SAR target robust identification method and system based on causal reasoning under background change

Abstract

The invention discloses an SAR target robust identification method and system based on causal reasoning under background change. According to the method, a causal model for SAR target identification is constructed, an SAR image, an observable background and unobservable hybrid variables are incorporated into a unified framework, and a causal path, a back door path and a front door path are distinguished. For an observable background, extracting a background sample by adopting a backdoor adjustment strategy, and carrying out weighted fusion on the background sample and an original image to generate a first intervention sample so as to eliminate false correlation between the background and a target category; and aiming at the unobservable hybrid variables, designing a front door adjustment strategy to construct a second intervention sample, calculating a total causal effect, and weakening the influence by using a progressive intervention strategy. A classification network is constructed based on a CNN, training optimization is carried out by combining two intervention samples, and finally the recognition performance is verified on an EOC-Scene data set. According to the method, the recognition robustness and generalization ability of the SAR target under the complex background change are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of synthetic aperture radar target recognition technology, and more specifically, to a robust SAR target recognition method and system based on causal reasoning under changing background conditions. Background Technology

[0002] Synthetic Aperture Radar (SAR) target recognition algorithms have significant application value in military target identification, vehicle and ship monitoring, and other tasks, providing reliable target identification information for scenarios such as surveillance, precision strikes, and emergency rescue. The core objective of target recognition algorithms is to achieve stable and accurate target discrimination under complex observation conditions, and one of the key challenges is robust target recognition under complex background changes.

[0003] In practical applications, SAR imaging systems offer advantages such as being unaffected by lighting and weather conditions, and possessing all-day, all-weather imaging capabilities, making them particularly suitable for target monitoring in complex surface and marine environments. However, the SAR imaging process is highly dependent on the electromagnetic scattering characteristics of the target and its surrounding background. When the observation scene changes, differences in surface type, environmental structure, and clutter distribution significantly affect the composition of the echo signal, altering the contrast between target scattering and background scattering. This poses a serious challenge to the stability and generalization ability of the recognition algorithm.

[0004] From the perspective of imaging mechanisms, SAR images are essentially a comprehensive reflection of the backscattering of radar electromagnetic waves by ground targets and their surrounding environment. Factors affecting backscattering characteristics mainly fall into two categories: one is radar system parameters, such as operating wavelength, incident angle, and polarization; the other is the characteristics of the ground objects and environment, including target structure, material properties, and the roughness, dielectric constant, and spatial distribution of the background area. Under the condition that radar system parameters remain consistent, changes in the background environment often introduce significant differences in scattering, resulting in targets exhibiting distinctly different imaging characteristics against different backgrounds.

[0005] Especially under complex or unknown background conditions, background clutter can be highly coupled with the target in terms of energy distribution, texture structure, or statistical properties, interfering with the target's discrimination features and even causing the target to be dominated by background features. This background correlation makes data-driven recognition models prone to learning background bias, resulting in significant performance degradation when the training and testing backgrounds are inconsistent. In practical applications, although sufficient labeled data of some targets under specific background conditions can be obtained, how to effectively utilize existing data, suppress background interference, and extract background-invariant discrimination features has become one of the key problems that urgently need to be solved in the field of SAR target recognition when facing application scenarios with diverse, frequently changing, or even unknown background types.

[0006] In target recognition, traditional SAR target recognition methods typically implicitly assume that training and test data have similar background distributions, failing to adequately consider the impact of complex background variations on classifier performance. These methods often treat samples under different background conditions uniformly, encoding both target-specific discriminative information and background-related information during feature extraction, without effectively distinguishing or constraining the two. The resulting feature representations inevitably contain a large amount of background-dependent components, making it easy for the model to learn biased features strongly correlated with specific backgrounds during training, rather than truly stable target identity features. This feature representation struggles to maintain discriminative consistency when the background distribution changes, leading to a significant decline in the model's recognition performance under unknown or changing backgrounds.

[0007] To address the issue of background variation, some studies have improved the adaptability of models by constructing multi-scene or multi-background training sets. This involves introducing target samples under various background conditions during the training phase to learn more robust feature representations. However, these methods typically rely on uniformly and sufficiently collecting training samples under multiple background conditions. Their data acquisition process is often based on idealized scenario assumptions, which are difficult to meet in practical applications. Especially for SAR systems with limited observation resources and high imaging costs, systematically acquiring high-quality labeled data covering multiple complex backgrounds is both costly and practically challenging.

[0008] On the other hand, existing multi-background or multi-scene methods still lack explicit decoupling between "target-related information" and "background-related information" at the feature modeling level. Most methods employ uniform feature extraction and constraint strategies, failing to apply differentiated modeling and suppression mechanisms for different types of features. This makes the models susceptible to background interference when faced with significant changes in background distribution or even unknown backgrounds, thus limiting their generalization ability in real-world complex environments. Therefore, how to effectively reduce the interference of background factors under changing background conditions and extract target discrimination features with background invariance has become a crucial problem that urgently needs to be solved in the field of SAR target recognition.

[0009] Traditional target recognition algorithms often fail to effectively distinguish between target-related information and background-related information during feature extraction, resulting in extracted features inevitably containing a large number of components influenced by background changes. These methods tend to model the target and its surrounding environment holistically, making the model prone to relying on discrimination patterns formed under specific background conditions during training. When the background distribution changes, the stability of the feature representation decreases, leading to a significant degradation in target identity discrimination performance and making it difficult to extract robust target representations under complex or unknown background conditions.

[0010] While deep learning-based discriminative networks can automatically learn feature representations in an end-to-end manner, mitigating the limitations of manual feature design to some extent, their feature learning process often lacks explicit modeling and constraints on background interference, resulting in insufficient interpretability. Because the network may implicitly encode background-related patterns, its recognition decisions are often difficult to explain clearly from a physical or semantic perspective, and unstable recognition results may still occur when background conditions change.

[0011] In multi-scene or multi-background learning tasks, some methods improve the model's adaptability by introducing training samples under various background conditions. These methods achieve good results under ideal conditions with relatively uniform background distribution and sufficient coverage, but their training data typically assumes that different background types appear evenly in the samples. However, in practical applications, especially in SAR target detection and recognition tasks, a large number of target samples are often only available under limited background conditions, while effective training data is lacking in other complex or rare backgrounds, leading to a significant imbalance or even continuous absence of background distribution. When background types that do not appear in the training set or are insufficiently covered appear during the testing phase, existing multi-background methods often struggle to cope effectively.

[0012] In target recognition problems with widely varying or even unknown backgrounds, it is often necessary to jointly train the model using existing samples of known background categories and samples with limited background conditions to achieve stable target recognition under various background conditions. This places higher demands on feature representation, requiring the extraction of robust features from the training data that are as independent as possible of background changes and only represent the target's identity attributes, thereby ensuring that the model still has reliable discrimination ability under complex background conditions. Summary of the Invention

[0013] The main objective of this invention is to provide a robust SAR target identification method and system based on causal reasoning under changing background conditions, so as to at least solve the problem of weak generalization ability in the prior art and enhance the ability to reduce the interference of background factors under changing background conditions.

[0014] To achieve the above objectives, a robust SAR target identification method and system based on causal reasoning under changing background conditions is provided.

[0015] In a first aspect, the present invention provides a robust SAR target identification method based on causal reasoning under changing background conditions, the identification method comprising:

[0016] A causal model for SAR target recognition is constructed, and multiple variables in the SAR target recognition process are defined. These multiple variables are then integrated with the causal model. The multiple variables include at least: SAR image, observable background, and unobservable confounding variables.

[0017] For observable background, a backdoor adjustment strategy is designed to extract background samples from SAR images. The background samples are then weighted and fused with the SAR images to generate the first intervention sample. The backdoor adjustment strategy is used to eliminate spurious correlations between the background and target categories in the SAR images.

[0018] For unobservable confounding variables, a front-door adjustment strategy is designed. A second intervention sample is constructed through the front-door adjustment strategy, and the total causal effect is calculated based on the second intervention sample and SAR image. Based on the total causal effect, a progressive intervention strategy is used to eliminate the influence of unobservable confounding variables on the SAR target recognition process.

[0019] A classification network is constructed based on CNN. SAR images are input into the classification network to extract image features. The classification network is then intervened based on the first intervention sample and the second intervention sample. Backpropagation is used to update the parameters of the classification network to obtain the trained classification network.

[0020] Obtain the EOC-Scene dataset and use it to test the trained classification network to verify its effectiveness.

[0021] Specifically, multiple variables in the SAR target identification process are defined, and these variables are fused with a causal model, including:

[0022] Define the SAR image as variable X, the observable background as B, and the unobservable clutter variable as U;

[0023] Establish causal relationships between multiple variables and the causal model, including causal paths, backdoor paths, and frontdoor paths.

[0024] Specifically, causal path represents the path by which SAR images influence prediction results through feature extraction;

[0025] The backdoor path represents the path through which the observable background simultaneously influences image formation and category prediction;

[0026] The front door path represents the path where unobservable confounding variables simultaneously affect both the input and the label.

[0027] Specifically, for the observable background, a backdoor adjustment strategy is designed to extract background samples from the SAR image, including:

[0028] Calculate the channel average intensity map of the SAR image;

[0029] A preset quantile threshold is used to calculate the intensity threshold based on the quantile threshold and the channel average intensity map.

[0030] A binary mask is generated based on the intensity threshold, and background samples of the SAR image are extracted based on the binary mask.

[0031] Specifically, the background samples and SAR images are weighted and fused to generate the first intervention sample, including:

[0032] The fusion weights are obtained by fusing SAR images and background samples using an exponential weighting method.

[0033] The first intervention sample is generated based on the fusion weights.

[0034] Specifically, for unobservable confounding variables, a front-door adjustment strategy is designed to construct a second intervention sample. The total causal effect is then calculated based on the second intervention sample and the SAR image, including:

[0035] The mean of the SAR images is calculated to construct a second intervention sample;

[0036] The outputs of the classification network, which uses SAR images and second intervention samples as inputs, are calculated separately to obtain the first output and the second output.

[0037] The total causal effect is calculated based on the first and second output results.

[0038] Specifically, the progressive intervention strategy uses the Sigmiod function to control the intensity of intervention by unobservable confounding variables on the SAR target identification process.

[0039] Specifically, a classification network is built based on CNN, including:

[0040] A feature extraction layer for a classification network based on CNN;

[0041] It facilitates the construction of the output layer of a classification network using fully connected layers and the Softmax function;

[0042] Cross-entropy loss is used as the loss function for the classification network.

[0043] Specifically, the trained classification network was tested using the EOC-Scene dataset to verify its effectiveness, which was expressed as target recognition accuracy, top-five accuracy, recall, and F1 score.

[0044] In a second aspect, the present invention provides a robust SAR target identification system based on causal reasoning under changing background conditions. The identification system is applied to the identification method of the first aspect, and the identification system includes:

[0045] The causal model building module is used to build a causal model in the SAR target recognition process. It defines multiple variables in the SAR target recognition process and integrates these multiple variables with the causal model. The multiple variables include at least: SAR image, observable background and unobservable confounding variables.

[0046] The backdoor adjustment strategy design module is connected to the causal model construction module. The backdoor adjustment strategy design module is used to design backdoor adjustment strategies for observable backgrounds. The backdoor adjustment strategy extracts background samples from SAR images and performs weighted fusion with SAR images to generate the first intervention sample. The backdoor adjustment strategy is used to eliminate false correlations between background and target categories in SAR images.

[0047] The front-door adjustment strategy design module is connected to the back-door adjustment strategy design module. The front-door adjustment strategy design module is used to design a front-door adjustment strategy for unobservable confounding variables, construct a second intervention sample through the front-door adjustment strategy, calculate the total causal effect based on the second intervention sample and SAR image, and use a progressive intervention strategy based on the total causal effect to eliminate the influence of unobservable confounding variables on the SAR target recognition process.

[0048] The classification network construction and training module is connected to the front door adjustment strategy design module. The classification network construction and training module is used to build a classification network based on CNN. The SAR image is input into the classification network to extract image features. The classification network is then intervened based on the first intervention sample and the second intervention sample. Backpropagation is used to update the parameters of the classification network to obtain the trained classification network.

[0049] The classification network testing and evaluation module is connected to the classification network construction and training module. The classification network testing and evaluation module is used to obtain the EOC-Scene dataset and use the EOC-Scene dataset to test the trained classification network to verify the effect of the trained classification network.

[0050] This application provides a robust SAR target recognition method and system based on causal reasoning under changing background conditions. The method constructs a causal model for SAR target recognition, distinguishes between observable background and unobservable confounding variables, and designs back-door adjustment and front-door adjustment strategies respectively: back-door adjustment extracts background samples and weightedly fuses them with the original image to generate intervention samples, eliminating spurious correlations between background and target categories; front-door adjustment constructs another intervention sample and calculates the total causal effect, using a progressive intervention strategy to reduce the influence of unobservable confounding variables. Subsequently, a classification network is constructed based on CNN, and feature extraction and network parameter updates are performed using both intervention samples. Finally, the recognition performance is evaluated on the EOC-Scene dataset, thus achieving robust SAR target recognition under changing background conditions. Attached Figure Description

[0051] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0052] Figure 1 A flowchart illustrating a robust SAR target identification method based on causal reasoning under changing background conditions, provided in this application;

[0053] Figure 2 A schematic diagram of the overall framework of a robust SAR target identification method based on causal reasoning under changing background conditions provided in this application;

[0054] Figure 3 A schematic diagram of the backdoor adjustment process for a robust SAR target identification method based on causal reasoning under changing background conditions provided in this application;

[0055] Figure 4 A schematic diagram of the target recognition subnetwork structure for a robust SAR target recognition method based on causal reasoning under changing background conditions, provided in this application;

[0056] Figure 5 A schematic diagram of the target transformation subnetwork structure for a robust SAR target recognition method based on causal reasoning under changing background conditions, provided in this application;

[0057] Figure 6 The image shows the generation effect of the reconstruction task and the transformation task in the target transformation subnetwork of a SAR target robust recognition method based on causal reasoning under background changes provided in this application.

[0058] Figure 7 The optimal parameter search graph for a robust SAR target identification method based on causal reasoning under changing background conditions is provided in this application.

[0059] Figure 8 This is a connection diagram of a robust SAR target identification system based on causal reasoning under changing background conditions, provided in this application. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0061] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.

[0062] In this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0063] This application provides a robust SAR target recognition method and system based on causal inference under changing background conditions. The method first constructs a causal model encompassing SAR images, observable background, and unobservable confounding variables. For the observable background, a backdoor adjustment strategy is designed, extracting background samples from the image and weightedly fusing them with the original image to generate intervention samples, thereby severing spurious associations between the background and target categories. For unobservable confounding variables, a frontdoor adjustment strategy is used to construct another intervention sample, calculating the total causal effect, and employing a progressive intervention strategy to weaken its influence. Based on this, a CNN classification network is trained, and the model parameters are optimized by incorporating the aforementioned intervention mechanisms. Finally, the improved recognition accuracy, recall, and generalization ability are validated on the EOC-Scene dataset.

[0064] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0065] Figure 1 A flowchart illustrating a robust SAR target identification method based on causal reasoning under changing background conditions, as provided in this application, is shown below. Figure 1 As shown in this embodiment, a robust SAR target identification method based on causal reasoning under changing background conditions is provided. This identification method includes:

[0066] A causal model for SAR target recognition is constructed, and multiple variables in the SAR target recognition process are defined. These multiple variables are then integrated with the causal model. The multiple variables include at least: SAR image, observable background, and unobservable confounding variables.

[0067] For observable background, a backdoor adjustment strategy is designed to extract background samples from SAR images. The background samples are then weighted and fused with the SAR images to generate the first intervention sample. The backdoor adjustment strategy is used to eliminate spurious correlations between the background and target categories in the SAR images.

[0068] For unobservable confounding variables, a front-door adjustment strategy is designed. A second intervention sample is constructed through the front-door adjustment strategy, and the total causal effect is calculated based on the second intervention sample and SAR image. Based on the total causal effect, a progressive intervention strategy is used to eliminate the influence of unobservable confounding variables on the SAR target recognition process.

[0069] A classification network is constructed based on CNN. SAR images are input into the classification network to extract image features. The classification network is then intervened based on the first intervention sample and the second intervention sample. Backpropagation is used to update the parameters of the classification network to obtain the trained classification network.

[0070] Obtain the EOC-Scene dataset and use it to test the trained classification network to verify its effectiveness.

[0071] This application provides a robust SAR target recognition method based on causal inference under changing background conditions. This method establishes a causal model for SAR target recognition, incorporating SAR images, observable background, and unobservable confounding variables into a unified analytical framework. For the observable background, a backdoor adjustment strategy is employed, extracting background samples and weightedly fusing them with the original image to generate intervention samples, thereby blocking spurious correlations between the background and target categories. For unobservable confounding variables, a frontdoor adjustment strategy is designed, constructing new intervention samples and calculating the total causal effect, then suppressing their interference through a progressive intervention strategy. Subsequently, a classification network is constructed based on a CNN, incorporating the aforementioned intervention samples into the training process. Backpropagation is used to optimize the network parameters, and finally, the model's recognition performance is evaluated on the EOC-Scene dataset.

[0072] This method systematically introduces front-door and back-door adjustment strategies for causal inference into the field of SAR target recognition, theoretically distinguishing and handling two different types of interference factors: observable background and unobservable confounding variables. Through proactive data generation and a progressive optimization mechanism, it effectively eliminates spurious correlations unrelated to the target's essence, enhancing the model's generalization ability and robustness under complex background changes, thereby improving the accuracy and reliability of SAR target recognition.

[0073] Specifically, multiple variables in the SAR target identification process are defined, and these variables are fused with a causal model, including:

[0074] Define the SAR image as variable X, the observable background as B, and the unobservable clutter variable as U;

[0075] Establish causal relationships between multiple variables and the causal model, including causal paths, backdoor paths, and frontdoor paths.

[0076] This application provides a robust SAR target recognition method based on causal reasoning under changing background conditions. This method establishes a formalized causal model to analyze the SAR target recognition process. Specifically, it defines the input data as key variables: the SAR image is variable X, the observable background is B, and the unobservable confounding factors (such as imaging conditions, target internal state, etc.) are U. Based on this, the method explicitly distinguishes and models three key causal paths: the direct causal path from X to the prediction result; the backdoor path where the background B simultaneously influences the formation and final prediction of X; and the frontdoor path where the confounding variable U simultaneously influences the input X and the true label. This structured modeling lays the theoretical foundation for subsequently targeted removal of non-causal spurious associations.

[0077] Construct a structured causal model and define the following variables: U: unobservable variables (such as weather, noise interference); B: background bias variables (observable); X: input SAR image samples; F: features extracted from the network; Y: predicted class label.

[0078] Directed acyclic graphs (DAGs) are used to represent causal relationships between variables and to identify interference paths of background factors on the identification process.

[0079] Figure 2 A schematic diagram illustrating the overall framework of a robust SAR target identification method based on causal reasoning under changing background conditions, as provided in this application, is shown below. Figure 2 The diagram shown is a schematic representation of the overall framework provided in this embodiment. This overall framework includes:

[0080] Observable variables are represented by solid circles, while unobservable (latent) variables are represented by dashed circles. Variable U represents unobservable factors that are difficult to measure precisely, such as weather conditions or noise interference. Variable B reflects the inherent background bias in the dataset. Variable X represents the input SAR sample to be identified, F represents the features extracted from the network, and Y represents the predicted label. Arrows in the DAG indicate causal relationships between these variables. This structured causal model provides a high-level abstraction for analyzing the impact of background interference on the identification process, laying the foundation for developing effective anti-detection strategies.

[0081] LINK 1 X→F→Y: The causal relationship between the input sample X and the predicted label Y is moderated by the feature representation F, reflecting an indirect causal effect. In real-world recognition tasks, the target category cannot be directly determined from the raw samples alone; what truly serves as the basis for prediction are abstracted high-level features, such as shape, structure, or motion patterns. This process reflects human cognition, where observed sensory input must be transformed into structured, generalized "knowledge" representations. For example, humans do not merely react to pixel-level changes; they interpret and internalize abstract concepts from raw visual data, such as "wings," "turrets," or "ship outlines," and associate them with learned categories. Similarly, in deep models, meaningful feature extraction enables the learning system to bridge the perceptual gap between raw measurements and semantic understanding. Therefore, in the SCM framework, the recognition process is modeled through the indirect path X→F→Y, where F captures the attributes of the basic, semantic information required for classification. Thus, there is no direct causal relationship between X and Y, reiterating that effective recognition depends on intermediate feature abstraction.

[0082] LINK 2 X←B→Y: This link illustrates how background bias can act as a confounding factor between input samples and predicted labels in recognition tasks. Ideally, extracted features should only represent inherent properties of the target object. Because specific backgrounds frequently coexist with specific target categories in the training data, the model inadvertently learns associations between background elements and category labels. These background-related features have no causal relationship with the target itself, but they influence the recognition process by introducing misleading associations. Therefore, the model may rely on irrelevant contextual cues, leading to shortcut learning. For example, even if the background has no causal influence on the target category, the learned associations may cause the model to incorrectly associate background patterns with specific categories. This confounding effect ultimately impairs the model's robustness, especially in the face of changing backgrounds and generalizing to out-of-distribution scenarios.

[0083] LINK 3 X←U→Y: This link reflects the confounding effects of the unobserved variable U, which simultaneously influences both the input X and the label Y. In recognition tasks, U may represent factors that are difficult to observe or quantify directly, such as weather conditions or sensor noise variables. These factors alter the appearance of the input and also affect the labeling process, thus creating spurious correlations between X and Y. Since the model cannot explicitly capture U, its effects cannot be directly controlled or corrected by traditional training strategies. Therefore, models trained under the influence of U are prone to overfitting to dataset-specific noise and exhibit poor generalization, especially when deployed in new environments where the distribution of U is shifted.

[0084] In the constructed causal relationship diagram, two types of confounding factors exist along the causal path from X to Y. The first type involves background bias, which can be observed relatively accurately. According to intervention theory, this confounding can be addressed using the method of shell adjustment. The second type includes factors that are difficult to observe, such as environmental disturbances, where appropriate mediating variables must be identified to reveal the causal effects along the path. Here, we slightly misuse the notation doX to represent doX=x:

[0085]

[0086]

[0087] By combining the intervention strategies described above, we can ultimately obtain an unbiased estimate of the causal effect of X on Y under changing context. This makes robust learning based on context possible.

[0088] This method explicitly decomposes traditionally difficult-to-distinguish complex sources of interference into "observable background (B)" and "unobservable confounding variables (U)," corresponding to the classic causal problems of "backdoor path" and "frontdoor path," respectively. This distinction is not empirical but based on a rigorous causal reasoning framework, thus giving subsequent intervention strategies (backdoor adjustment and frontdoor adjustment) clear theoretical relevance and explanatory power. It fundamentally avoids conflating biases of different natures, providing a clear roadmap for achieving precise, interpretable, and robust improvements.

[0089] Specifically, causal path represents the path by which SAR images influence prediction results through feature extraction;

[0090] The backdoor path represents the path through which the observable background simultaneously influences image formation and category prediction;

[0091] The front door path represents the path where unobservable confounding variables simultaneously affect both the input and the label.

[0092] This application provides a robust SAR target identification method based on causal reasoning under changing background conditions. This method deconstructs the SAR target identification process into three core paths through a causal model: first, the "causal path," where the SAR image directly influences the model's prediction through its own target features—this is the correct association the model should learn; second, the "backdoor path," where the observable background not only affects the formation of the SAR image but may also be incorrectly used by the model for category prediction, thus introducing spurious correlations; and third, the "frontdoor path," where unobservable confounding variables simultaneously affect both the SAR image and the ground truth label, constituting a hidden confusion mechanism. The core of this method lies in identifying and systematically blocking these two non-causal interference paths—the backdoor and the frontdoor—thereby ensuring that the model makes robust judgments based solely on the true causal path (target features).

[0093] The causal path is X→F→Y; the back door path is X←B→Y; and the front door path is X←U→Y.

[0094] This method does not simply attribute recognition errors to "complex backgrounds," but rather decomposes the problem causally: for "backdoor paths," it cuts off the influence of the background as a confounding factor through backdoor adjustments; for "frontdoor paths," it separates the influence of confounding variables through frontdoor adjustments. This path-analysis-based intervention strategy allows the model to systematically isolate biases from different sources, rather than relying on empirical methods such as data augmentation for overall optimization. This significantly improves the method's principle and generalization ability, enabling it to maintain stable and reliable recognition performance even when facing new, unknown backgrounds or confounding scenarios.

[0095] Specifically, for the observable background, a backdoor adjustment strategy is designed to extract background samples from the SAR image, including:

[0096] Calculate the channel average intensity map of the SAR image;

[0097] A preset quantile threshold is used to calculate the intensity threshold based on the quantile threshold and the channel average intensity map.

[0098] A binary mask is generated based on the intensity threshold, and background samples of the SAR image are extracted based on the binary mask.

[0099] This application provides a robust SAR target identification method based on causal reasoning under changing background conditions. This method employs a backdoor adjustment strategy for the observable background, with its core being the automatic and quantitative separation of the target from the background. Specifically, it first calculates the channel average intensity map of the SAR image to characterize the energy distribution of each region. Then, by pre-setting a quantile threshold (e.g., selecting a lower percentile of intensity), this statistical value is established as the intensity threshold for distinguishing between the background and the target. Finally, a binary mask is generated based on this threshold to accurately identify the background region in the image, thereby extracting the clean background sample. This process achieves automated and reproducible removal of background interference information from the original SAR image.

[0100] For the input image, calculate its channel average intensity map using the following formula:

[0101]

[0102] in, This represents the channel average intensity map. Indicates the number of channels. This represents the input image.

[0103] The intensity threshold is determined based on a preset quantile threshold s (e.g., 90%), calculated using the following formula:

[0104]

[0105] in, Indicates the intensity threshold. Indicates the quantile threshold. .

[0106] The formula for generating a binary mask is as follows:

[0107]

[0108] in, Represents a binary mask. This represents the two-dimensional value of the input image.

[0109] The formula for extracting background samples is as follows:

[0110]

[0111] in, This represents the background sample.

[0112] The advantages of this method lie in the objectivity, adaptability, and interpretability of its separation strategy. Compared to relying on subjective annotation or fixed thresholds, it utilizes the global intensity statistics (quantiles) of the image itself to determine the segmentation threshold, adapting to changes in intensity contrast across different scenes and targets. This data-driven quantitative method reduces human bias, ensures consistency in background extraction, and provides a reliable and clean source of background information for subsequent weighted fusion to generate intervention samples. It is a key foundation for effectively implementing backdoor adjustments and cutting off spurious correlations.

[0113] Specifically, the background samples and SAR images are weighted and fused to generate the first intervention sample, including:

[0114] The fusion weights are obtained by fusing SAR images and background samples using an exponential weighting method.

[0115] The first intervention sample is generated based on the fusion weights.

[0116] This application provides a robust SAR target recognition method based on causal inference under changing background conditions. Instead of simply replacing the extracted background samples, this method uses an exponential weighting method to fuse the original SAR image with the extracted background samples. By setting different weighting coefficients, this method can systematically generate a series of images with varying fusion strengths, i.e., "first intervention samples." This process is equivalent to controlling the adjustment of the background components of the input image in causal intervention, thereby simulating various situations under changing background conditions and providing rich and controllable training data for the model to learn how to eliminate background interference.

[0117] K-1 background samples are randomly selected from the dataset to construct a background noise set, calculated using the following formula:

[0118]

[0119] in, This represents a background clutter set.

[0120] The original sample and background sample are fused using an exponential weighting method, calculated as follows:

[0121]

[0122] in, This represents the fusion weight of the k-th sample. Indicates the attenuation parameter. This represents the total number of samples fused. This indicates the sample index.

[0123] The first intervention sample is generated based on the fusion weights, and the calculation formula is as follows:

[0124]

[0125] in, This represents the first intervention sample generated. This represents the original input sample. This represents the k-th mixed background sample.

[0126] The advantages of this method lie in its gradual, controllable, and causal effectiveness. The exponential weighting method allows for a smooth and continuous adjustment of the background's intensity, avoiding training instability or image distortion caused by sudden, complete background replacement. This gradual fusion generates a complete data spectrum from the "original background" to the "completely new background," enabling the model to gradually learn to detach from its dependence on specific backgrounds during training. This allows for a more robust grasp of the target's essential characteristics, ultimately effectively cutting off spurious correlations caused by backdoor paths.

[0127] Specifically, for unobservable confounding variables, a front-door adjustment strategy is designed to construct a second intervention sample. The total causal effect is then calculated based on the second intervention sample and the SAR image, including:

[0128] The mean of the SAR images is calculated to construct a second intervention sample;

[0129] The outputs of the classification network, which uses SAR images and second intervention samples as inputs, are calculated separately to obtain the first output and the second output.

[0130] The total causal effect is calculated based on the first and second output results.

[0131] This application provides a robust SAR target identification method based on causal inference under changing background conditions. This method employs a mean-based intervention-based front-door adjustment strategy to address unobservable confounding variables. Its core operation involves calculating the average value of all SAR images to construct an "averaged" second intervention sample. Subsequently, the original SAR image and this averaged image are input into the current classification network, yielding two outputs. By calculating the difference between these two outputs, the total causal effect predicted by the model after excluding specific image content (individual features) can be estimated. This effect largely reflects the overall impact of unobservable confounding variables.

[0132] The second intervention sample is constructed using the mean of all samples in the current training batch as the intervention input. The calculation formula is as follows:

[0133]

[0134] in, This represents the second intervention sample. Indicates the number of samples.

[0135] The model outputs for the original samples and the intervention samples are calculated separately using the following formulas:

[0136] ;

[0137] ;

[0138] in, This indicates the first output result. This indicates the second output result.

[0139] The formula for calculating the total causal effect is:

[0140]

[0141] in, Indicates the total causal effect. This indicates the progressive intervention weight, which increases with the number of training rounds.

[0142] The key advantage of this method lies in its ingenious approximation and feasibility. Traditional methods struggle to handle unobservable confounding variables, which cannot be directly measured or controlled. This method creatively utilizes the "image mean" statistic as an intervention tool. Its core insight lies in the fact that the mean image erases all target specificity and can be considered a standardization intervention on the input, thus approximately isolating the systematic biases caused by the confounding variables. This makes front-door adjustment, which is theoretically complex and difficult to calculate directly, operational, providing an effective computational proxy for identifying and gradually eliminating implicit, systematic confounding effects during training.

[0143] Specifically, the progressive intervention strategy uses the Sigmiod function to control the intensity of intervention by unobservable confounding variables on the SAR target identification process.

[0144] This application provides a robust SAR target recognition method based on causal inference under changing background conditions. After calculating the total causal effect caused by confounding variables, this method does not directly apply it all to network updates. Instead, it exerts influence through a gradual intervention strategy controlled by a Sigmoid function. This strategy dynamically adjusts the intervention intensity according to the training progress (e.g., the current iteration number or epoch): in the early stages of training, the intervention intensity is low, allowing the model to learn basic visual features and classification capabilities; as training progresses, the Sigmoid function smoothly and non-linearly increases the intervention intensity, gradually forcing the model to confront and overcome the systematic bias introduced by confounding variables, ultimately achieving robust recognition.

[0145] The Sigmoid function is used to control the intensity of the intervention; the calculation formula is as follows:

[0146]

[0147] in, Indicates the weight of gradual intervention. Indicates the current training round number. Indicates the total number of training rounds. Indicates the transition speed control parameter. This represents the Sigmoid activation function.

[0148] The advantage of this method lies in its balance between training stability and optimization. Directly applying full-intensity causal intervention can severely interfere with the model's basic learning in the early stages of training, even leading to non-convergence. However, by using the Sigmoid function for gradual adjustment, the two objectives of "learning features" and "correcting biases" are cleverly decoupled and balanced over time. It mimics the human learning process of "mastering the basics first, then refining and correcting," ensuring a smooth transition in training. This allows the model to robustly remove the influence of confounding factors without destroying already acquired effective knowledge, ultimately achieving more robust and generalized performance convergence.

[0149] Specifically, a classification network is built based on CNN, including:

[0150] A feature extraction layer for a classification network based on CNN;

[0151] It facilitates the construction of the output layer of a classification network using fully connected layers and the Softmax function;

[0152] Cross-entropy loss is used as the loss function for the classification network.

[0153] This application provides a robust SAR target recognition method based on causal reasoning under changing background conditions. This method employs a classic deep neural network architecture as the main framework of its classifier. Specifically, it leverages the powerful spatial feature extraction capabilities of convolutional neural networks (CNNs) to construct feature extraction layers, automatically learning multi-level discriminative features from SAR images. Subsequently, these high-level features are mapped to the target category space through fully connected layers and normalized using the Softmax function, outputting the predicted probability for each category. The entire network is trained using the cross-entropy loss function as the optimization objective, aiming to minimize the difference between the model's predicted distribution and the true label distribution, thereby driving the learning and updating of network parameters.

[0154] The advantages of this method lie in its mature, efficient, and adaptable architecture. Employing a CNN as the backbone effectively processes the texture and structural information of SAR images, automatically learning robust features suitable for target recognition, and avoiding tedious manual feature engineering. The combination of fully connected layers and the output structure of Softmax is a standard and efficient configuration for image classification tasks. Simultaneously, using cross-entropy loss as the optimization criterion provides the model with clear and differentiable gradient signals, ensuring stable and efficient convergence during training. This classic and reliable architectural choice allows the method to focus all its innovative efforts on causal intervention strategies without being distracted by the design of the base classifier, ensuring the overall simplicity, effectiveness, and reproducibility of the method.

[0155] Specifically, the trained classification network was tested using the EOC-Scene dataset to verify its effectiveness, which was expressed as target recognition accuracy, top-five accuracy, recall, and F1 score.

[0156] This application provides a robust SAR target recognition method based on causal inference under changing backgrounds. After model training, this method uses the publicly available and standard EOC-Scene dataset as a benchmark. It does not rely solely on overall accuracy but constructs a multi-dimensional comprehensive evaluation system: in addition to basic target recognition accuracy, it introduces the top five accuracies to evaluate the model's ranking ability on ambiguous or difficult samples; simultaneously, it calculates recall to examine the model's completeness in detecting various types of targets; and it combines precision and recall to calculate the F1 score to balance the model's overall performance on both positive and negative samples. These metrics collectively quantify the final recognition performance of the trained classification network in complex scenarios with changing backgrounds.

[0157] The advantages of this method lie in its comprehensiveness, rigor, and fair comparability in evaluation. It avoids model defects that may be masked by relying solely on a single accuracy rate, providing a panoramic performance profile from different perspectives through multiple indicators. More importantly, the use of the publicly available EOC-Scene dataset for testing ensures the reproducibility of the experiments and the horizontal comparability of the results. This allows for an objective and fair comparison of the robustness improvements of this method with other state-of-the-art work in the field, greatly enhancing the persuasiveness and credibility of the research findings.

[0158] Figure 3 A schematic diagram of the feature unwrapping network for a robust SAR target identification method based on causal reasoning under changing background conditions, as provided in this application, is shown below. Figure 3 The diagram shown is a schematic diagram of the feature unwrapping network provided in this embodiment. The schematic diagram of the feature unwrapping network includes: sampling different images to obtain different samples, sending the different samples to a parameter-shared encoding network to obtain different features, and sending the obtained features to a target recognition subnetwork and a target transformation subnetwork to obtain different category results, different reconstructed samples, and a transformed sample.

[0159] Figure 4 A schematic diagram of the target recognition subnetwork structure for a SAR target robustness recognition method based on causal reasoning under changing background conditions, as provided in this application, is shown below. Figure 4 The diagram shown is a schematic diagram of the target recognition sub-network structure provided in this embodiment. The target recognition sub-network structure includes: inputting different features of the same image into a convolutional layer, the convolutional layer sharing parameters, and processing different categories through the SoftMax function.

[0160] Figure 5 A schematic diagram of the target transformation subnetwork structure of a SAR target robustness recognition method based on causal reasoning under changing background conditions is provided in this application, as shown below. Figure 5 The diagram shown is a schematic diagram of the target transformation sub-network structure provided in this embodiment. The schematic diagram of the target transformation sub-network structure includes: sending multiple features of different images into a parameter-shared decoding network to obtain multiple reconstructed samples and one transformation sample.

[0161] To verify the effectiveness and applicability of the proposed SAR target recognition method based on causal intervention and robust to background changes under complex background conditions, the following embodiments are used. In this embodiment, the EOC-Scene dataset is selected as the experimental dataset. This dataset is used to simulate the distribution shift problem caused by changes in the background environment in SAR target recognition tasks. By introducing imaging backgrounds of different complexities while maintaining consistency in sensor parameters and target categories, the generalization ability of the target recognition model under background change conditions is verified. The dataset contains 19,584 training samples collected from scenes with relatively simple background structures, including sandstone surfaces and bare soil scenes; the test set contains 77,791 samples collected from scenes with complex background structures, including urban scenes, factory scenes, and woodland scenes. Complex backgrounds contain factors such as overlapping buildings, vegetation obstruction, and dense canopies, which significantly change the scattering characteristics of the target. The dataset contains 40 vehicle target categories. The target categories in the training and test sets are completely consistent, and the imaging conditions are also consistent, using X / Ku bands and a four-polarization scheme, with a spatial resolution of approximately 0.12 to 0.15 meters. This ensures that variations in recognition performance are mainly caused by background differences. To compare and verify the effectiveness of the proposed method, several existing target recognition methods are introduced as comparison schemes in the experiments. The comparison methods include general image classification networks and dedicated network models designed for SAR target recognition tasks. General image classification models include VGG16, ResNet50, Xception, Inception-V3, EfficientNet, ConvNeXt, and FastViT; models designed for SAR target recognition tasks include AConvNet, MANet, FENNet, ESENet, and HDANet. These multiple comparison methods are used to verify the applicability of the proposed method under different network structure conditions. During model training, the network parameters of all comparison methods and the proposed method are initialized using default settings, without using pre-trained models. The training process uniformly employs a momentum-based stochastic gradient descent algorithm for optimization. The initial learning rate for the general image classification model is set to 0.1, while the initial learning rate for the SAR-specific model is set to 0.05, the momentum parameter is set to 0.9, and the batch size is set to 128. Each model is trained for 100 training epochs, and the model with the best recognition performance on the validation set is selected for the testing phase. The key hyperparameters involved in this method are set as follows: the number of feature fusion iterations is set to 4, the intensity quantile threshold is set to 90, and the sigmoid control parameter is set to 0.3. These parameters were determined through experimental experience to balance background suppression and target discrimination capabilities. For performance evaluation, multiple evaluation metrics are used to quantitatively analyze the recognition performance of each method, including target recognition accuracy, top-five accuracy, recall, and F1 score.The above experimental setup and comparative analysis are used to verify the effectiveness and stability of the proposed method for SAR target identification under changing background conditions.

[0162] Table 1. Basic information about the dataset used in the validation process.

[0163]

[0164] To verify the effectiveness of the method, we demonstrate the image reconstructed by the decoding network. (Sample) Identity features extracted from the network and rotational features The reconstructed sample is obtained after decoding through the network. The same sample The reconstructed sample can be obtained. Rotation features of two samples , The input yields parameters representing the angular difference. And thus complete the rotation feature. To rotation feature The transformation yields the rotational features. Identity characteristics and rotational features The input is passed through a decoding network to obtain the transformed sample. .

[0165] Figure 6 The image shows the generation effect of the reconstruction task and the transformation task in the target transformation subnetwork of a SAR target robust recognition method based on causal reasoning under background changes, as provided in this application. Figure 6 As shown, the SAR images generated by the reconstruction and transformation tasks are displayed. Figure 6 The first two lines are the network input, i.e., the samples. , The third party uses identity characteristics and rotational features The reconstructed sample is obtained after decoding through the network. The fourth line is identity characteristics. and rotational features The input is passed through a decoding network to obtain the transformed sample. .from Figure 6 As can be seen from the sample Identity features extracted by the coding network and rotational features The decoding network can successfully complete the reconstruction task, demonstrating the effectiveness of both the encoding and decoding networks; simultaneously, based on the samples... Identity features extracted by the coding network and rotational features Excellent reconstruction was also achieved through a parameter-sharing decoding network, demonstrating that the target transformation network can handle rotational features. To rotation feature The transformation, i.e., rotation feature and rotational features They gradually become more consistent during the training process.

[0166] To verify the efficiency of our method, we compared it with several mainstream target recognition algorithms. For this application, in order to determine the parameters... We determined the parameter settings under the two experimental conditions through grid search and cross-validation. Figure 7 The optimal parameter search graph for a robust SAR target identification method based on causal reasoning under changing background conditions provided in this application is shown below. Figure 7 The diagram shows the optimal parameter search process. (a) illustrates the search process under five types of missing angle samples, and (b) shows the search process under nine types of missing angle samples. In the comparison algorithm selection, in addition to the classic Support Vector Machine (SVM) classifier and the Sparse Representations Classification (SRC) classifier, we also selected a high-performance deep convolutional neural network, AconvNet. Based on this, we used directly rotated and augmented samples as the training set to test the feasibility of directly rotating images to solve the angle missing problem. Furthermore, we combined the deep convolutional neural network with an STN module to obtain robust angle-invariant features. Compared to the above comparison algorithms, this application achieved the highest recognition accuracy in the azimuth angle missing sample set tests under both conditions. The specific recognition results are shown in Table 2, which fully demonstrates the efficiency of this application in target recognition under continuous azimuth angle missing conditions.

[0167] Table 2. Recognition accuracy of the present invention and the comparative method in two missing cases.

[0168]

[0169] Figure 8 A connection diagram of a robust SAR target identification system based on causal reasoning under changing background conditions is provided in this application, as shown in the figure. Figure 8 As shown in this embodiment, a robust SAR target identification system based on causal reasoning under changing background conditions is provided. The identification system includes:

[0170] The causal model building module is used to build a causal model in the SAR target recognition process. It defines multiple variables in the SAR target recognition process and integrates these multiple variables with the causal model. The multiple variables include at least: SAR image, observable background and unobservable confounding variables.

[0171] The backdoor adjustment strategy design module is connected to the causal model construction module. The backdoor adjustment strategy design module is used to design backdoor adjustment strategies for observable backgrounds. The backdoor adjustment strategy extracts background samples from SAR images and performs weighted fusion with SAR images to generate the first intervention sample. The backdoor adjustment strategy is used to eliminate false correlations between background and target categories in SAR images.

[0172] The front-door adjustment strategy design module is connected to the back-door adjustment strategy design module. The front-door adjustment strategy design module is used to design a front-door adjustment strategy for unobservable confounding variables, construct a second intervention sample through the front-door adjustment strategy, calculate the total causal effect based on the second intervention sample and SAR image, and use a progressive intervention strategy based on the total causal effect to eliminate the influence of unobservable confounding variables on the SAR target recognition process.

[0173] The classification network construction and training module is connected to the front door adjustment strategy design module. The classification network construction and training module is used to build a classification network based on CNN. The SAR image is input into the classification network to extract image features. The classification network is then intervened based on the first intervention sample and the second intervention sample. Backpropagation is used to update the parameters of the classification network to obtain the trained classification network.

[0174] The classification network testing and evaluation module is connected to the classification network construction and training module. The classification network testing and evaluation module is used to obtain the EOC-Scene dataset and use the EOC-Scene dataset to test the trained classification network to verify the effect of the trained classification network.

[0175] This application provides a robust SAR target recognition system based on causal reasoning under changing background conditions. The system consists of five functional modules forming a complete processing flow: First, the causal model construction module is responsible for building the core causal graph and identifying SAR images, observable background, and unobservable confounding variables; second, the backdoor adjustment strategy design module generates the first intervention sample through background extraction and weighted fusion to block spurious associations caused by the background; third, the frontdoor adjustment strategy design module weakens the influence of unobservable confounding factors by constructing mean intervention samples and progressive interventions; fourth, the classification network construction and training module integrates a CNN classifier and incorporates the intervention samples generated by the first two modules into the training, optimizing the network through backpropagation; finally, the classification network testing and evaluation module performs multi-index performance tests on the EOC-Scene dataset to verify the effectiveness of the overall method.

[0176] This system breaks down complex causal intervention techniques into logically clear and well-defined modules. Each module handles a specific sub-problem (modeling, background intervention, hybrid intervention, network training, performance evaluation), making the overall architecture highly organized and easy to understand and implement. This design not only ensures a complete closed loop from theoretical modeling to experimental verification but also facilitates future improvements or replacements of individual modules (e.g., using more advanced networks or intervention strategies) without affecting the overall framework. Clear connections between modules ensure the smooth flow of data and causal intervention information, ultimately achieving a robust identification process from problem diagnosis to solution.

Claims

1. A robust SAR target identification method based on causal reasoning under changing background conditions, characterized in that, The identification method includes: A causal model is constructed for the SAR target recognition process, and multiple variables in the SAR target recognition process are defined. The multiple variables are then fused with the causal model. The multiple variables include at least: SAR image, observable background and unobservable confounding variables. For the observable background, a backdoor adjustment strategy is designed. The backdoor adjustment strategy is used to extract background samples from the SAR image. The background samples are then weighted and fused with the SAR image to generate a first intervention sample. The backdoor adjustment strategy is used to eliminate spurious correlations between the background and target categories in the SAR image. For the unobservable confounding variables, a front-door adjustment strategy is designed. A second intervention sample is constructed through the front-door adjustment strategy. The total causal effect is calculated based on the second intervention sample and the SAR image. Based on the total causal effect, a progressive intervention strategy is used to eliminate the influence of the unobservable confounding variables on the SAR target recognition process. A classification network is constructed based on CNN. The SAR image is input into the classification network to extract image features. The classification network is then intervened based on the first intervention sample and the second intervention sample. Backpropagation is used to update the parameters of the classification network to obtain the trained classification network. Obtain the EOC-Scene dataset and use it to test the trained classification network to verify its effectiveness.

2. The robust SAR target identification method based on causal reasoning under changing background conditions as described in claim 1, characterized in that, The definition of multiple variables in the SAR target identification process, and the fusion of these variables with the causal model, includes: The SAR image is defined as variable X, the observable background as B, and the unobservable clutter variable as U; Establish causal relationships between various variables and the causal model, including causal paths, backdoor paths, and frontdoor paths.

3. The robust SAR target identification method based on causal reasoning under changing background conditions as described in claim 2, characterized in that, The causal path refers to the path by which the SAR image influences the prediction result through feature extraction; The backdoor path represents the path through which the observable background simultaneously affects image formation and category prediction; The front door path represents the path through which the unobservable confounding variable simultaneously affects both the input and the label.

4. The robust SAR target identification method based on causal reasoning under changing background conditions according to claim 1, characterized in that, The step involves designing a backdoor adjustment strategy for the observable background, and extracting background samples from the SAR image using this strategy, including: Calculate the channel average intensity map of the SAR image; A preset quantile threshold is used to calculate an intensity threshold based on the quantile threshold and the channel average intensity map. A binary mask is generated based on the intensity threshold, and the background sample of the SAR image is extracted based on the binary mask.

5. A robust SAR target identification method based on causal reasoning under changing background conditions, as described in claim 1, is characterized in that... The step of weighted fusion of the background sample and the SAR image to generate the first intervention sample includes: The SAR image and the background sample are fused using an exponential weighting method to obtain the fusion weight; The first intervention sample is generated based on the fusion weights.

6. The robust SAR target identification method based on causal reasoning under changing background conditions according to claim 1, characterized in that, The step involves designing a front-door adjustment strategy for the unobservable confounding variables, constructing a second intervention sample using the front-door adjustment strategy, and calculating the total causal effect based on the second intervention sample and the SAR image, including: The mean of the SAR images is calculated to construct the second intervention sample; The output results of the classification network, which uses the SAR image and the second intervention sample as inputs, are calculated respectively to obtain the first output result and the second output result; The total causal effect is calculated based on the first output result and the second output result.

7. A robust SAR target identification method based on causal reasoning under changing background conditions, as described in claim 1, is characterized in that... The progressive intervention strategy utilizes the Sigmiod function to control the intensity of the intervention of the unobservable confounding variables on the SAR target identification process.

8. A robust SAR target identification method based on causal reasoning under changing background conditions, as described in claim 1, is characterized in that... The classification network built based on CNN includes: The feature extraction layer of the classification network is constructed based on the CNN; The fully connected layer and the Softmax function are used to construct the output layer of the classification network; Cross-entropy loss is used as the loss function for the classification network.

9. A robust SAR target identification method based on causal reasoning under changing background conditions, as described in claim 1, is characterized in that... The effectiveness of the trained classification network is verified by testing it using the EOC-Scene dataset, which is represented by target recognition accuracy, top-five accuracy, recall, and F1 score.

10. A robust SAR target identification system based on causal reasoning under changing background conditions, characterized in that, The identification system is applied to the identification method according to any one of claims 1 to 9, and the identification system comprises: A causal model construction module is used to construct the causal model in the SAR target identification process, define multiple variables in the SAR target identification process, and fuse the multiple variables with the causal model. The multiple variables include at least: the SAR image, the observable background, and the unobservable confounding variables. A backdoor adjustment strategy design module is connected to the causal model construction module. The backdoor adjustment strategy design module is used to design the backdoor adjustment strategy for the observable background, extract the background sample of the SAR image through the backdoor adjustment strategy, and perform weighted fusion of the background sample and the SAR image to generate the first intervention sample. The backdoor adjustment strategy is used to eliminate the false correlation between the background and the target category in the SAR image. A front-door adjustment strategy design module is connected to the back-door adjustment strategy design module. The front-door adjustment strategy design module is used to design the front-door adjustment strategy for the unobservable confounding variables, construct the second intervention sample through the front-door adjustment strategy, calculate the total causal effect based on the second intervention sample and the SAR image, and use the progressive intervention strategy based on the total causal effect to eliminate the influence of the unobservable confounding variables on the SAR target recognition process. The classification network construction and training module is connected to the front door adjustment strategy design module. The classification network construction and training module is used to construct the classification network based on the CNN, input the SAR image into the classification network to extract the image features, intervene in the classification network based on the first intervention sample and the second intervention sample, and update the parameters of the classification network using the backpropagation to obtain the trained classification network. The classification network testing and evaluation module is connected to the classification network construction and training module. The classification network testing and evaluation module is used to obtain the EOC-Scene dataset and use the EOC-Scene dataset to test the trained classification network to verify the effect of the trained classification network.

Images