Producing Plankton Classifiers that are Robust to Dataset Shift

TL;DR

This paper tackles dataset shift (DS) in plankton classification by building ZooLake2.0 and 10 OOD test cells to simulate deployment. It introduces a three-step pipeline—identify, diagnose, cure—through which DS is quantified via distributional and compositional changes in class abundances and image appearances, respectively. The authors compare CNNs and vision transformers, finding that BEiT-based ensembles with targeted augmentations and test-time rotation augmentation (the BEsT model) achieve the best OOD performance (83% on aggregated OOD data) and more faithful abundance estimates, while RGB reweighting and simple quantification methods underperform. The study provides a reproducible framework and practical strategies (ensembling, architecture choice, targeted augmentation, TTA) to bolster plankton classifiers against DS, with implications for reliable ecosystem monitoring and broader open-set/open-domain recognition tasks.

Abstract

Modern plankton high-throughput monitoring relies on deep learning classifiers for species recognition in water ecosystems. Despite satisfactory nominal performances, a significant challenge arises from Dataset Shift, which causes performances to drop during deployment. In our study, we integrate the ZooLake dataset with manually-annotated images from 10 independent days of deployment, serving as test cells to benchmark Out-Of-Dataset (OOD) performances. Our analysis reveals instances where classifiers, initially performing well in In-Dataset conditions, encounter notable failures in practical scenarios. For example, a MobileNet with a 92% nominal test accuracy shows a 77% OOD accuracy. We systematically investigate conditions leading to OOD performance drops and propose a preemptive assessment method to identify potential pitfalls when classifying new data, and pinpoint features in OOD images that adversely impact classification. We present a three-step pipeline: (i) identifying OOD degradation compared to nominal test performance, (ii) conducting a diagnostic analysis of degradation causes, and (iii) providing solutions. We find that ensembles of BEiT vision transformers, with targeted augmentations addressing OOD robustness, geometric ensembling, and rotation-based test-time augmentation, constitute the most robust model, which we call BEsT model. It achieves an 83% OOD accuracy, with errors concentrated on container classes. Moreover, it exhibits lower sensitivity to dataset shift, and reproduces well the plankton abundances. Our proposed pipeline is applicable to generic plankton classifiers, contingent on the availability of suitable test cells. By identifying critical shortcomings and offering practical procedures to fortify models against dataset shift, our study contributes to the development of more reliable plankton classification technologies.

Figures (38)

• Figure 1: Schematic flowchart of our pipeline to address dataset shift. Identification: Models are usually trained, validated and tested on in-dataset (ID) data, which, due to dataset shift (DS), differs from the data found at deployment. By using 10 out-of-dataset (OOD) test cells, we simulate the performance at deployment, which is lower than the nominal ID test performance. Characterization: We quantify dataset shift and the sensitivity of our models to dataset shift. Cure: We study the effect of several modeling choices on combating DS.
• Figure 2: (a): Distribution of class abundances in the ZooLake2.0 and OOD dataset. (b): Dates of collection of the images in the ZooLake2.0 dataset and in the test cells.
• Figure 3: Distribution of classes in the 10 test cells. The column on the right indicates the day of sampling that the test cell comes from. The ID dataset is not assigned a date, because it comes from different days (Fig. \ref{['fig:dataset']}b).
• Figure 4: Graphical examples of basic augmentations compared to original images, including random flipping and rotation.
• Figure 5: Four different performance metrics (accuracy, F1 score, Bray-Curtis dissimilarity, Normalized Mean Absolute Error) in the ID training and test set (blue), in the aggregated OOD data (red), and in the individual OOD test cells, obtained with a MobileNet (see Sec. \ref{['app:models']}).