Few-Shot Learning (FSL) aims to enable models to recognize novel categories using only a few labeled samples per class (Tsoumplekas et al., 2024). This paradigm is particularly important in scientific domains such as environmental monitoring and biological taxonomy, where data collection is expensive and sample acquisition is constrained (Bennequin et al., 2022; Lim et al., 2025).

Existing FSL methods mainly follow three approaches. First, data augmentation-based methods expand limited datasets by generating or transforming samples (e.g., via semantic priors or adversarial generation). Second, meta-learning-based methods (e.g., MAML-style learning-to-learn) optimize for rapid task adaptation across episodes (Hospedales et al., 2021). Finally, graph-based methods explicitly model relationships between support and query samples for relational reasoning (Zhou X. et al., 2022).

Although these methods have achieved success on standard benchmarks such as Omniglot and miniImageNet, they are mostly evaluated on static, clean, and well-separated tasks, which can deviate from realistic fine-grained and evolving deployments (Bennequin et al., 2022). This differs substantially from real-world scenarios (e.g., environmental microbiology) where new morphotypes continually emerge, and the learner must adapt under extreme sample scarcity over time; consequently, conventional FSL alone is often insufficient for continuously evolving environments (Lim et al., 2025).

Class-Incremental Learning (CIL) enables machine learning models, most commonly neural network, to recognize an expanding set of classes introduced progressively over time. Unlike FSL, CIL typically assumes that each newly introduced class is supported by a relatively sufficient number of training samples, while requiring the model to preserve discriminative performance on previously learned classes as new ones are acquired. The central challenge of CIL is mitigating catastrophic forgetting, a phenomenon in which the model's recognition performance on earlier classes degrades significantly as it learns new ones.

Existing CIL approaches can be broadly categorized into several methodological paradigms. Exemplar-based replay methods (e.g., iCaRL; Rebuffi et al., 2017 ) retain a small set of historical samples and leverage knowledge distillation to maintain decision boundaries for previously learned classes. Constraint or regularization-based methods (e.g., LUCIR; Hou et al., 2019 ) restrict updates to network parameters or feature representations in order to enforce representational consistency across incremental learning stages. Representation-level distillation methods perform feature-space distillation to ensure the stability of the embedding space throughout the incremental learning process. Although these methods have demonstrated strong performance on large-scale natural image benchmarks, they generally rely on the availability of sufficient, clean, and well-curated labeled data at each incremental stage and assume well-curated visual domains. Such assumptions limit their direct applicability to scientific domains such as environmental microbiome analysis, where samples from novel classes are often extremely scarce and are further characterized by substantial noise and high intra-class variability.

FSCIL combines the strengths of FSL and CIL, enabling models to continually learn new classes using only a small number of labeled samples while retaining previously acquired knowledge (Zhang et al., 2025). This task was formally defined by, who established a benchmark protocol and highlighted the compounded challenges introduced by simultaneous data scarcity and class-incremental expansion.

Representative FSCIL methods address this challenge from multiple perspectives: TOPIC (Tao et al., 2020) introduces a topology-preserving mechanism (via neural gas) to stabilize the feature-space structure across incremental stages; CEC proposes continuously evolving classifiers to mitigate forgetting and maintain coherent decision boundaries (Zhang C. et al., 2021) Meta-learning-based frameworks (e.g., MetaFSCIL; Chi et al., 2022) and forward-compatible training (e.g., FACT; Zhou D. et al., 2022) enhance adaptability by improving cross-session generalization and by calibrating representations or classifiers for future updates.

These methods have achieved significant progress on widely used benchmark datasets, which are CUB-200-2011, miniImageNet, and CIFAR-100. However, most FSCIL baselines are still primarily benchmarked on general-purpose natural-image datasets under relatively controlled acquisition conditions. Such settings can differ substantially from scientific or ecological microscopy imagery, where fine-grained morphology, non-uniform illumination, imaging noise, background impurities, and skewed class distributions are common and can stress robustness and stability. The demonstration of the FSCIL workflow is shown in Figure 1.

Environmental microorganism recognition has progressed rapidly from classical pipelines based on hand-crafted descriptors (Ciranni et al., 2024) to modern end-to-end learning frameworks (Zhang et al., 2023). Recent advances increasingly leverage deep convolutional networks and vision transformers, often coupled with metric learning and attention mechanisms to better capture fine-grained morphology. To reduce reliance on exhaustive annotations, self-supervised and contrastive pretraining, semi-weakly supervised learning, and domain adaptation have become common strategies for extracting transferable visual representations from large collections of unlabeled or heterogeneously labeled microbial images (Bendidi et al., 2024). In parallel, data-centric solutions such as strong augmentation, synthetic sample generation, and uncertainty-aware sampling further improve robustness under class imbalance and imaging variability–conditions that are especially prevalent in environmental microbiology (Panaïotis et al., 2025).

Despite these methodological improvements, environmental monitoring imposes a fundamental operational constraint: models are typically trained in a static, closed-set manner and therefore can only recognize categories present in the training data, while real deployments often encounter samples from categories that were not included during training due to task diversity and incomplete coverage. Moreover, when such out-of-training categories are encountered, only a small number of labeled instances are usually available, making adaptation under sparse supervision essential. In contrast, many standard CIL settings implicitly assume that each incremental phase provides a non-trivial amount of training data, which can be unrealistic for environmental monitoring scenarios where newly encountered categories may be represented by only a handful of specimens. Conversely, FSL is designed for extreme data scarcity but does not, by itself, address the sequential acquisition of new categories and the need to preserve performance on previously learned classes (Wang et al., 2020).

This mismatch creates a critical gap for real-world environmental microorganism recognition: systems must simultaneously (i) learn novel microbial classes from very few images and (ii) retain prior knowledge over a long horizon without catastrophic forgetting (Zhou et al., 2024). FSCIL directly targets this dual requirement, making it particularly important for environmental applications where pronounced intra-class variability and subtle inter-class differences are common. Consequently, positioning FSCIL as a core paradigm–and establishing a dedicated FSCIL benchmark for environmental microbial imagery–is essential for aligning advanced AI recognition capabilities with the dynamic, long-term monitoring demands of environmental science.

All 41 categories are partitioned into a base session and multiple incremental sessions. The base session includes 20 categories, with 25 samples per category, for fully supervised representation learning to establish a robust feature space for subsequent increments. The remaining 21 categories are organized into seven incremental sessions, each introducing three novel categories under a 3-way 3-shot setting (Zhang et al., 2024), which mirrors the sparse, progressive inclusion of previously unseen categories encountered in practical microbial monitoring. The choice of the 3-shot incremental configuration follows the standard FSCIL protocol introduced by (Zhang et al., 2024), where each new class is represented by only three support samples to simulate extreme data scarcity during deployment. To reduce evaluation bias, the category grouping is constructed to maintain taxonomic diversity and comparable difficulty across sessions. The sampling ratio used in the base session follows the widely adopted 60%–40% train–test division described by Tao et al. (2020) for CIFAR-100 and miniImageNet, ensuring that EMDS-7 is aligned with canonical FSCIL dataset construction principles and remains directly comparable to prior work. Overall, this session design follows standard FSCIL protocols and enables controlled yet realistic assessment of how models adapt to new categories while preserving prior knowledge.

In this benchmark, EMDS-7 is converted into an image-level classification setting. Specifically, each original image is treated as one classification sample, without performing bounding-box cropping or object-level filtering. This design choice avoids introducing object-detection–dependent biases and ensures that all methods receive identical, unaltered visual evidence. All images are normalized and resized to a fixed resolution of 224 × 224 before being used for training and evaluation. This unified preprocessing avoids introducing design-dependent biases and ensures a standardized input format across all FSCIL methods. To ensure reproducibility, all processed images, session-wise class partitions, and fixed train/test split lists are included in the released benchmark resources.

We follow a widely adopted and challenging FSCIL evaluation protocol to ensure rigor and comparability (Rebuffi et al., 2017). For fair comparison, all baselines use an ImageNet-pretrained ResNet-18 backbone. In the base stage, the model is trained on all images from the base categories for 100 epochs using SGD with an initial learning rate of 0.1 and standard learning rate decay. During incremental learning, we employ a frozen-backbone setting: backbone parameters remain fixed across sessions, and only the classifier is updated using the current session's 3-shot samples (Zhang C. et al., 2021). The base-stage training configuration (100 epochs with an initial learning rate of 0.1) follows widely adopted settings in existing FSCIL benchmarks and the original implementations of representative methods, and provides stable convergence on EMDS-7 without altering the relative performance ranking among methods. After each session, the model is evaluated on a cumulative test set containing all seen classes. We report Average Accuracy (AA), defined as the mean accuracy across all test sessions (Tao et al., 2020), and Performance Drop (PD), measured as the difference in accuracy between the base and final sessions, to quantify long-term stability and forgetting. This evaluation protocol corresponds to the Session-ID setting, where performance is assessed cumulatively after each incremental session, ensuring a consistent and transparent measurement of class-incremental behavior. All results are averaged over five random runs and summarized with performance evolution curves and tables.

We adopt the Session-ID evaluation protocol, where after each incremental sessions, the model is evaluated on a cumulative test set containing all classes observed up to that session. The Session-ID accuracy is defined as the top-1 classification accuracy on this cumulative set. Average Accuracy (AA) is computed as the mean of Session-ID accuracies across all sessions, and Performance Drop (PD) is defined as the difference between the base-session accuracy and the final-session accuracy. These metrics jointly quantify overall effectiveness and long-horizon stability under class-incremental learning.