DINOv3 Improves Class-Balanced Recovery but Not Global Accuracy in Harpa Shell Classification

Methods

Dataset scope and taxonomic classes

This study used the same 13-class Harpa Röding, 1798 species-classification dataset as the preceding leakage-aware Harpa label-curation study [3]. The operational classes were Harpa amouretta, H. articularis, H. cabriti, H. costata, H. crenata, H. davidis, H. doris, H. goodwini, H. gracilis, H. harpa, H. kajiyamai, H. kolaceki, and H. major. Labels were treated as dataset labels inherited from the previous Harpa workflow, not as new taxonomic determinations made in the present study.

The fixed expanded leakage-aware dataset contained 17,923 image records, partitioned into 14,348 training records and 3,575 validation records. Species support was imbalanced. The largest classes were H. harpa, H. major, and H. articularis, whereas H. gracilis, H. kolaceki, and H. goodwini had the smallest image counts. Because of this imbalance, the evaluation was designed to consider both global accuracy and class-balanced metrics.

Image preprocessing

The image substrate was the YOLO-transformed Harpa dataset produced in the previous study [3]. Raw heterogeneous shell photographs had been processed with a custom Ultralytics YOLO shell-detection/segmentation pipeline. The selected shell instance was cropped from the source image using the segmentation-mask region, padded to avoid cutting the specimen too tightly, and placed on a square black background. Each transformed image was then treated as one classifier input.

This preprocessing was not interpreted as a separate causal factor in the present DINOv3 comparison. It defined the fixed image representation on which both the EfficientNetV2B2 baseline and DINOv3 conditions were evaluated. Validation images were not resampled, relabelled, or class-balanced.

Leakage-aware train-validation split

All model comparisons used the same leakage-aware validation framework. The split was generated before the present experiments with split seed 20260606 and a target validation ratio of approximately 20%. Related images were grouped so that repeated specimen views, transformed outputs, near-duplicates, and images derived from the same source record were assigned to the same partition. Grouping took precedence over independent image-level randomization, reducing the risk that a model could learn specimen- or source-specific shortcuts.

The validation partition was kept fixed throughout the study. Model differences were therefore interpreted as differences caused by architecture, imbalance handling, or training-data curation, not by changes in the validation substrate.

Experimental design

The study compared three main model conditions. First, an EfficientNetV2B2 baseline was trained as a three-run replicated reference condition on the uncurated leakage-aware Harpa training set. Second, a DINOv3 classifier was trained as a three-run replicated condition on the same uncurated training set, using the DINOv3 configuration selected from preliminary tuning. Third, DINOv3 was trained once after applying the very-strict statistical label-curation profile selected in the previous Harpa audit study.

The DINOv3 tuning and imbalance-strategy experiments were treated as screening analyses. They were used to select the final DINOv3 comparison condition, but they were not treated as the primary replicated architecture comparison.

EfficientNetV2B2 baseline protocol

The baseline classifier used an EfficientNetV2B2 convolutional backbone initialized from ImageNet-pretrained weights and adapted to the 13-class Harpa task by replacing the original ImageNet classification head with a genus-specific softmax classifier. Input images were resized to 400 × 400 pixels. The classifier head followed the previous Harpa framework and included global average pooling, batch normalization, dropout, and a dense 13-class softmax output layer.

The baseline protocol used Adam optimization, focal loss, a learning rate of 0.0004, top dropout of 0.4, and batch size 64. No class-balanced sampling and no class weights were used during baseline fitting. Three independent baseline runs were trained from different stochastic initializations and evaluated on the fixed leakage-aware validation records.

DINOv3 classifier protocol

DINOv3 was evaluated as an alternative pretrained vision backbone for the same 13-class closed-set Harpa classification task. The DINOv3 backbone was adapted by attaching a trainable species-classification head producing 13 softmax probabilities. Fine-tuning was partial rather than complete: only the upper part of the DINOv3 model was unfrozen during the selected comparison condition.

A preliminary DINOv3 fine-tuning screen varied the number of unfrozen top layers while keeping the dataset and validation framework fixed. The tested settings were 2, 4, 8, 16, and 32 unfrozen top layers. The 4-layer setting was retained for subsequent DINOv3 comparisons because it gave the strongest screening result without evidence that deeper unfreezing systematically improved performance. This tuning step was used only to select a training configuration; it was not interpreted as a taxonomic or biological result.

The selected DINOv3 comparison condition used categorical cross-entropy and class-balanced sampling. It did not use focal loss and did not use class weights. Class-balanced sampling oversampled minority species with replacement to equalize training exposure across classes without creating new image files. This choice was based on the imbalance-strategy screen described below.

DINOv3 imbalance-strategy screen

After selecting the 4-layer fine-tuning setting, four single-run DINOv3 imbalance-handling conditions were compared: no explicit imbalance correction, class weights only, class-balanced sampling only, and focal loss only. All four conditions used the same DINOv3 fine-tuning setting and the same leakage-aware Harpa validation framework.

This screen separated two possible optimization goals. No-balancing and focal-loss-only conditions were used to test accuracy-oriented behaviour. Class weights and class-balanced sampling were used to test whether weak or low-support species could be recovered more evenly. The class-balanced sampling condition was selected for the replicated DINOv3 comparison because it gave the strongest class-balanced behaviour in the screening analysis, rather than because it maximized single-run accuracy.

Very-strict curated DINOv3 follow-up

The final follow-up condition tested whether the very-strict statistical label-curation profile selected in the previous Harpa audit study also benefited DINOv3. This profile was not redefined in the present study. It was used as a fixed training-set intervention derived from the previous EfficientNetV2B2 audit analysis.

The very-strict profile removed 25 of the original 14,348 training records, leaving 14,323 natural training files before class-balanced sampling. The validation manifest was unchanged. The curated DINOv3 run used the same selected DINOv3 fine-tuning setting, categorical cross-entropy, class-balanced sampling, no class weights, and no focal loss. Because this curated DINOv3 condition was represented by a single run, it was treated as directional follow-up evidence rather than as a replicated curation-effect estimate.

Evaluation records and comparability checks

The validation manifest contained 3,575 files. The reported model-comparison metrics were calculated from the explicit evaluated records available in the merged evaluation artifacts. For the main EfficientNetV2B2 and DINOv3 comparison, these artifacts contained 1,602 evaluated records.

Some preliminary DINOv3 tuning and imbalance analyses used a smaller shared evaluation subset of 1,202 records to keep comparisons between screening conditions aligned. These subset analyses were interpreted only as screening and comparability checks. The main replicated comparison used the full explicit 1,602-record evaluation set.

For the very-strict curated DINOv3 run, a common-subset check was also performed because the explicit merged evaluation records were not perfectly identical to those of the earlier DINOv3 class-balanced runs. The common-subset check used only records shared between the curated run and the previous DINOv3 class-balanced cohort. This check was used to avoid overinterpreting apparent metric changes caused by differences in available explicit evaluation records.

Evaluation metrics

Model performance was evaluated on the fixed leakage-aware validation framework using top-1 accuracy, macro precision, macro recall, macro F1, and weighted F1. Top-1 accuracy was defined as the proportion of evaluated validation records for which the predicted species matched the dataset label.

Species-level precision, recall, and F1 were calculated from the validation confusion matrix. Macro metrics were calculated as unweighted averages across the 13 species and were used to assess performance across rare and common species equally. Weighted F1 was calculated by weighting species-level F1 by validation support and therefore reflected the dominant validation-class distribution more strongly.

Per-species recall was used to interpret whether changes in aggregate metrics reflected uniform improvement or redistribution of errors across classes. This was important because the previous Harpa analysis and the EfficientNetV2B2 baseline indicated that high global accuracy could coexist with weak recovery of rare or difficult species.

Statistical reporting

Three-run model cohorts were summarized by mean and sample standard deviation. Differences between the replicated DINOv3 condition and the replicated EfficientNetV2B2 baseline were reported in percentage points. Single-run screening and curated-data follow-up results were reported descriptively and were not treated as replicated evidence.

The main interpretation therefore distinguished between replicated architecture comparison and directional follow-up evidence. A DINOv3 condition was considered operationally stronger than EfficientNetV2B2 only if it improved the class-balanced metrics without unacceptable losses in strong or high-support species. A single curated DINOv3 run was interpreted as promising only when it improved the relevant metrics under the same fixed validation framework, but replication was required before treating it as a stable curation effect.

Reproducibility artifacts

The reproducibility record consisted of the fixed training and validation manifests, model-condition definitions, run seeds, training artifacts, evaluation JSON files, merged prediction tables, and species-level metric summaries. For the curated DINOv3 follow-up, the exported very-strict training manifest and the quarantine list from the previous Harpa audit workflow were also part of the reproducibility record. The validation manifest was preserved unchanged across all conditions so that model differences could be traced to architecture, imbalance handling, or training-set curation rather than to a changed evaluation set.

Results

Evaluation design and compared model conditions

The experiments evaluated species-level classification of Harpa images across 13 species classes. All model comparisons used the same leakage-aware validation framework, so that differences between conditions reflect differences in model architecture, imbalance handling, or training-data curation rather than a change in validation design.

The main comparison was made between two replicated model conditions. The first was an EfficientNetV2B2 baseline trained three times from independent stochastic initializations. The second was a DINOv3 classifier trained three times with class-balanced sampling. These two replicated cohorts form the central comparison and are reported using mean performance and between-run variation.

A third condition evaluated DINOv3 after applying the very-strict statistical label-curation profile to the training data. This condition used the same DINOv3 class-balanced strategy.

Some earlier tuning and screening analyses used smaller shared prediction subsets to keep direct comparisons fair. Those subsets are treated only as comparability checks. The main result pattern is based on averaged performance across the replicated cohorts, together with the class-level changes that explain those averages.

Role in Results	Model condition	Training-data condition	Replication level	Purpose in the analysis
Primary baseline	EfficientNetV2B2 baseline	Uncurated leakage-aware Harpa dataset	Three independent runs	Establishes the reference level for global accuracy, macro metrics, and class-level recall.
Primary DINOv3 comparison	DINOv3 with class-balanced sampling	Uncurated leakage-aware Harpa dataset	Three independent runs	Tests whether DINOv3 changes the balance between overall accuracy and species-level performance.
Curated-data follow-up	DINOv3 with class-balanced sampling after very-strict curation	Very-strict statistically curated Harpa dataset	Single run	Tests whether the conservative statistical label-curation profile also improves DINOv3 performance.
Screening context	DINOv3 tuning and imbalance-strategy variants	Uncurated leakage-aware Harpa dataset	Mostly single-run screening	Used to select the DINOv3 condition for the replicated comparison; not treated as the final model comparison.

Structure of the comparison

EfficientNetV2B2 baseline stability and class imbalance in performance

The EfficientNetV2B2 baseline was evaluated as a replicated three-run condition on the leakage-aware Harpa validation framework. Across the three independent runs, the model reached a mean top-1 validation accuracy of 89.99% ± 0.18 percentage points. The small between-run variation indicates that the baseline was stable with respect to stochastic training variation.

The global accuracy was higher than the class-balanced metrics. Macro recall was 79.73% ± 0.55 percentage points and macro F1 was 82.69% ± 1.31 percentage points, whereas weighted F1 remained close to accuracy at 89.64% ± 0.27 percentage points. This separation between global and macro-level metrics indicates that the model performed consistently well on the dominant or visually easier species, but less evenly across all species.

The class-level results explain this imbalance. Several species showed high and stable recall under the EfficientNetV2B2 baseline, including amouretta, articularis, costata, doris, and harpa. In contrast, davidis, gracilis, and kolaceki were weakly resolved. The larger classes major and cabriti were not weak in absolute terms, but they remain diagnostically important because residual errors in these classes can strongly influence the overall error profile.

Metric	Mean ± SD	Result pattern
Top-1 accuracy	89.99% ± 0.18 pp	Stable global validation performance across the replicated EfficientNetV2B2 runs.
Macro recall	79.73% ± 0.55 pp	Lower than accuracy, indicating uneven recovery of species across the 13-class problem.
Macro F1	82.69% ± 1.31 pp	Confirms that class-balanced performance is weaker than the global accuracy suggests.
Weighted F1	89.64% ± 0.27 pp	Close to accuracy because high-support classes contribute more to the weighted score.

Baseline metric profile

Class group	Species	Baseline recall pattern	Interpretation for later comparison
Strong baseline classes	amouretta, articularis, costata, doris, harpa	High recall, mostly above 94%.	These classes define the performance level that later DINOv3 comparisons should preserve.
Weak baseline classes	davidis, gracilis, kolaceki	Low recall, approximately 33–38%.	These classes are the main test of whether an alternative backbone improves difficult species rather than only matching global accuracy.
Diagnostic residual-confusion classes	cabriti, major	Moderate-to-high recall, approximately 87–88%.	These classes are not the weakest, but remain important because residual errors may influence the overall model comparison.

Selected baseline recalls by species group

These results establish EfficientNetV2B2 as a stable and high-performing baseline, but not as a uniformly strong classifier across all species. The later DINOv3 comparisons should therefore be evaluated not only by overall accuracy, but also by whether they improve the weak or diagnostic classes without losing performance in the strong baseline classes.

DINOv3 fine-tuning screen

A short DINOv3 fine-tuning screen was used to choose the number of unfrozen top layers for subsequent experiments. This screen varied the number of unfrozen layers while keeping the experiment within the same leakage-aware Harpa validation framework. The result did not show a monotonic relationship between deeper unfreezing and higher validation accuracy.

Accuracy remained within a narrow interval, from 88.69% to 89.18%. The best screening result was obtained with 4 unfrozen top layers, and this setting was therefore retained for the later DINOv3 comparisons. This screen should be interpreted only as a tuning step; it does not by itself establish a biological or taxonomic result.

Unfrozen top layers	Accuracy	Macro F1	Result pattern
2	88.69%	83.48%	Within the same narrow performance range as the other settings.
4	89.18%	84.09%	Highest accuracy in the fine-tuning screen; selected for later comparison.
8	88.85%	83.15%	No improvement over the 4-layer setting.
16	88.94%	82.67%	No evidence that deeper unfreezing improved performance.
32	88.85%	83.48%	No evidence of a systematic gain from still deeper unfreezing.

DINOv3 accuracy across unfrozen-layer settings

This screen was used only to select a DINOv3 fine-tuning setting for subsequent experiments. The narrow accuracy range indicates that, in this dataset, the number of unfrozen top layers was not the main determinant of the later model-comparison results.

DINOv3 imbalance-strategy ablation

After the DINOv3 fine-tuning screen, a small imbalance-strategy ablation was used to determine whether the observed DINOv3 error profile depended mainly on the backbone or on the way class imbalance was handled during training. Four single-run DINOv3 conditions were compared: no explicit imbalance correction, class weights only, class-balanced sampling only, and focal loss only.

The ablation showed that the DINOv3 backbone is sensitive to imbalance handling. The no-balancing and focal-loss-only conditions gave the highest single-run top-1 accuracy, both reaching 90.77%. However, these accuracy-oriented settings did not produce the strongest class-balanced metrics. The class-balanced sampling condition gave the highest macro recall (84.08%) and macro F1 (84.96%).

The class-level pattern explains this trade-off. Class weights and class-balanced sampling improved weak species such as davidis, kolaceki, and gracilis, but they reduced recall for major, a high-support class. Therefore, the ablation does not identify a single universally superior DINOv3 configuration. Instead, it separates two behaviours: an accuracy-oriented behaviour with no balancing or focal loss only, and a class-balanced behaviour with class-balanced sampling.

DINOv3 imbalance strategy	Accuracy	Macro recall	Macro F1	Weighted F1	Result pattern
No explicit balancing	90.77%	79.95%	82.61%	90.35%	Highest accuracy and weighted F1, but weaker macro performance than the sampling condition.
Class weights only	89.52%	83.91%	83.49%	89.89%	Improved macro recall, but with lower precision and no accuracy gain.
Class-balanced sampling only	89.85%	84.08%	84.96%	90.09%	Best class-balanced result; selected for replicated DINOv3 comparison.
Focal loss only	90.77%	78.99%	81.73%	90.25%	High accuracy, but weaker macro recall and macro F1 than the sampling condition.

Accuracy and macro F1 under DINOv3 imbalance strategies

Species	No explicit balancing	Class weights only	Class-balanced sampling only	Focal loss only	Interpretation
davidis	36.6%	58.5%	58.5%	31.7%	Class weights and sampling substantially improved this weak class.
kolaceki	50.0%	70.0%	60.0%	40.0%	The largest recall value was obtained with class weights, but support is very small.
gracilis	16.7%	33.3%	33.3%	16.7%	Class weights and sampling doubled recall, although the class remained weak.
major	89.5%	81.0%	81.4%	89.1%	Class weights and sampling reduced recall for this high-support class.

The ablation therefore justifies carrying class-balanced sampling forward into the replicated DINOv3 comparison. It was not selected because it maximized single-run accuracy, but because it produced the strongest class-balanced behaviour, which directly addresses the weak-species pattern observed in the EfficientNetV2B2 baseline.

Replicated comparison of EfficientNetV2B2 and DINOv3 class-balanced sampling

The central comparison was made between the replicated EfficientNetV2B2 baseline and the replicated DINOv3 condition using class-balanced sampling. Both conditions were evaluated as three-run cohorts under the same leakage-aware Harpa validation framework. Results are reported as mean ± standard deviation across runs.

EfficientNetV2B2 remained slightly stronger in mean top-1 accuracy (89.99% ± 0.18 percentage points) than DINOv3 with class-balanced sampling (89.53% ± 0.26 percentage points). The difference was small, at −0.46 percentage points for DINOv3 relative to EfficientNetV2B2.

In contrast, DINOv3 with class-balanced sampling was stronger on the class-balanced metrics. Macro recall increased from 79.73% ± 0.55 to 84.76% ± 0.97, and macro F1 increased from 82.69% ± 1.31 to 86.15% ± 0.33. Weighted F1 was effectively tied between the two model families.

Metric	EfficientNetV2B2 baseline	DINOv3 class-balanced sampling	DINOv3 minus EfficientNetV2B2	Result interpretation
Top-1 accuracy	89.99% ± 0.18 pp	89.53% ± 0.26 pp	−0.46 pp	EfficientNetV2B2 remained slightly stronger in global accuracy.
Macro recall	79.73% ± 0.55 pp	84.76% ± 0.97 pp	+5.03 pp	DINOv3 recovered more examples across the species classes.
Macro F1	82.69% ± 1.31 pp	86.15% ± 0.33 pp	+3.46 pp	DINOv3 gave the stronger class-balanced overall score.
Weighted F1	89.64% ± 0.27 pp	89.70% ± 0.36 pp	+0.06 pp	Weighted performance was effectively unchanged.

Replicated model comparison

The replicated comparison therefore shows a trade-off rather than a simple overall replacement of one backbone by the other. EfficientNetV2B2 remains the slightly stronger global-accuracy model, whereas DINOv3 with class-balanced sampling is the stronger class-balanced model. Weighted F1 is effectively tied, indicating that the main difference is how performance is distributed across species rather than a large change in overall weighted performance.

Class-level explanation of the replicated comparison

The difference between EfficientNetV2B2 and DINOv3 with class-balanced sampling was mainly a redistribution of recall across species rather than a uniform increase in accuracy. The DINOv3 condition improved several species that were weak under the EfficientNetV2B2 baseline, especially davidis, kolaceki, goodwini, and gracilis.

The largest recall gain was observed for davidis, which increased by 23.96 percentage points. Recall also increased strongly for kolaceki (+20.00 percentage points), goodwini (+14.81 percentage points), and gracilis (+12.50 percentage points). The gains for goodwini, gracilis, and kolaceki should be interpreted cautiously because these classes have very small validation support.

These gains were offset by losses in higher-support or previously strong classes. Recall for major decreased by 7.30 percentage points, while harpa and costata decreased by 2.24 and 2.08 percentage points, respectively. Because major and harpa are large validation classes, these losses reduce the chance of a global accuracy gain, even though the macro-level metrics improve.

Species	Support	EfficientNetV2B2 recall	DINOv3 class-balanced recall	Recall difference	Diagnostic meaning
davidis	64	32.81%	56.77%	+23.96 pp	Largest DINOv3 gain; this was one of the weakest baseline classes.
kolaceki	10	36.67%	56.67%	+20.00 pp	Strong gain, but support is very small.
goodwini	9	85.19%	100.00%	+14.81 pp	Large apparent gain, but support is very small.
gracilis	8	37.50%	50.00%	+12.50 pp	Improved relative to the baseline, but remains weak in absolute recall.
major	338	88.26%	80.97%	−7.30 pp	Main DINOv3 loss; important because this is a high-support class.
harpa	328	94.72%	92.48%	−2.24 pp	Small loss, but relevant because this is also a high-support class.
costata	48	100.00%	97.92%	−2.08 pp	Small loss from a previously perfect baseline recall.

Recall redistribution under DINOv3 class-balanced sampling

This class-level redistribution explains the metric pattern reported in the replicated comparison. DINOv3 with class-balanced sampling improves macro recall and macro F1 because it recovers more examples from weak or low-support species. It does not improve global accuracy because the gains in smaller classes are offset by recall losses in larger or previously strong classes, especially major.

Very-strict curated DINOv3 result

As a side result, DINOv3 was also evaluated after applying the very-strict statistical label-curation profile that had been selected in the previous EfficientNetV2B2 audit analysis. This intervention was deliberately small: only 25 of 14,348 training files were removed before training. The purpose of this run was not to replace the replicated DINOv3 comparison, but to test whether the same conservative curation profile also showed a directional benefit when used with the DINOv3 backbone.

The very-strict curated DINOv3 run reached 90.64% accuracy, 86.53% macro recall, 87.66% macro F1, and 90.80% weighted F1. These are the strongest single-run DINOv3 values in the current result set. However, this condition is represented by only one run, whereas the main EfficientNetV2B2 and DINOv3 comparisons are based on replicated three-run cohorts.

The result therefore provides directional evidence that the very-strict curation profile may also improve DINOv3 performance. It should not yet be presented as a confirmed curation effect or as a replacement for the replicated DINOv3 class-balanced result.

Condition	Training-data condition	Replication level	Accuracy	Macro recall	Macro F1	Weighted F1	Interpretation
DINOv3 class-balanced sampling	Uncurated leakage-aware Harpa dataset	Three independent runs	89.53% ± 0.26 pp	84.76% ± 0.97 pp	86.15% ± 0.33 pp	89.70% ± 0.36 pp	Main replicated DINOv3 reference condition.
DINOv3 after very-strict curation	Very-strict statistically curated Harpa dataset	Single run	90.64%	86.53%	87.66%	90.80%	Strongest single DINOv3 result so far, but not yet replicated.

Side comparison with the replicated DINOv3 reference

Common-subset comparability check

The very-strict DINOv3 result was also checked on the subset of explicit evaluation records that was shared with the previous DINOv3 class-balanced cohort. Although the validation manifest itself was unchanged, the merged explicit DINOv3 evaluation records were not perfectly identical between the very-strict run and the earlier DINOv3 class-balanced runs. The shared overlap contained 1,274 of 1,602 evaluated records.

On this common subset, the very-strict DINOv3 run remained higher in accuracy and weighted F1. However, it did not remain higher in macro F1 relative to the previous DINOv3 class-balanced mean. This check therefore supports a cautious interpretation: the very-strict result is promising, but the apparent macro-F1 improvement from the full evaluation should not yet be treated as a stable curation effect.

Condition	Accuracy	Macro recall	Macro F1	Weighted F1	Common-subset interpretation
Previous DINOv3 class-balanced mean	90.66% ± 0.16 pp	87.20% ± 0.31 pp	88.04% ± 0.70 pp	90.69% ± 0.17 pp	Higher common-subset macro recall and macro F1.
Very-strict DINOv3 single run	90.97%	86.53%	87.30%	91.06%	Higher accuracy and weighted F1, but not higher macro F1.
Difference: very-strict minus previous mean	+0.31 pp	−0.67 pp	−0.74 pp	+0.37 pp	Accuracy and weighted-F1 gains persist, but the macro-F1 gain does not.

Common-subset metric differences

Discussion

DINOv3 changes the performance distribution rather than replacing EfficientNetV2B2

The replicated comparison does not support a simple replacement narrative in which one architecture is uniformly superior to the other. Instead, the results show a redistribution of performance across the Harpa species classes. EfficientNetV2B2 remained slightly stronger in mean top-1 accuracy, reaching 89.99% ± 0.18 percentage points, compared with 89.53% ± 0.26 percentage points for DINOv3 with class-balanced sampling. This difference was small, but it indicates that DINOv3 did not improve the global accuracy of the replicated model cohort.

The opposite pattern was observed for class-balanced metrics. DINOv3 increased macro recall from 79.73% ± 0.55 to 84.76% ± 0.97, and macro F1 from 82.69% ± 1.31 to 86.15% ± 0.33. Weighted F1 was effectively unchanged between the two model families. The main effect of DINOv3 was therefore not a general increase in correct predictions, but a change in how performance was distributed across species.

This result is consistent with the architectural distinction between the two backbones. EfficientNetV2 is a supervised convolutional architecture optimized for efficient end-to-end image classification, using training-aware architecture search, Fused-MBConv blocks, MBConv blocks, and progressive learning. DINOv3 follows a different design logic: it is a self-supervised vision foundation model intended to learn transferable visual representations, with particular emphasis on reusable dense and semantic features. These different optimization targets make a direct winner-takes-all interpretation inappropriate.

For the present Harpa task, DINOv3 is therefore better only under a class-balanced definition of performance. It improved macro recall and macro F1, which are important when rare or difficult species should contribute equally to the evaluation. However, EfficientNetV2B2 remained at least as practical when the objective was raw top-1 accuracy and stable performance on the dominant validation distribution. The most defensible interpretation is therefore that DINOv3 changes the operating point of the classifier: it shifts performance toward more even species recovery, but it does not simply replace EfficientNetV2B2 as the globally stronger classifier.

Macro metrics reveal rare-species recovery hidden by global accuracy

The comparison between EfficientNetV2B2 and DINOv3 shows why global accuracy alone is not sufficient for evaluating fine-grained shell identification. A high top-1 accuracy can indicate that the model performs well on the dominant or visually easier species, but it can still hide weak recovery of rare, small-support, or diagnostically difficult classes. For a taxonomic classifier, this distinction is important because the scientific value of the model does not depend only on how often it recognizes the most common species.

In the EfficientNetV2B2 baseline, the separation between global performance and macro-level performance indicates that classification success was unevenly distributed across the 13 Harpa classes. This does not make the baseline weak; rather, it shows that the baseline was strongest where the dataset gave it the most favourable conditions: larger species classes, visually clearer classes, and classes with more stable representation in the training data. The weaker species therefore become the more informative test of whether an alternative backbone changes the biological usefulness of the classifier.

DINOv3 with class-balanced sampling appears to shift part of the classifier capacity toward those weaker classes. The improvement of species such as H. davidis, H. kolaceki, H. goodwini, and H. gracilis suggests that the DINOv3 condition did not merely reproduce the majority-class behaviour of EfficientNetV2B2. Instead, it recovered more examples from classes that were previously under-resolved. This is the main reason why macro recall and macro F1 are central to the interpretation of the experiment.

This gain has to be interpreted together with the losses in larger or previously stronger classes, especially H. major. The DINOv3 result therefore reflects a redistribution of recall rather than an unqualified improvement. From an operational perspective, reduced recall in a high-support class can affect many more user predictions than a gain in a rare class. From a scientific perspective, however, improved recovery of rare or difficult species may be more important than a small change in the dominant validation classes.

The biological implication is that shell identification should not be treated as a majority-class recognition task. A model that performs well on common species such as H. harpa or H. major, but fails to recover rare or difficult species, may still be useful in a public identification workflow, but it is less complete as a taxonomic classification system. Macro metrics are therefore not secondary diagnostics in this study; they are part of the definition of model quality.

Relationship to self-supervised representation literature

The present result fits within a broader pattern reported for self-supervised visual representations. DINOv3 is not primarily designed as a conventional supervised classifier, but as a representation backbone whose features can be reused across downstream tasks [10]. This distinction matters for interpreting the Harpa experiment: the relevant question is not only whether DINOv3 produces a higher top-1 accuracy, but whether its representations change how difficult or low-support visual categories are recovered.

A useful conceptual comparison is the WikiArt benchmark, where supervised EfficientNetV2 was compared with DINO-family self-supervised representations for artistic style and genre classification [6]. In that setting, EfficientNetV2 remained a strong supervised classifier, but DINO-based linear-probe and nearest-neighbour representations were competitive and, for some classification targets, stronger than the supervised EfficientNetV2 baseline [6]. The important point is not that artwork classification is directly equivalent to shell taxonomy, but that both tasks depend on subtle visual structure rather than only on broad object identity.

This is relevant for seashell identification because diagnostic information is distributed across multiple visual scales. Species-level shell recognition may depend on global outline, aperture shape, sculpture, colour pattern, ribbing, growth stage, and viewpoint-dependent characters. These cues are not always captured equally well by a classifier optimized mainly for the dominant closed-set decision boundary. A self-supervised representation may retain visual information that is useful for rare or ambiguous classes, even when this does not translate into a higher global accuracy [10, 6].

The Harpa results are therefore consistent with literature in which self-supervised features improve the distribution of performance in visually subtle tasks. DINOv3 did not simply dominate EfficientNetV2B2 as a classifier, but it improved the class-balanced behaviour of the model. This supports the interpretation that its value lies less in replacing a supervised CNN for all cases, and more in providing a different representation space that can improve recovery of difficult species.

This interpretation should remain cautious. EfficientNetV2 and DINOv3 are usually evaluated under different adaptation protocols, including full fine-tuning, frozen linear probing, and nearest-neighbour evaluation [4, 10, 6]. The present study is therefore best viewed as a domain-specific contribution: it shows that, under a controlled leakage-aware shell-classification protocol, the expected advantage of self-supervised representation learning appears mainly as improved macro-level species recovery, not as an unconditional gain in top-1 accuracy.

Architecture versus training-policy effects

The DINOv3 result should not be interpreted as an architecture-only effect. The main replicated comparison evaluates a DINOv3 backbone after a specific downstream adaptation choice: class-balanced sampling. This matters because class imbalance is not a neutral property of the dataset. Re-sampling and re-weighting strategies are established responses to long-tailed class distributions, and they can change which classes contribute most strongly to the fitted decision boundary [7]. Similarly, focal loss was introduced to change the training emphasis by reducing the contribution of easy examples and focusing optimization on harder examples in imbalanced dense detection [8].

The DINOv3 imbalance-strategy screen in the present study showed this interaction directly. The no-balancing and focal-loss-only settings produced the strongest accuracy-oriented single-run behaviour, whereas class-balanced sampling produced the strongest macro-F1 behaviour on the full evaluated set. Therefore, the macro-metric advantage observed for DINOv3 should not be attributed solely to the self-supervised backbone. It is more accurate to state that DINOv3, when combined with the selected class-balanced sampling policy, shifted the classifier toward stronger class-balanced recovery.

This interpretation is also consistent with how both architecture families were originally designed. EfficientNetV2 was not proposed merely as a static CNN topology, but as a combination of training-aware neural architecture search, Fused-MBConv design, compound scaling, and progressive learning with adaptive regularization [4]. DINOv3 is likewise best understood as a representation learning pipeline rather than only as a transformer backbone: its reported performance depends on large-scale self-supervised pretraining, DINO/iBOT/KoLeo objectives, Gram anchoring, high-resolution adaptation, and distillation into deployable model variants [10].

The practical consequence is that the comparison in this study is between complete training systems, not isolated network diagrams. The EfficientNetV2B2 baseline represents a supervised CNN classifier trained with its own loss, image size, optimizer, and regularization choices. The selected DINOv3 condition represents a self-supervised backbone adapted with categorical cross-entropy and class-balanced sampling. The observed difference in macro recall and macro F1 therefore reflects an interaction between architecture, pretrained representation, and imbalance policy.

This also limits the generality of the conclusion. The study does not show that every DINOv3 training policy improves class-balanced performance, nor that DINOv3 automatically outperforms EfficientNetV2B2 on the Harpa task. It shows that the selected DINOv3 adaptation produced a more class-balanced operating point than the EfficientNetV2B2 baseline, while other DINOv3 imbalance settings were more accuracy-oriented. The architecture and the training policy therefore have to be interpreted together.

Species-level gains and losses: weak-species recovery versus high-support class erosion

The species-level results clarify why the DINOv3 comparison should be interpreted as a redistribution of performance rather than as a uniform improvement. The DINOv3 class-balanced condition improved several species that were weak or low-support under the EfficientNetV2B2 baseline, including H. davidis, H. kolaceki, H. goodwini, and H. gracilis. At the same time, this gain was partly offset by reduced recall in larger or previously stronger classes, most notably H. major. This pattern is important because a model can become more useful for difficult species while becoming less favourable for the most frequent validation classes.

A plausible explanation is that the two model families emphasize different representational properties. EfficientNetV2 was designed as a supervised convolutional architecture for efficient end-to-end image classification, using training-aware neural architecture search, Fused-MBConv blocks, compound scaling, and progressive learning with adaptive regularization [4]. DINOv3 was designed as a self-supervised vision foundation model, with large-scale pretraining, DINO/iBOT/KoLeo objectives, Gram anchoring, high-resolution adaptation, and distillation to smaller model variants [10]. In the DINOv3 report, Gram anchoring is specifically introduced to preserve dense feature quality during long self-supervised training [10].

This representational difference is relevant to Harpa classification because shell recognition is a fine-grained visual task. In the previous Harpa study, the genus was described as requiring species-level recognition from subtle combinations of shell outline, sculpture, colour pattern, aperture structure, varices, wear state, and viewpoint-dependent morphology [3]. Conchological treatments of Harpidae likewise rely on shell form, ribbing or varices, colour pattern, and related shell characters for species-level distinction [9]. A representation that preserves both local texture-like evidence and broader shell geometry could therefore be advantageous for classes whose diagnostic signal is weak, variable, or underrepresented in the training set.

The present results are consistent with that expectation, but they do not prove the mechanism. The improvement of weak species under DINOv3 with class-balanced sampling is compatible with the idea that self-supervised transformer features retain useful local and global morphology [10]. However, the observed gains could also reflect the class-balanced training policy, stochastic variation, or interactions between species support and the classifier head. Conversely, the reduced recall for H. major suggests that improving rare-species recovery can move the decision boundary away from a high-support class that was better served by the EfficientNetV2B2 baseline.

The biological interpretation should therefore remain conservative. DINOv3 appears to improve the recovery of some weak or low-support species, but this comes with a cost in at least one high-support class. For a scientific identification system, that trade-off may be acceptable when the priority is more even taxonomic coverage. For an operational public classifier, the loss in a frequent species may be more consequential because it affects a larger share of expected predictions. Additional embedding-space analysis, nearest-neighbour inspection, and feature-visualization work would be needed before attributing the weak-species gains directly to specific DINOv3 feature properties rather than to the combined architecture-and-training protocol.

Comparison with task-specific DINO/EfficientNet literature

The present Harpa results are consistent with a broader pattern in which the relative value of DINO-style self-supervised representations and EfficientNet-style supervised CNNs depends on the target domain, adaptation protocol, and evaluation objective. EfficientNetV2 was developed as a supervised convolutional architecture for fast and parameter-efficient image classification, with training-aware neural architecture search, Fused-MBConv blocks, model scaling, and progressive learning with adaptive regularization [4]. DINOv3, in contrast, was developed as a self-supervised vision foundation model intended to produce transferable visual representations, including dense features that can be reused across multiple downstream vision tasks [10]. Because these model families were optimized for different purposes, task-specific comparisons should not be reduced to a single architecture ranking.

A useful counterpoint to a DINO-favourable interpretation is provided by the fetal-ultrasound benchmark of Barrientos, Du, and Dvornek [11]. That study compared ultrasound-specific foundation models with natural-image-pretrained baselines, including ResNet50, EfficientNetV2, and DINOv3, using both frozen linear probing and full fine-tuning [11]. The strongest model depended on the adaptation protocol: FetalCLIP performed best under linear probing, whereas USFM performed best under full fine-tuning [11]. This supports the interpretation that architecture choice is conditional on the domain and training protocol, rather than being determined solely by whether the backbone is a CNN or a self-supervised transformer.

The same caution applies to the present shell-classification comparison. DINOv3 did not dominate EfficientNetV2B2 in global top-1 accuracy, but it improved the class-balanced behaviour of the model. This places the Harpa result between two extremes: it does not show that DINOv3 is universally superior for closed-set species classification, but it also does not support treating EfficientNetV2B2 accuracy as the only meaningful outcome. Instead, the result is compatible with task-specific literature in which different backbones become preferable under different definitions of success, such as accuracy, macro-level recovery, transferability, or cross-domain generalization [4, 10, 11].

DINOv3 is often most compelling in tasks that reward transferable dense visual features rather than only closed-set image-level classification. The DINOv3 report emphasizes dense feature quality, transfer across downstream tasks, and the use of Gram anchoring to preserve dense representations during long self-supervised training [10]. An indirect example is the aerial roof-structure study of Versteeg, Wijnhoven, and Oswald, which used a DINOv3 ConvNeXt-Base backbone for joint instance segmentation and geometric attribute regression from aerial imagery [12]. That task required instance-level roof masks together with geometric predictions for building height, roof slope, and roof azimuth, which is a different problem type from closed-set species classification [12].

This dense-feature literature is relevant to the Harpa findings, but only indirectly. Shell identification may also depend on distributed visual evidence, including global form and local surface pattern, but the present task remains a 13-class species-classification problem rather than segmentation, correspondence, or geometry regression. The observed DINOv3 benefit should therefore be interpreted as a smaller and more conditional effect: transferable representations may help recover weak or visually difficult species, but that benefit does not necessarily translate into a higher global accuracy for the whole validation distribution.

The species-level trade-off is central to the practical interpretation. In the present results, DINOv3 with class-balanced sampling improved weak or low-support species such as H. davidis, H. kolaceki, H. goodwini, and H. gracilis, but this was offset by reduced recall in higher-support classes, especially H. major. For a scientific classifier, improved recovery of difficult species may be valuable because it gives a more balanced taxonomic model. For a public identification system, however, reduced performance in common species can affect a larger number of expected user predictions. The choice between EfficientNetV2B2 and DINOv3 is therefore not only an architectural choice; it is also a choice between majority-case reliability and more balanced recovery of rare or difficult species.

Effect of very-strict label curation on DINOv3

The very-strict DINOv3 follow-up should be interpreted as a training-data curation result, not as a new replicated architecture comparison. The curation profile was inherited from the previous leakage-aware Harpa audit study, where conservative statistical evidence was used to exclude only the most strongly supported label-inconsistency candidates from the training set [3]. This type of intervention is methodologically related to confident-learning approaches, in which model-predicted probabilities and observed labels are used to estimate label uncertainty and identify candidate label errors [13].

In the present DINOv3 follow-up, the intervention was deliberately small. Only 25 of the original 14,348 training records were removed, leaving 14,323 natural training files before class-balanced sampling. The validation manifest was not changed, and the DINOv3 training configuration remained the same as in the selected class-balanced condition. The resulting run produced the strongest single DINOv3 result in the present experiment, with higher accuracy, macro recall, macro F1, and weighted F1 than the replicated uncurated DINOv3 mean.

This result is encouraging, but it should not be treated as definitive evidence that very-strict curation improves DINOv3. The curated condition was represented by a single run, whereas the main EfficientNetV2B2 and DINOv3 comparisons were based on replicated three-run cohorts. Because DINOv3 training includes stochastic initialization, class-balanced sampling, early stopping, and fine-tuning of a pretrained backbone, a single improved run cannot separate a true curation effect from ordinary run-to-run variation.

The common-subset check makes the interpretation more cautious. Although the validation manifest was unchanged, the explicit merged evaluation records were not perfectly identical between the very-strict run and the previous DINOv3 class-balanced cohort. On the shared subset of evaluated records, the very-strict run remained higher in accuracy and weighted F1, but it did not remain higher in macro F1 relative to the previous DINOv3 class-balanced mean. Therefore, the full-evaluation macro-F1 gain should not yet be interpreted as a stable consequence of label curation.

The most defensible interpretation is that conservative label curation may also be beneficial for DINOv3, but the present evidence is directional rather than conclusive. The result is consistent with the general rationale that removing high-confidence label inconsistencies can improve the training signal [13], and it is also consistent with the earlier Harpa audit result obtained with EfficientNetV2B2 [3]. However, confirmation requires a replicated DINOv3 very-strict cohort, ideally evaluated on the same explicit record set as the replicated uncurated DINOv3 cohort. Until that replication is available, the very-strict DINOv3 result should be presented as promising follow-up evidence, not as a confirmed curation effect.

Operational implications for seashell identification

The present results have different implications for a production identification system and for a research-oriented taxonomic workflow. For production-oriented top-1 classification, EfficientNetV2B2 remains the safer default in the current Harpa pipeline. It produced the slightly stronger replicated global accuracy, showed low between-run variation, and is already the established reference model in the preceding leakage-aware Harpa workflow [3]. This practical conclusion is also consistent with the original EfficientNetV2 design goal: the architecture was developed as a supervised convolutional model family optimized for fast training, parameter efficiency, and strong image classification performance [4].

DINOv3 is more attractive when the operational objective is not limited to the most probable species label. In the present study, its value was most visible in class-balanced behaviour and weak-species recovery rather than in global top-1 accuracy. This is consistent with the purpose of DINOv3 as a self-supervised vision foundation model intended to provide transferable visual representations and high-quality dense features across downstream vision tasks [10]. For seashell identification, this makes DINOv3 especially relevant for secondary evidence: difficult species-pair analysis, rare-species inspection, embedding-space exploration, and cases where the classifier output alone is insufficient.

A hybrid deployment strategy is therefore scientifically and operationally defensible. EfficientNetV2B2 can remain the main closed-set classifier when the priority is stable top-1 prediction on the dominant validation distribution, while DINOv3 can be used as a complementary representation backbone for cases requiring additional evidence. In such a workflow, DINOv3 embeddings could support nearest-neighbour comparison, retrieval of visually similar training examples, disagreement triage, or review of difficult species pairs. The general use of nearest-neighbour evidence is supported by the classical nearest-neighbour classification literature [14], while DINOv3 provides a modern self-supervised feature space in which such comparisons may be more informative for visually subtle images [10].

This hybrid role also fits the label-audit context of the project. The previous Harpa study used statistical evidence to identify conservative label-inconsistency candidates [3], and confident-learning methodology provides a general framework for estimating label uncertainty and identifying probable label errors from model predictions and observed labels [13]. DINOv3 should therefore not be viewed only as an alternative classifier head. Its embeddings may provide independent evidence for future label-audit decisions, especially when EfficientNetV2B2 and DINOv3 disagree or when a specimen lies near a persistent species-pair boundary.

The immediate operational recommendation is therefore not to replace EfficientNetV2B2 outright. A conservative deployment path would keep EfficientNetV2B2 as the default production classifier and use DINOv3 as an auxiliary representation model. DINOv3 predictions or embeddings could be surfaced when the EfficientNetV2B2 prediction confidence is low, when the predicted class is rare, when the top alternatives form a known difficult species pair, or when nearest-neighbour evidence is needed for audit or interpretation. This preserves the stability of the established supervised classifier while exploiting the main advantage observed for DINOv3: improved recovery of weak classes and a representation space that is useful beyond a single softmax prediction.

Limitations

This study is a controlled comparison within a single genus-level classification problem, not a general benchmark of EfficientNetV2B2 and DINOv3 across biological image recognition. The results therefore apply most directly to the present leakage-aware Harpa dataset and should not be generalized without additional testing on other shell genera, other image sources, and other taxonomic boundaries. This limitation is important because image-based species identification is known to depend on the taxonomic scope, training-set composition, image quality, and sampling of visual variation [1, 2].

A second limitation is that the DINOv3 comparison evaluates one selected adaptation strategy rather than the full space of possible DINOv3 training protocols. The selected DINOv3 condition used partial fine-tuning and class-balanced sampling, and the ablation showed that alternative imbalance strategies produced different operating points. The comparison therefore tests DINOv3 under the best-performing class-balanced protocol identified in this study, not DINOv3 as an architecture independent of downstream training policy. This matters because DINOv3 was developed as a self-supervised foundation model whose downstream performance depends on how its learned representation is adapted to a target task [10].

The evaluation was also limited by the use of explicit merged evaluation records rather than a perfectly identical prediction table for every comparison. The validation manifest was kept unchanged, but the very-strict DINOv3 run and the previous DINOv3 class-balanced cohort did not have identical explicit evaluated records. The common-subset analysis partly addressed this issue, but it also showed why the curated result should not be overinterpreted. Future comparisons should ensure that all replicated conditions are evaluated on the same explicit record set whenever possible.

Finally, the present paper evaluates DINOv3 mainly as a closed-set classifier. This does not fully exploit the representation-learning rationale of DINOv3. The results suggest that DINOv3 may be useful for weak-species recovery, but this study did not yet analyze the embedding space, nearest-neighbour structure, species-pair geometry, or disagreement cases in detail. Because DINOv3 is designed to provide transferable and dense visual representations [10], a complete assessment of its value for shell identification will require analyses beyond top-1 classification metrics.

Future work

A first priority is to evaluate DINOv3 embeddings directly. The present results show that DINOv3 changes class-balanced behaviour, but they do not show whether this is caused by a more informative visual representation of shell morphology. Embedding-space analysis could test whether weak species, difficult species pairs, and suspected label inconsistencies occupy meaningful positions in DINOv3 feature space. Nearest-neighbour retrieval should be especially useful here, because nearest-neighbour methods provide a direct way to inspect similarity structure around individual examples [14].

A second direction is to use DINOv3 as auxiliary evidence in the existing label-audit workflow. The previous Harpa audit used conservative statistical evidence to identify likely label inconsistencies [3], and confident learning provides a general framework for estimating label uncertainty from model predictions and observed labels [13]. DINOv3 embeddings could add an independent representation-based signal to this process, especially for cases where EfficientNetV2B2 and DINOv3 disagree, where a prediction falls near a recurrent species-pair boundary, or where the nearest neighbours are visually inconsistent with the assigned label.

A third direction is to test a hybrid classifier-and-retrieval workflow. EfficientNetV2B2 can remain the default classifier for stable top-1 predictions, while DINOv3 can provide secondary evidence for rare species, low-confidence predictions, and difficult alternatives. This would match the empirical pattern observed here: EfficientNetV2B2 remains the safer production classifier for global accuracy, whereas DINOv3 is more informative when the task requires class-balanced recovery or representation-level inspection.

Future work should also extend the comparison beyond Harpa. Testing the same EfficientNetV2B2 and DINOv3 workflow on other genera would show whether the observed pattern is specific to this genus or reflects a broader property of shell classification. This is important because automated species identification systems are affected by taxonomic scope, class imbalance, rare-species sampling, and image-domain variation [1, 2]. A multi-genus evaluation would therefore provide stronger evidence for deciding when DINOv3 should be used as a classifier, when it should be used as an embedding model, and when EfficientNetV2B2 remains the more appropriate operational default.