Few-Shot Class-Incremental SAR Target Recognition with a Forward-Compatible Prototype Classifier

Highlights

What are the main findings?

• We propose a Forward-Compatible Prototype Classifier (FCPC) by emphasizing the model’s forward compatibility to continually learn new concepts from limited samples without forgetting the previously learned ones.
• A Nearest-Class-Mean (NCM) classifier is proposed for prediction by comparing the semantics of unknown targets with prototypes of all classes based on the cosine criterion.

What are the implications of the main findings?

• The proposed method can continually learn new concepts from limited samples without forgetting the previously learned ones, which can improve the SAR ATR capability.
• The proposed method powers the DL-based SAR ATR systems with Few-Shot Class-Incremental Learning (FSCIL) ability to satisfy real-world SAR ATR scenarios.

Abstract

In practical Synthetic Aperture Radar (SAR) applications, new-class objects can appear at any time as the rapid accumulation of large-scale and high-quantity SAR imagery and are usually supported by limited instances in most cooperative scenarios. Hence, powering advanced deep-learning (DL)-based SAR Automatic Target Recognition (SAR ATR) systems with the ability to continuously learn new concepts from few-shot samples without forgetting the old ones is important. In this paper, we tackle the Few-Shot Class-Incremental Learning (FSCIL) problem in the SAR ATR field and propose a Forward-Compatible Prototype Classifier (FCPC) by emphasizing the model’s forward compatibility to incoming targets before and after deployment. Specifically, the classifier’s sensitivity to diversified cues of emerging targets is improved in advance by a Virtual-class Semantic Synthesizer (VSS), considering the class-agnostic scattering parts of targets in SAR imagery and semantic patterns of the DL paradigm. After deploying the classifier in dynamic worlds, since novel target patterns from few-shot samples are highly biased and unstable, the model’s representability to general patterns and its adaptability to class-discriminative ones are balanced by a Decoupled Margin Adaptation (DMA) strategy, in which only the model’s high-level semantic parameters are timely tuned by improving the similarity of few-shot boundary samples to class prototypes and the dissimilarity to interclass ones. For inference, a Nearest-Class-Mean (NCM) classifier is adopted for prediction by comparing the semantics of unknown targets with prototypes of all classes based on the cosine criterion. In experiments, contributions of the proposed modules are verified by ablation studies, and our method achieves considerable performance on three FSCIL of SAR ATR datasets, i.e., SAR-AIRcraft-FSCIL, MSTAR-FSCIL, and FUSAR-FSCIL, compared with numerous benchmarks, demonstrating its superiority and effectiveness in dealing with the FSCIL of SAR ATR.

1. Introduction

Synthetic Aperture Radar (SAR) is a landmark achievement in the field of Earth remote sensing science. Thanks to its capability for observing targets in all day and weather conditions, SAR images have been important data sources for various applications in both civil and military fields, like hazard warning [1], change detection [2], and target surveillance [3,4,5]. As a fundamental task of SAR, SAR Automatic Target Recognition (SAR ATR), which aims to identify potential targets in SAR images, can provide diverse and complementary information for decision-making and has been a longstanding and intensive research topic in the past several decades. When entering the big data era, deep-learning (DL) approaches have shown superior performance to their precursors, igniting a new revolution in various SAR ATR tasks [6,7]. After years of innovation, numerous airborne and spaceborne high-resolution SAR imaging systems have been manufactured and are in service, making the collection of targets of interest from SAR images more convenient and timely, which provides more opportunities for researchers to explore solutions to realistic problems. In particular, given the ever-changing worlds humans and SAR ATR systems face, more efforts should be devoted to pushing them forward to satisfy practical demands.

Traditionally, most current SAR ATR systems are usually designed for ideal recognition tasks, where training data is collected in advance from a closed scene and assumed to follow the nearly i.i.d distributions with testing samples. However, real-world SAR ATR scenarios are non-stationary, where new classes can appear at any time with the acquisition of a large quantity of large-scale SAR imagery. A static SAR ATR system that is not adaptive nor robust to changes can quickly become outdated or unreliable. Therefore, powering a DL-based SAR ATR system with the ability to continually learn new concepts without forgetting the previous ones is important. To address this deficiency, incremental learning (IL), especially Class-IL (CIL), which aims at learning new classes in a stream format, has been borrowed from the computer vision (CV) areas and actively explored in the SAR ATR field in recent works. Certainly, these CIL-powered SAR ATR solutions have shown supervising performance in learning new concepts at non-stationary situations compared with traditional ones. However, standard CIL still assumes that there is sufficiently labeled data for new classes, which inevitably impedes its practical applications since it is expensive and even impossible to collect a large amount of annotated data in most SAR ATR scenarios, like boundary management and battlefield surveillance. Henceforth, powering SAR ATR systems with the ability to continuously learn new classes from limited samples is more practical and worth investigating.

In response, Few-Shot Class-Incremental Learning (FSCIL) has been explored and has attracted much attention from the CV communities. The goal of the FSCIL is to power the classifiers with the ability to continually learn new classes from few-shot samples without losing much performance in judging the old ones. As a specific setting of the CIL, two intrinsic issues, including Catastrophic forgetting [8] and Overfitting, should be addressed carefully. The former refers to the fact that a DL-based classifier learning on the new targets can easily encounter an irreversible performance drop when judging old ones, resulting from the inaccessibility to previous data due to data privacy or storage limitations. The latter, however, suggests that a classifier trained by new-class limited samples tends to capture instance-specific but not general patterns of new classes, due to the data scarcity of novel classes in most non-cooperative scenarios, leading to a substantial performance gap in identifying training and testing samples.

Recently, numerous solutions to the problem have been proposed, and significant performance has been achieved in CV fields. Among them, methods following the forward-compatible paradigm, which aims at improving the model’s adaptability to unknown categories in advance, have exhibited significant performance. Typically, FACT [9] pre-allocated and learned numerous virtual prototypes, which were synthesized from real-class semantics features, to sequence the embedding of known classes and reserve them for the new ones. ALICE [10] solved the problem from the open-set perspective by introducing auxiliary classes via mixing pairs of real-class samples. In terms of the FSCIL in the SAR ATR field, DSSC [11] tentatively solved the problem from the forward-compatible aspect by designing two self-supervised learning (SSL) tasks, which include the Scattering Mixup Module (SMM) and the Rotation-aware Transformation Module (RTM), for virtual-class generation while decoupling the model’s parameters with dynamic worlds for rapid knowledge transferring. Despite the tremendous success, all virtual classes in the DSSC were synthesized based on real-class pixel-level scattering information and can be easily interfered with by background clutter. Also, the decoupling operation implicitly weakens the classifier’s continual adaptability to new concepts, leading to a gradual feature misalignment for representing both old and current targets. Furthermore, due to the unique imaging mechanisms of SAR and the complexity and diversity of target structures and postures, their signatures can exhibit significant intra-class variability and inter-class similarity, leading to severe interference for accurate target recognition. Henceforth, domain knowledge of targets in SAR imagery should be carefully considered for stable inference.

In this paper, a novel Forward-Compatible Prototype Classifier (FCPC) is proposed to improve the model’s forward-compatibility to new targets before and after deployment. In the FCPC, compared with scattering mixup, a Virtual-class Semantic Synthesizer (VSS) is designed for virtual-class generation by using base-class semantics. As a result, the model’s representation ability for diverse features of unknown targets can be further improved in advance. After deployment, a Decoupled Margin Adaptation (DMA) strategy is designed to merely update the model’s high-level semantic parameters by improving the consistency of new-class marginal features with respect to class prototypes while enlarging the difference to inter-class ones. For inference, a Nearest-Class-Mean (NCM) classifier is employed for prediction by comparing the cosine distances between features of testing samples and all-class prototypes. Extensive experiments on three FSCIL of SAR ATR datasets, i.e., MSTAR-FSCIL, SAR-AIRcraft-FSCIL, and FUSAR-FSCIL, demonstrate the superiority of our method compared with numerous task-specific benchmarks. Overall, our contributions are three-fold.

• We explored the FSCIL problem in the SAR ATR field and designed an FCPC framework to further improve the model’s forward compatibility with unpredictable targets in the dynamic world before and after deployment.
• We designed a VSS for virtual-class synthesis based on real-class semantics, a DMA for making our method continually evolve, and an NCM classifier for general prediction.
• We showed the state-of-the-art performance of the FCPC in comparison to several advanced benchmarks on three derived FSCIL of SAR ATR datasets.

The rest of the paper is organized as follows. The related works are presented in Section 2. The problem set-up, the motivations, and our method are given in Section 3. The experimental settings, results, and analysis are detailed in Section 4. Section 5 discusses the potential limitations of the proposed model and its future research directions. Section 6 concludes the work.

2. Related Works

2.1. SAR Target Recognition

The methods of SAR target recognition can be broadly categorized into template-based, model-based, and machine-learning-based paradigms [6]. In template-based solutions [12,13,14], unknown targets are classified by matching them to pre-defined templates based on low-level, hand-crafted features such as gray scale, length, edge, and region moment; it is simple and intuitive yet easily influenced by the environment. In model-based approaches [15,16,17], parameterized models with rigorously physical priors and hypotheses are formulated to estimate target electromagnetic scattering structures for similarity comparison. The models have a strong explainability yet are very complicated, caused by complex SAR imagery mechanisms. Recently, machine-learning-based, especially DL-based, solutions [6,18,19,20,21,22,23] have dominated the field, benefiting from their powerful feature-representing and -discriminating abilities. However, most are data-hungry and merely designed for an ideally closed-world recognition scenario, which is incompatible with reality.

Our method still follows the DL-based paradigm and aims to continuously learn new knowledge from scarce samples without seriously forgetting the old ones.

2.2. Few-Shot Class-Incremental Learning

Few-Shot Class-Incremental Learning (FSCIL) [24] as an extension of the CIL aims at teaching a classifier to continually learn new concepts from few-shot samples without forgetting the old ones and has been an active exploration area in recent years. Typically, TOPIC [24] first defined the problem and preserved feature manifolds of old classes by a neural gas network. ERDIL [25] selected old-class exemplars using an exemplar relation graph (ERG) and distilled relational features for knowledge preservation. IDLVQC [26] represented class knowledge as quantized reference vectors and continually adjusted locations of the old vectors for misalignment reduction. Although their effectiveness has been verified, most of them focus on mitigating the forgetting of the old knowledge, i.e., backward compatibility, but overlook the pre-perceiving ability to the new knowledge, weakening its adaptability to dynamic worlds. In contrast, methods with forward compatibility mainly focus on preparing the base stage training to facilitate the better acceptance of novel classes in the future. Typically, CEC [27] proposed a continually evolving classifier leveraged by a meta-learned graph attention network for new knowledge adaptation. FACT [9] pre-allocated numerous orthogonal prototypes as virtual classes to squeeze the embedding of known targets and reserve for the new. ALICE  [10] solved the problem from an open-set perspective by generating and pre-assigning several virtual classes generated by mixing samples of different classes.

In this paper, our method explores the FSCIL in the SAR ATR field following the forward-compatible scheme for better knowledge incorporation and identification.

2.3. FSCIL of SAR ATR

Recently, some works have been carried out on the FSCIL in the SAR ATR field. Among them, Wang et al. proposed a HEIEN network [28] with an adaptive class-incremental learning (ACIL) module. Due to the tight connection between the cosine criterion and target azimuth-aware structures, the CPL framework [29] designed a couple of losses and a Nearest-Class-Mean (NCM) classifier to learn and identify targets at cosine space. Furthermore, the AASC framework [30] proposed losses at both semantic and manifold facets and designed a subspace classifier on the Grassmannian manifold for prediction. ODF [31] proposed an orthogonal distribution optimization method for the FSCIL of SAR ATR. DILHyFS  [32] proposed a dual-branch architecture that focuses on local feature extraction and leverages the discrete Fourier transform and global filters to capture long-term spatial dependencies. The effectiveness of the method was verified on numerous settings derived from the MSTAR dataset. Unlike the above approaches focusing on the model’s backward compatibility, in which the performance on learned targets is highly considered by stabilizing the drift of feature spaces passively, DSSC  [11] and DSAC  [33] addressed the problem from the forward-compatible angle by pre-allocating and learning several virtual classes in advance. Nevertheless, these classes, generated by the scattering mix-up, can easily interferred with by surroundings, making class supervisions inaccurate and unstable. However, in our method, not only are the semantics of base classes adopted for virtual-class generation but the classifier is also decoupled and evolved after deployment for better knowledge retention and transfer.

3. Materials and Methods

3.1. Problem Statement

Few-Shot Class-Incremental Learning (FSCIL) [24] necessitates a method to learn continually from scarcely labeled samples, which usually contains a base and numerous incremental steps, a.k.a. sessions, to mimic a practical recognition scenario. At the base session, classes with sufficient training samples are provided. At incremental sessions, new-class few-shot samples continually appear to serve as novel knowledge collected in dynamic worlds. An ideal FSCIL model should perform well on all seen categories during evaluation once optimized on the current data. Formally, assuming that we have a sequence of labeled data sets ${\mathcal{D}}^1,{\mathcal{D}}^2,\dots$, ${\mathcal{D}}^t$, where ${\mathcal{D}}^t={\left\{(x_i^t,y_i^t)\right\}}_{i=1}^N$. $x_i^t$ and $y_i^t$ refer to a training sample and its category, respectively. ${\mathcal{C}}^t$ is the class set of ${\mathcal{D}}^t$. $|{\mathcal{C}}^t|$ is the number of categories of ${\mathcal{C}}^t$. Notably, $\forall j,k\in t,j\ne k$, ${\mathcal{C}}^j\cap {\mathcal{C}}^k=⌀$. The base session dataset ${\mathcal{D}}^1$ usually contains large-scale training samples for each class. For each incremental session $t,t>1$, ${\mathcal{D}}^t$ contains several new classes with few-shot examples and are provided in an N-way K-shot format. At session t, only ${\mathcal{D}}^t$ is accessed for optimization, and the model is expected to classify all seen categories $\mathcal{C}={\cup }_{i=1}^t{\mathcal{C}}^i$ during evaluation.

3.2. Motivations

As the intrinsic obstacles of the FSCIL and characteristics of targets in SAR imagery, the Forward-Compatible and Stable-Discriminating capabilities should be considered for mitigating the FSCIL in the SAR ATR field.

• Forward-Compatible ability means that a DL-based SAR ATR system can not only represent targets of known categories but can also incorporate new concepts rapidly once optimization has taken place. According to the Attribute Scattering Center (ASC) theory, target information in SAR imagery can be regarded as the specific composition of various basic scattering parts, e.g., dihedral, trihedral, cylinder. Furthermore, target deep semantics can be seen as collections of various convolutional activations, which implicitly represent high-level class-agnostic structures. Henceforth, the SAR ATR system with a strong forward compatibility should possess the ability to capture diverse cues of incoming targets.
• Stable-Discriminating ability means that an SAR ATR system can accurately identify targets of different classes captured in diverse scenarios and postures. In contrast to rich and stable target cues in optical images, only target-specific structures and parts can be observed in SAR images, providing rare and inconsistent information, caused by the particular SAR imaging mechanisms. Unlike single target instances, prototypes derived from the average of target features at diverse postures, can provide class-related information in a general and stable way for better identification.

3.3. Overall Framework

The overall framework of our method is given by Figure 1 and can be divided into three stages.

• At the base learning stage ($t=1$), a CNN-based feature extractor parameterized by $f(x;\mathsf{\Phi })$ is trained on sufficient data ${\mathcal{D}}^1$ to learn an embedding space for identifying base classes while extracting general patterns of unknown ones. Like the dynamically real-world SAR ATR scenarios and the prominent partiality of targets in SAR imagery, the model’s forward compatibility with unknown targets is promoted by a Virtual-class Semantic Synthesizer (VSS), in which numerous virtual classes with soft labels are synthesized from pre-encoded real-class embeddings $h(x;\varphi )$. After being post-encoded by $g(·;\phi )$ on both real and virtual classes and optimized by a cross-entropy (CE) loss, a well-spanned $f(x;\mathsf{\Phi })$ with base-class prototypes can be obtained.
• At incremental learning stages ($t>1$), where new-class data ${\mathcal{D}}^t,(t>1)$ continually appear with few-shot samples, the model’s forward compatibility is dynamically released by a Decoupled Margin Adaptation (DMA) strategy by merely fine-tuning high-level semantic parameters in a timely manner. Therefore, the similarities of few-shot samples of novel classes to class prototypes and the discrepancy with interclass ones can be improved effectively. After optimization, the prototypes of novel classes are obtained for further inference.
• For the inference of session t, like the general patterns provided by class prototypes, a Nearest-Class-Mean (NCM) classifier is formulated for identification by comparing semantic features of unknown targets with all-class prototypes.

3.4. Forward Compatibility

In FCPC, a Virtual-class Semantic Synthesizer (VSS) and a Decoupled Margin Adaptation (DMA) strategy are designed to improve the model’s forward compatibility with novel targets at the base and incremental stages, respectively.

3.4.1. Virtual-Class Semantic Synthesizer

At the base stage, a Virtual-class Semantic Synthesizer (VSS) is designed to generate numerous virtual classes with soft labels for enhancing the model’s forward compatibility with incoming targets in advance, inspired by the latent connections between physical-aware scattering parts of targets in SAR imagery and the representing learning paradigm of the DL. The former, supported by the Attribute Scattering Center (ASC) theory [34], formulates targets in SAR images as the composition of class-agnostic physical parts and structures, e.g., dihedral, trihedral, cylinder, and so on. The latter serve as the cornerstone of DL-based learners, representing target semantic cues as the collections of convolutional activations. Therefore, the richer semantic patterns captured by our method, the more diverse scattering parts it can perceive. Algorithm 1 gives the process of the VSS concretely.

• Virtual Class Generation (VCG): As the rich and diverse target parts provided by ${\mathcal{D}}^1$, numerous virtual classes are synthesized by mixing up real-class semantic parts within a batch data $\mathcal{B}$. In Algorithm 1, $\left|\mathcal{B}\right|$ is the sample number of $\mathcal{B}$. The larger the repeated time M, the more diverse the components of virtual classes that can be generated. $\mathcal{S}$ saves the virtual-class instances. In Equation (1), we sample $\lambda$ from a $Beta(\alpha ,\alpha )$ to control the overlaps between real and virtual classes. Notably, instead of synthesizing targets in input space, the intermediate spatial manifolds $h_{\mathcal{B}}$ with different categories pre-encoded by $h^1(x;\varphi )$ are adopted, as are their distinctly component-aware characteristics and limited background interference.
• Soft Label (SL): The more mixing-up cues provided by virtual labels, the more diversified connections the model can learn from the virtual samples. In response, different from assigning virtual targets with newly one-hot labels ${\widehat{y}}_i$, a Gaussian soft-labeling function varying with the $\lambda$ given by Equation (2) is designed to generate soft supervisions. In particular, the labels for virtual targets reach the highest values as the $\lambda$ reaches 0.5.

Algorithm 1 The procedure of VSS.

Input A batch data $\mathcal{B}={\{x_i,y_i\}}_i^{\left|\mathcal{B}\right|}$ sampled from ${\mathcal{D}}^1$ Require Repeated time M, target embeddings $h_{\mathcal{B}}={\left\{h(x_i;\varphi )\right\}}_i^{\left|\mathcal{B}\right|}$ of $\mathcal{B}$, $\alpha$ of $Beta(\alpha ,\alpha )$ for m = 1 to M do       1. get ${\tilde{h}}_{\mathcal{B}}=\left\{h({\tilde{x}}_i;\varphi )\right\}$ by shuffling $h_{\mathcal{B}}$ randomly.       2. get one-hot labels ${\left\{y_i\right\}}_i^{\left|\mathcal{B}\right|}$ of $h_{\mathcal{B}}$ and shuffled one-hot ones ${\left\{{\tilde{y}}_i\right\}}_i^{\left|\mathcal{B}\right|}$ of ${\tilde{h}}_{\mathcal{B}}$.       3. get virtual-class embeddings ${\widehat{h}}_{\mathcal{B}}=\left\{h({\widehat{x}}_i;\varphi )\right\}$ from $h_{\mathcal{B}}$ and ${\tilde{h}}_{\mathcal{B}}$ following Equation (1).       4. get corresponding soft labels ${\left\{{\widehat{y}}_i\right\}}_i^{\left|\mathcal{B}\right|}$ from $\left\{y_i\right\}$ and $\tilde{y}$ following Equation (2)       5. append ${\widehat{h}}_{\mathcal{B}}$ and ${\left\{{\widehat{y}}_i\right\}}_i^{\left|\mathcal{B}\right|}$ to $\mathcal{S}$. end for Output$\mathcal{S}$

$$h({\widehat{x}}_i;\varphi )=\lambda ·h(x_i;\varphi )+(1-\lambda )h({\tilde{x}}_i;\varphi )$$

(1)

$$\begin{array}{cc}\hfill \Gamma \left(\lambda \right)& ={\displaystyle \frac{1}{\sigma \sqrt{2\pi }}}exp(-{\displaystyle \frac{{\lambda }^2}{2{\sigma }^2}})\hfill \\ \hfill {\widehat{y}}_i& =1-{\displaystyle \frac{\Gamma \left(\lambda \right)-\Gamma (1-\lambda )}{(1-\Gamma (1\left)\right)}}\hfill \end{array}$$

(2)

$${\mathcal{L}}_{VSS}={\displaystyle \frac{1}{|N_r|}}{\mathcal{L}}_{CE}(g\left(h(x_i;\varphi )\right),y_i)+{\displaystyle \frac{1}{|N_v|}}{\mathcal{L}}_{BCE}(g\left(h({\widehat{x}}_i;\varphi )\right),{\widehat{y}}_i)$$

(3)

3.4.2. Decoupled Margin Adaptation

Given highly confused and unstable target cues provided by few-shot samples, a Decoupled Margin Adaptation (DMA) strategy, which contains a Decoupled Adaptation (DA) strategy and a Prototype Margin Loss (PML), is designed to make the model continually evolve to adapt new concepts while relieving the forgetting problem after deployment.

• Decoupled Adaptation: As target cues provided by few-shot samples are extremely rare and unstable, directly tuning whole parameters of our method can easily lead to the overfitting problem. In response, a decoupled adaptation (DA) strategy is introduced to divide the method’s whole parameters into low-level and high-level parameters based on the layers. During incremental learning, only the high-level parts containing abstract information are tuned while freezing the low-level ones with general cues to balance the representation of class-agnostic patterns and the rapid adaptation to class-specific patterns.
• Prototype Margin Loss: Given the rare and unstable target cues provided by few-shot instances, a Prototype Margin Loss (PML) is designed to update the model’s parameters when necessary, avoiding inappropriate adaptation. Unlike the cross-entropy (CE) loss, the PML ensures robust feature learning under limited data. As defined in Equation (4), $f^t(x_c;\mathsf{\Phi })$ and $v_c$ represent the deep embeddings of a novel class $c,c\in {\mathcal{C}}^t$, and the parameterized class-related prototypes, respectively. Additionally, another class prototype $v_j$, corresponding to the top J nearest distances to the $f^t(x_c;\mathsf{\Phi })$, is selected for comparison. The cosine distance d is adopted for judgment due to its strong connection to azimuth-aware target structures, as demonstrated in [29]. The parameter m controls the discrepancy margin. By minimizing ${\mathcal{L}}_{PML}$, the compact intraclass feature spaces and well-separated interclass feature spaces can be learned properly.

$${\mathcal{L}}_{PML}=\sum _{(x_c,y_c)}^{{\mathcal{D}}^t}\sum _j^{J}max(d(f^t(x_c;\mathsf{\Phi })-v_c)-d(f^t(x_c;\mathsf{\Phi })-v_j)+m,0)$$

(4)

3.5. Nearest-Class-Mean Classifier

At inference, a Nearest-Class-Mean (NCM) classifier is maintained for stable discriminability. Specifically, as shown by the bottom subfigure of Figure 1, semantic features of an unknown target x are first extracted by the current feature extractor $f^t(x;\mathsf{\Phi })$. Then, the cosine distances d between $f^t(x;\mathsf{\Phi })$ and all-class prototypes ${\left\{p_i\right\}}_{i=1}^{c^t}$ are compared. Here, the class prototype $p_i$ is calculated by Equation (5), where $N_i$ is the total number of few-shot samples with class i. Finally, the prediction $\widehat{y}$ is calculated by Equation (6) and can be regarded as the class with the closest distance between the prototype and the target features.

$$\begin{array}{cc}\hfill p_i& ={\displaystyle \frac{1}{N_i}}\sum _{n=1}^{N_i}1(y_n=i)f^t(x_n;\mathsf{\Phi })\hfill \end{array}$$

(5)

$$\begin{array}{cc}\hfill \widehat{y}& =arg\underset{i\in c^t}{max}d\langle p_i,f^t(x;\mathsf{\Phi })\rangle \hfill \end{array}$$

(6)

Notably, instead of using Euclidean distance for evaluation, the cosine distance is employed to measure the normalized similarity between prototypes and test samples. This approach effectively captures target structural characteristics, thereby enhancing discrimination accuracy and reducing background interference.

4. Results

In this part, the experimental datasets, the implementation details, and the evaluation metrics are first depicted. Afterward, ablation studies for the proposed modules and extensive results are provided.

4.1. Dataset Preparation

In the experiments, two FSCILs of the SAR ATR datasets, which contain the MSTAR-FSCIL and the SAR-AIRcraft-FSCIL, are leveraged for validation.

4.1.1. MSTAR-FSCIL

Following the [30], the MSTAR-FSCIL dataset derived from the Standard Operation Condition (SOC) of the MSTAR [12] dataset is leveraged to verify the effectiveness of our method for training fine-grained ground vehicles captured by airborne platform. According to the types of the vehicles, four types of targets, i.e., BTR70, 2S1, BRDM2, and BMP2, with full-azimuth samples imaged at a 17$°$ depression angle are adopted to serve as the base session data ${\mathcal{D}}^1$. The remaining six with randomly selected few-shot samples are used for incremental learning. Furthermore, the same categories supported by full-azimuth instances imaged at a 15$°$ depression angle are used for validation.

Table 1. Configurations of the MSTAR-FSCIL dataset.

Session	Order	Type	Serial No.	Train	Test
Base	1	BTR70	c71	233	196
	2	2S1	b01	299	274
	3	BRDM2	E-71	298	274
	4	BMP2	9563	233	196
Incremental	5	ZIL131	E12	5	274
	6	T62	A51	5	273
	7	D7	92v13015	5	274
	8	BTR60	k10yt7532	5	195
	9	T72	132	5	196
	10	ZSU234	d08	5	274

shows the dataset configurations in detail. Figure 2 shows the targets in SAR and optical images.

4.1.2. SAR-AIRcraft-FSCIL

The SAR-Aircraft-FSCIL dataset was constructed from the public dataset named SAR-Aircraft-1.0 [35], which was developed and released by the Chinese Academy of Sciences (CAS) in 2023. This dataset comprises seven distinct classes of aircraft, including the A220, A320, A330, ARJ21, Boeing737, Boeing787, and others. All samples were acquired from three international airports using data captured by the Chinese Gaofen-3 satellite under C-band single polarization in spotlight mode, achieving a spatial resolution of 1 m. Following the framework established in [30], four classes with the largest sample sizes, namely, Other, A220, Boeing787, and Boeing737 were used for base session learning. The remaining three classes (A320, ARJ21, and A330) were designated as incremental classes ${\mathcal{D}}^t,t>1$. For training, each base class was represented by 2000 samples, while the incremental classes received only five samples each. During validation, a random selection of 200 samples was drawn across all classes to ensure balanced testing. A detailed overview of the SAR-Aircraft-FSCIL dataset is provided in

Table 2. Configurations of the SAR-AIRcraft-FSCIL dataset.

Session	Order	Type	Train	Test
Base	1	Other	2000	200
	2	A220	2000	200
	3	Boeing787	2000	200
	4	Boeing737	2000	200
Incremental	5	A320	5	200
	6	ARJ21	5	200
	7	A330	5	200

, which includes specific configurations such as imaging parameters and class distributions. Additionally, Figure 3 illustrates representative examples of targets in both SAR and optical imagery, offering a comparative perspective on the dataset’s characteristics.

4.1.3. FUSAR-FSCIL

The FUSAR-Ship [36] dataset is an open SAR-AIS matchup dataset of Gaofen-3 satellite released by Fudan University for ship and marine target detection and recognition. The FUSAR-Ship dataset is constructed from over 100 GF-3 scenes covering a large variety of sea, land, coast, river and island scenarios. It includes over 5000 ship image chips covering about 10 types of marine targets. Figure 1 shows the SAR and optical images of the ten selected types of ships in the FUSAR-Ship dataset. Following the process of the MSTAR-FSCIL and SAR-AIRcraft-FSCIL datasets, a FUSAR-FSCIL dataset is constructed from the FUSAR-Ship dataset for task-related ship recognition evaluation. Here, four types of ships, i.e., Cargo, Other, Fishing, and BulkCarrier, with abundant samples are selected for base-session training. The six remaining types of ships, i.e., Tanker, Unspecificied Container, Dredger, Tug, and GeneralCargo, are used for incremental learning.

Table 3. Configurations of the FUSAR-FSCIL dataset.

Session	Order	Type	Train	Test
Base	1	Cargo	240	30
	2	Other	240	30
	3	Fishing	240	30
	4	BulkCarrier	240	30
Incremental	5	Tanker	5	30
	6	Unspecificied	5	30
	7	Container	5	30
	8	Dredger	5	30
	9	Tug	5	30
	10	GeneralCargo	5	30

shows the configurations of the constructed FUSAR-FSCIL dataset. Additionally, Figure 4 presents both SAR and optical imagery of selected ships of the FUSAR-FSCIL dataset.

4.2. Implementation Details

In all experiments, we randomly selected five-shot samples for each incremental class and conducted $T=10$ trials for statistical evaluation.Following the established protocol in [24], a ResNet-18 network pre-trained on ImageNet serves as the feature extractor $f(x;\mathsf{\Phi })$ for all compared methods for balancing the model’s performance and computational efficiency. Its architecture comprises 18 convolutional layers enhanced with skip connections to mitigate the vanishing gradient problem. Regarding hyperparameters, for the VSS method, the number of repetitions M is set to 4, and the value of $\alpha$ in the $Beta(\alpha ,\alpha )$ is 0.5. For the DMA, the margin m for the PML is set as 0.5, and features from the first residual block of the ResNet-18 are utilized for the synthesis process. For the base session training, all benchmarks are trained using the Stochastic Gradient Descent (SGD) optimizer for 50 epochs. The initial learning rate is set to $1\times {10}^{-2}$ and is decayed to $1\times {10}^{-3}$ at epoch 30 and further to $1\times {10}^{-4}$ at epoch 40. The mini-batch size is set to 32. For incremental learning, the learning rate is initialized to $1\times {10}^{-4}$ and decayed by a factor of 0.1 at epoch 30. All hyperparameters for the compared benchmarks are configured as described in their original papers. The input images are normalized to the range [0, 1] and resized to 64 × 64 pixels. For data augmentation, all inputs are randomly rotated within $\pm 5^{\circ }$. All experiments are conducted on a NVIDIA GTX 3080Ti GPU (NVIDIA Corporation, Santa Clara, CA, USA) with CUDA 11.1.

4.3. Evaluation Metrics and Benchmarks

In experiments, three metrics are used for evaluation. First, the classification accuracy (Acc.) is reported to evaluate the benchmark’s performance on categories seen at each session. Second, the average accuracy (Avg. Acc), calculated by averaging the Acc. of all sessions, reflects the model’s comprehensive performance on all categories seen. Third, the performance drop (PD) rate [27] is provided to measure the absolute performance deterioration by subtracting the accuracy of the last session from that of the first.

Following [30], numerous task-specific benchmarks are adopted for evaluation, as described below. Traditional DL-based methods: Ft-CNN and Oracle are employed. The former fine-tunes a CNN-based classifier on the data of the current session using a cross-entropy loss, while the latter is an offline learner trained on all data from seen classes. IL solutions: Three typical incremental learning (IL) classifiers—iCaRL [37], EEIL [38], and LUCIR [39]—are utilized. These methods follow a replay-based paradigm, where all samples of new classes are stored for learning in subsequent sessions to ensure a fair comparison. FSCIL solutions: Methods including TOPIC [24], ERDIL [25], IDLVQC [26], CEC [27], FACT [9], ALICE [10], SAVC [40], CPL [29] and AASC [30] are adopted for comprehensive evaluation. The key features of these algorithms are discussed in the Related Works (FSCIL) section (Section 2.2). Notably, the last two methods (CPL, AASC) are specifically designed for the FSCIL in SAR ATR.

4.4. Ablation Study

The model’s forward compatibility supported by the two modules, named the VCG and the SL, within the VSS and stable discriminability supported by the DA, the PML, and the NCM classifier are explored in this section. The overall numerical performance is summarized in

Table 4. Effects of the proposed modules evaluated on the MSTAR-FSCIL dataset.

Forward-Compatible		Stable Discriminating			MSTAR-FSCIL (%)
VCG	SL	DA	PML	NCM	Avg. Acc	Avg. HA	PD ↓
-	-	-	-	✓	75.26	61.37	41.25
✓	-	-	-	✓	79.28	71.26	35.51
-	✓	-	-	✓	77.64	65.11	36.40
✓	✓	-	-	✓	82.49	76.59	32.83
✓	✓	✓	-	✓	82.75	77.64	32.50
✓	✓	✓	✓	✓	83.03	78.13	32.15

. For a fair comparison, a baseline model trained solely on the base session data with the NCM classifier is utilized, with its results presented in the first column of the table.

4.4.1. Effects of VSS

The VSS is employed to enhance the model’s ability to perceive unknown categories in advance. The quantitative and qualitative results of this enhancement are presented in the first four rows of Table 4, as well as in Figure 5 and Figure 6.

• Quantitative performance: The contributions of the two submodules, namely the VCG and the SL, are progressively verified. Compared to the baseline results presented in the first row of the table, our method with the VCG achieves 79.28%, 71.26%, and 35.51% in Avg Acc, Avg. HA, and PD, respectively, as shown in the second row of the table. Additionally, the corresponding results for the method with the SL are 77.64%, 65.11%, and 36.40%. Finally, as indicated in the fourth row of the table, the scores of our method with the VSS, i.e., with both the VCG and SL, reach 82.49%, 76.59%, and 32.83%. These results are 7.23%, 15.22%, and 4.85% superior to the baseline, demonstrating the effectiveness of the VSS and its submodules.
• Qualitative performance: The feature diversity and significance are given by Figure 5. The former is quantified by its non-sparsity, measured as the ratio of non-zero feature channels to the total number of channels. The latter is represented by the L2-norm, which corresponds to the overall magnitude of the target features. Larger values correspond to features with more diversity and importance. Here, our method, facilitated by the VCG and the SL, can achieve progressively competitive scores in non-sparsity (Figure 5a) and normalization (Figure 5b). Notably, the much more diverse features can be captured by our method by using both the VCG and the SL. More intuitively, the t-SNE embeddings of class features extracted by our method with the VSS are shown in Figure 6. Here, compared to the baseline, features extracted by our method with the VSS can be distributed more uniformly, implicitly reflecting the diverse and rich target cues acquired by our method.

4.4.2. Effects of DMA

Although the model’s forward compatibility with unpredictable incoming classes is prompted by the VSS, directly applying the model trained on the base session data can inevitably result in a performance drop due to the semantic drift between the current feature space and the few-shot new classes encountered in the dynamic world. For complementarity, the designed DMA, composed of the DA and PML, is leveraged for the model’s rapid adaptation during incremental learning stage.

• DA: The DA is employed to balance the model’s stability for class-agnostic knowledge and its plasticity for class-specific knowledge by leveraging the hierarchical features from different convolutional layers. As shown in the fifth row of Table 4, our method with the DA achieves more competitive results compared to the version without this module, with Avg Acc, Avg. HA, and PD values of 82.75%, 77.64%, and 32.50%, respectively. Additionally, the effects of varying the locations of the DA are explored in Figure 7a; the method with fewer and deeper trainable parameters achieves higher performance in terms of Avg Acc and Avg HA compared to configurations with shallower and more trainable layers. Furthermore, as demonstrated in Figure 7b, the model with the trainable layer4 achieves the best performance in classifying old classes at each session. This is attributed to its ability to balance static extraction of low-level general features and dynamic adaptation for class-aware semantic features, resulting in lower drifts in old-class features compared to other variants.
• PML: Our method’s proper adaptability to incoming targets is guaranteed by the PML, for which the contributions are shown in the last row of Table 4. Notably, the Avg Acc, the Avg. HA, and the PD of our method with the PML reach 83.03%, 78.13%, and 32.15%, 0.35%, which is superior to the method without the PML, demonstrating the effectiveness of the module. Furthermore, more investigations into session-wise class separations (R) [40] and interclass distances are presented in Figure 8a. Here, scores of the R among all classes learned by our method with the PML are consistently higher than those of methods without the module or with the CE loss across all sessions. Furthermore, as shown in Figure 8b, benefiting from the clear margin constraint between target samples and selected class weights considered by the PML, interclass distances learned by the loss are significantly superior to those learned by the methods without or with the CE. Henceforth, clear separations can be learned by the PML.

4.5. Benchmark Performance

In this section, comprehensive experiments on three derived benchmarks, namely MSTAR-FSCIL, SAR-AIRcraft-FSCIL, and FUSAR-FSCIL, are conducted to evaluate the effectiveness of our method.

4.5.1. Quantitative Evaluation

The quantitative performance of the compared benchmarks, evaluated on the MSTAR-FSCIL, SAR-AIRcraft-FSCIL, and FUSAR-FSCIL datasets, is presented in

Table 5. Comparison of benchmarks on the MSTAR-FSCIL dataset.

Methods	Sessions							Avg Acc	PD ↓
	1	2	3	4	5	6	7
Ft-CNN	98.94	76.21	62.01	52.02	47.55	43.08	38.12	59.70	60.82
Oracle	98.94	85.05	77.17	78.30	73.86	70.34	67.10	78.68	31.84
iCaRL [37]	92.12	83.08	73.46	66.41	65.04	60.42	54.27	70.69	37.85
EEIL [38]	98.94	81.76	70.77	62.78	63.35	61.55	56.68	70.83	42.26
LUCIR [39]	99.89	90.07	76.97	72.60	70.33	66.67	60.89	76.77	39.00
TOPIC [24]	91.10	85.27	72.77	63.25	61.24	56.03	50.03	68.53	41.07
ERDIL [25]	98.94	87.92	76.02	70.14	68.66	64.17	57.94	74.83	41.00
IDLVQC [26]	97.02	83.75	71.12	62.27	59.54	54.71	49.20	68.23	47.82
CEC [27]	90.54	80.52	72.27	72.53	66.97	61.76	57.97	71.79	32.57
FACT [9]	98.85	87.36	67.14	64.24	49.50	46.69	45.01	65.54	53.84
ALICE [10]	97.55	86.83	72.36	66.73	63.05	59.36	54.44	71.47	43.11
SAVC [40]	97.10	89.43	79.64	75.66	72.47	68.48	64.64	78.20	32.46
CPL [29]	99.89	90.41	78.32	74.93	73.04	69.66	63.66	78.56	36.23
AASC [30]	99.89	89.73	76.33	73.64	71.46	66.97	62.31	77.19	37.58
Ours (FCPC)	99.47	94.48	85.28	82.10	78.31	74.26	67.32	83.03	32.15

,

Table 6. Comparison of benchmarks on the SAR-AIRcraft-FSCIL dataset.

Methods	Sessions				Avg Acc	PD ↓
	1	2	3	4
Ft-CNN	99.62	82.93	67.66	59.32	77.38	40.30
Oracle	99.47	81.24	69.83	64.31	78.71	35.16
iCaRL [37]	99.50	88.82	76.20	65.37	82.47	34.13
EEIL [38]	99.62	85.00	73.13	67.26	81.25	32.36
LUCIR [39]	99.75	87.68	75.76	68.23	82.86	31.52
TOPIC [24]	99.64	89.09	76.95	67.06	83.18	32.58
ERDIL [25]	99.62	86.50	75.51	68.80	82.61	30.82
IDLVQC [26]	99.37	85.53	72.80	64.10	80.45	35.27
CEC [27]	78.97	65.64	57.76	51.75	63.53	27.22
FACT [9]	99.44	79.55	66.36	56.88	75.56	42.56
ALICE [10]	95.63	74.62	62.30	55.66	72.05	39.97
SAVC [40]	99.04	83.44	73.74	67.97	81.05	31.07
CPL [29]	99.37	88.55	77.98	72.06	84.49	27.31
AASC [30]	99.50	89.12	77.18	69.28	83.77	30.22
Ours (FCPC)	99.37	91.09	79.15	71.80	85.35	27.57

, and

Table 7. Comparison of benchmarks on the FUSAR-FSCIL dataset.

Methods	Sessions							Avg Acc	PD ↓
	1	2	3	4	5	6	7
Ft-CNN	75.83	58.53	48.11	42.29	35.96	32.22	30.30	46.18	45.53
Oracle	79.50	65.20	56.44	49.33	45.75	39.48	36.40	53.16	43.10
iCaRL [37]	75.83	73.40	59.33	52.57	47.13	42.41	39.13	55.69	36.70
EEIL [38]	75.83	62.53	50.94	46.76	40.33	35.41	34.73	49.50	41.10
UCIR [39]	80.83	71.47	60.28	53.19	46.67	41.11	37.73	55.90	43.10
TOPIC [24]	75.67	67.00	57.11	50.86	44.46	40.41	37.20	53.24	38.47
ERDIL [25]	77.50	60.40	43.72	36.62	26.04	22.59	21.30	41.17	56.20
IDLVQC [26]	75.00	65.60	54.56	44.90	39.08	35.56	30.83	49.36	44.17
CEC [27]	84.17	71.53	60.61	54.62	48.75	44.30	40.43	57.77	43.74
FACT [9]	77.50	75.00	63.83	56.52	50.00	45.04	40.10	58.28	37.40
ALICE [10]	60.00	56.60	49.00	45.10	40.42	36.15	33.23	45.79	26.77
SAVC [40]	78.83	72.60	60.72	54.29	49.42	44.56	40.73	57.31	38.10
CPL [29]	80.17	70.53	59.61	52.62	46.25	43.23	40.13	56.08	40.04
AASC [30]	82.23	70.92	60.53	53.21	47.35	43.55	39.33	56.73	42.90
Ours (FCPC)	80.17	76.33	64.94	57.86	51.46	47.85	44.20	60.40	35.97

respectively. Several conclusions can be drawn from the results.

• Firstly, owing to the specially designed techniques aiming to enhance forward compatibility and stable discriminability, our method achieves competitive performance in terms of Avg Acc compared to other benchmarks. On the MSTAR-FSCIL dataset, the Avg Acc of our method reaches 83.03%, surpassing Oracle, SAVC, and CPL by 4.35%, 4.83%, and 4.47%, respectively. Similarly, on the SAR-AIRcraft-FSCIL dataset, our method outperforms Oracle, SAVC, and CPL by 6.64%, 2.49%, and 0.86%, respectively. On the FUSAR-FSCIL dataset, the score of the Avg Acc of our method reaches 60.40%, which is 2.12% higher than the second-ranked solution, demonstrating the effectiveness of our method for coping with the FSCIL problem in SAR ATR field.
• Secondly, our method effectively addresses the catastrophic forgetting issue, resulting in superior performance on the PD metric. Specifically, our method achieves PD scores of 32.15% and 27.57% on the MSTAR-FSCIL and SAR-AIRcraft-FSCIL datasets, respectively. Unlike most IL and FSCIL methods, which commonly rely on less forgetting losses and replay strategies to mitigate forgetting, our approach with the DMA strategy reaches a balance between static feature representation and dynamic class adaptation, overcoming the limitations of sparse and confused information within limited samples.
• Thirdly, our method achieves the competitive generalization ability in solving the FSCIL in comparison to public benchmarks. For further validation, a combined dataset with ten sessions (one base + nine incremental) is constructed by combining the FSCIL-MSTAR with the SAR-AIRcraft-FSCIL datasets. Methods are first trained on base classes of the two datasets and then optimized on novel classes of the FSCIL-MSTAR and the SAR-AIRcraft-FSCIL incrementally. The results are given in

  Table 8. Comparison of benchmarks on the Combined dataset.

  Methods	Sessions										Avg Acc	PD ↓
  	1	2	3	4	5	6	7	8	9	10
  Ft-CNN	90.13	80.29	72.01	65.24	59.85	55.22	50.96	47.82	44.99	42.40	60.89	47.73
  Oracle	91.77	88.48	83.35	81.09	77.58	71.17	70.97	67.53	63.77	61.28	75.70	30.49
  iCaRL [37]	85.31	82.21	73.11	67.42	65.62	62.09	58.18	56.42	53.57	50.72	65.46	34.59
  EEIL [38]	90.13	80.03	73.08	68.74	63.97	60.20	57.72	56.08	53.92	51.98	65.59	38.15
  UCIR [39]	91.19	89.07	83.54	78.03	75.79	72.68	70.88	68.28	65.21	62.88	75.76	28.31
  TOPIC [24]	87.50	84.22	73.50	68.09	65.46	63.15	59.61	58.00	55.12	51.29	66.59	36.21
  ERDIL [25]	89.30	85.12	74.54	67.59	64.16	62.01	59.11	58.24	56.32	53.01	66.94	36.29
  IDLVQC [26]	84.22	82.39	73.59	65.43	62.26	60.93	58.93	58.13	55.02	53.84	65.47	30.38
  CEC [27]	91.18	88.98	79.86	73.52	70.79	66.62	64.35	61.92	58.78	57.15	71.31	34.03
  FACT [9]	90.19	87.61	81.15	76.00	73.46	68.46	65.11	61.83	59.59	58.35	72.18	31.84
  ALICE [10]	90.19	86.51	83.35	77.65	74.54	70.53	66.19	64.34	60.87	59.33	73.35	30.86
  SAVC [40]	88.37	85.11	81.2	81.45	77.71	74.42	73.68	70.5	68.66	66.21	76.73	22.16
  Ours (FCPC)	91.19	87.87	81.80	80.56	78.50	75.16	73.53	71.06	68.52	65.83	77.40	25.36

  . Our method achieves competitive performance on the combined dataset compared to other public benchmarks. Furthermore, our method achieves the average accuracy (Avg Acc) of 77.40%, which is 1.7% higher than the second-ranked solution, verifying its strong generalization ability for continual learning across diverse categories.

4.5.2. Qualitative Evaluation

The qualitative performance of compared benchmarks and the corresponding analysis are given in this section.

• Session performance curves: The accuracy (Acc) line charts for all compared methods evaluated on of the three datasets are illustrated in Figure 9a–c. Overall, our method with the designed modules consistently achieves the most competitive performance in terms of the Acc across all sessions. Furthermore, the richer and clearer the target-discriminating cues from abundant base classes, the stronger the forward compatibility that could obtained with our method. For instance, the MSTAR-FSCIL dataset, which contains more diverse target components provided by its full-azimuth targets, enables our method’s forward compatibility. Consequently, more competitive performance can be obtained by our method on the MSTAR-FSCIL dataset than that on the SAR-AIRcraft dataset.
• Confusion matrix: The normalized confusion matrices on the final session of the two datasets are shown in Figure 10, Figure 11 and Figure 12, respectively. Overall, benefiting from the emphasis on both the model’s forward compatibility with incoming classes and the stable discriminability based on limited samples, the colors of diagonal blocks of matrices for both base and incremental classes predicted by our method are more harmonic and bright than those by other benchmarks. For example, most approaches perform well on base classes but fail to judge the new ones, inducing biased confusion matrices. In addition, the more clear and discerning cues provided by limited samples, the more competitive and balanced results the compared benchmarks can reach. For example, the compared benchmarks widely perform better on the MSTAR-FSCIL dataset than on the others, thanks to more distinct cues of targets under ideal conditions than that on the SAR-AIRcraft-FSCIL.
• Session-wise t-SNE results: Considering the limited testing samples of the FUSAR-FSCIL dataset, the t-SNE are solely conducted on the MSTAR-FSCIL and SAR-AIRcraft-FSCIL datasets, and the results are also shown in Figure 13 and Figure 14, respectively. First, our method, with special consideration on its forward compatibility, can reserve more space for the new, leading to more separated distributions of interclass high-dimensional features than those produced by the baseline. Second, the clearer the target-discerning cues provided by new-class samples, the better distinguishing ability our method can possess. Also, as shown by Figure 13 and Figure 14, the t-SNE results for new classes in the MSTAR-FSCIL are more separated than those in the SAR-AIRcraft-FSCIL. Specifically, interclass features are more separated, while the intraclass ones are more compacted from our method, verifying the importance of unleashing the model’s forward compatibility before deployment.

5. Discussion

Analysis of Experimental Results. The comprehensive experimental results clearly demonstrate that both forward compatibility and stable discriminability are crucial for solving the FSCIL in the SAR ATR field. Due to the specific imaging mechanisms of the SAR and the limited novel samples, new-class target cues in openly dynamic environments are certainly rare and instable. Unlike existing algorithms focusing on identifying current classes or solely being optimized on new instances, our method with forward compatibility supported by the VCG and SL can proactively learn class-agnostic generalized information from sufficient base classes. Meanwhile, the model’s stable discriminability, supported by the DA, PML, and NCM, can learn and classify novel targets rapidly by leveraging few-shot samples. As a result, the model’s stability and plasticity can be balanced properly.

Potential Limitations. The model’s representability and discriminability to diverse incoming targets heavily relly on the high-quality features of both base and novel classes. First, as unknown targets are obtained by the mixing of base-class features and labels, significant feature discrepancies may exist between these two types of categories once the base-class features are highly incomplete or biased. Second, although target general features can be provided by class prototypes of the NCM classifier, which are derived from the average of the class-aware features, the model’s stable discriminability may still deteriorate due to the loss of fine-grained details and the diversity of target azimuth-aware features. Thus, effectively integrating both base-class and new-class features remains crucial for enhancing the model’s perception and discrimination of unknown targets.

Future Research. Since current methods primarily rely on semantic information, future work should prioritize scattering characteristics for target representation and discrimination, given the unique complexities of SAR imaging. Moreover, existing approaches largely depend on task-specific learning from new categories, inherently restricting their adaptability in open-world scenarios. Instead, a task-agnostic learning paradigm, e.g., meta-learning, would be far more suitable for enabling rapid and flexible discrimination.

6. Conclusions

We proposed a Forward Compatible Prototype Classifier (FCPC) to power the DL-based SAR ATR systems with Few-Shot Class-Incremental Learning (FSCIL) ability to satisfy real-world SAR ATR scenarios. The FCSC’s forward compatibility and stable discriminability were emphasized and promoted. For forward compatibility, a VSS was designed to synthesize virtual features with soft labels to unleash the ability before deployment leveraged by the intrinsic links between target partiality and DL’s representing learning paradigm. For stable discriminability, our method was decoupled and evolved from knowledge-oriented new-class fingerprints using an DMA strategy to balance the representation of class-agnostic patterns and the prompt adaptation to class-specific ones. An NCM classifier is maintained for identification without losing generalization. In experiments, the contributions of the designed modules were verified. Extensive experiments on two task-related datasets, i.e., MSTAR-FSCIL and SAR-AIRcraft-FSCIL, showed the effectiveness of our method for the FSCIL in openly dynamic SAR ATR scenarios compared with numerous latest benchmarks.