Reliability-calibrated contrastive learning improves plant leaf recognition

A new training method called ContrastPlant improved recognition of plant leaves across three public benchmarks and a fixed laboratory-to-field test, according to a 2026 research article. The approach combines supervised contrastive learning with prototype-reliability calibration, or PRC, a rule designed to prevent visually similar but unreliable stored examples from dominating model training. On the Flavia dataset, the method achieved 96.03% ± 0.31% accuracy, compared with 93.31% ± 0.16% for standard cross-entropy training. The largest practical gain appeared in the more demanding LeafSnap evaluation, where fixed laboratory-to-field accuracy rose from 71.49% ± 0.62% to 75.87% ± 0.67% without using field images for training or model selection.

Plant leaf identification is a fine-grained visual classification problem: individual leaves of the same species can look different because of lighting, viewpoint or growth stage, while leaves from different species may share nearly identical outlines, textures and vein patterns. The study examined category recognition only and did not test disease diagnosis, symptom localization, severity estimation or broader crop monitoring. Its central concern is not merely finding the most visually similar negative examples, as conventional hard-negative mining does, but determining whether those examples are dependable representatives of the species to which they have been assigned. This distinction matters when a memory bank contains a mislabeled, atypical or outdated leaf embedding.

ContrastPlant uses a shared encoder to produce both a conventional class prediction and a contrastive embedding. During training, that embedding queries a bounded memory divided into separate partitions for each leaf category. Same-class entries provide a positive relation, while embeddings stored under other labels form the candidate negative set. The system can then retain the 100 most similar out-of-class examples for each query, using cosine similarity to find visually close leaves. Because retrieval is restricted by label, correctly annotated examples of the query species are not treated as negatives. A separate exponential moving average, or EMA, maintains a class-level anchor that supplies a more stable positive representation than relying on one randomly selected image.

PRC then assigns every selected negative a continuous weight based on two quantities. The first is class consistency: it measures the margin between the similarity of a stored example to its recorded class anchor and its similarity to the nearest competing class anchor. A low consistency score suggests that the example does not align well with its labeled category. The second quantity is class ambiguity, calculated from the similarity between the query class anchor and the negative class anchor. Classes that appear particularly close to the query class receive greater relevance. Their product determines how strongly each negative contributes to the contrastive loss, with weights normalized separately for each query so that the average selected-negative weight remains approximately one.

The separation between difficulty and reliability has a direct mathematical effect on optimization. Similarity still controls how strongly a negative affects the model through the exponential term in the contrastive objective, preserving the value of hard examples. PRC changes the relative gradient assigned to those examples: reliable negatives from visually confusing classes can receive more influence, while prototype-inconsistent entries receive less. The weights are detached during back-propagation, meaning they act as training-time estimates rather than an additional learnable network. The authors also derived a variance-reduction result for the EMA state under assumptions of independent observations, but described this as a mechanism-level characterization rather than a convergence or generalization guarantee.

Across the three within-dataset benchmarks, PRC produced mean accuracy gains of 2.72 percentage points on Flavia, 1.58 points on Swedish Leaf and 2.43 points on Folio relative to cross-entropy. The paired 95% confidence intervals for these differences were above zero, and macro-F1 followed the same pattern. On Flavia, ContrastPlant also exceeded SupCon by 1.18 points, hard-negative contrastive learning by 0.77 points and a similarity-only Top-K method with the same EMA anchor by 0.35 points. That last comparison is particularly informative because it holds the memory, negative count, positive anchor, loss schedule and optimization protocol constant while assigning every selected negative a weight of one.

Ablation experiments indicated that both proposed factors contributed positively on average, although the statistical evidence differed. The complete method beat a reliability-only version by 0.35 percentage points and an ambiguity-only version by 0.42 points. After Holm correction, however, the advantage over the ambiguity-only variant remained significant at p = 0.01165, whereas the advantage over the reliability-only variant did not, at p = 0.06822. Replacing the EMA class state with an ordinary memory centroid produced a smaller mean difference of 0.14 points, but its confidence interval included zero. In a controlled test that randomly corrupted 10% of Flavia training labels, PRC reduced the accuracy loss from 2.09 to 1.28 points relative to cross-entropy and lowered the share of absolute negative-gradient mass assigned to injected noisy entries from 18.9% to 8.9%. The authors cautioned that this did not establish robustness to arbitrary or unknown forms of label corruption.

The fixed LeafSnap protocol provided the strongest test of acquisition shift. Models were trained and selected using 23,027 laboratory images from 184 species, while all 7,719 retained field images were reserved for final testing and excluded from training, validation, memory construction, class-anchor updates and hyperparameter selection. ContrastPlant reached 75.87% ± 0.67% field accuracy, outperforming cross-entropy by 4.38 points, SupCon by 2.77 points and Top-K with EMA by 1.97 points. The study also tested a Swin-Tiny transformer and an EfficientFormer-L1 architecture on Flavia, where PRC added 0.70 and 0.81 percentage points, respectively, without changing either model’s inference parameters or GFLOPs. These results support compatibility with the tested architectures rather than universal independence from the choice of backbone.

Analysis of the learned representation space showed that ContrastPlant produced the smallest mean intra-class cosine distance, the largest distance to the nearest other-class centroid and the highest cosine silhouette score on Flavia. Its advantage was also associated with fewer pooled cross-entropy confusions among five representative species pairs, with descriptive reductions ranging from 60% to 80%. Nevertheless, field failures persisted under cluttered backgrounds, strong shadow, blur, partial occlusion, unusual viewpoints and low resolution. The study also identified important limitations: Flavia, Swedish Leaf and Folio are relatively small and clean datasets, image-level splits could not exclude near-duplicate or same-specimen leakage, and the Folio test set contained only 96 images, making each additional correct prediction worth roughly one percentage point. The method has not been evaluated at very large taxonomic scale.

Reliability calibration adds cost during training but disappears when the classifier is exported. On the reported V100 platform, epoch time was 1.55 times that of cross-entropy and peak training memory increased from 2,152 to 3,165 MB. At inference, the memory, projection head, retrieval process, class anchors and PRC weights are all removed, leaving the same ResNet18-and-classifier graph used by the cross-entropy model. Measurements on an 8 GB Jetson Orin Nano using FP16 TensorRT inference found near-identical latency, throughput, memory use and energy consumption for the two exported models, although the authors said this was not a formal equivalence test. The result offers a practical way to improve leaf-category recognition without enlarging the deployed network, but the authors said future work must address anchor contamination, stale memory entries, scalable retrieval and more diverse field conditions.

Subject of Research: Reliability-calibrated supervised contrastive learning for fine-grained plant leaf recognition

Article Title: Contrastive learning with prototype reliability calibration for plant leaf recognition

Article References: Zeng, X., Cai, W., Lin, J., Yang, W., & Gao, Y. (2026). Contrastive learning with prototype reliability calibration for plant leaf recognition. Discover Artificial Intelligence, 6(1), Article 1151. https://doi.org/10.1007/s44163-026-02196-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02196-x

Keywords: plant leaf recognition, contrastive learning, prototype calibration, hard-negative mining, fine-grained classification, machine learning, LeafSnap, edge deployment, Contrastive, learning, prototype, reliability