; ultimately, the hidden state output by the Temporal Memory Module (TMM) in the final stage is used as the basis for this process. This serves as the final result of the change detection.

5. The remote sensing change detection deep network based on the DDM brain-like decision-making mechanism according to claim 4, characterized in that: In the loss-guided unit (LGU), binary cross-entropy (BCE) loss is used to supervise the prediction map. Considering the multi-stage progressive localization characteristic of this network, in order to fully utilize the features of intermediate layers and avoid gradient vanishing or overfitting during training, the final hidden state of each stage is... With constraints applied, the BCE loss for stage i is defined as follows: ,in, This represents the pixel value of the label at position (h, w). This represents the probability that the pixel at position (h,w) in the model prediction at stage i is a changed pixel; and to integrate the constraints of multiple losses, a joint loss function is constructed. , is defined as: .

Images