Uncertainty-Aware Semi-Supervised Framework with Supervised Contrastive Learning for Sugarcane Leaf Disease Detection

1. Introduction

• Sugarcane disease symptoms vary within classes and overlap visually across diseases and healthy tissue, while field conditions intensify these challenges.Limited feature diversity and overfitting risk can restrict practical use of deep learning models.Source: §1 · p. 3
• Expert annotation is costly and time-consuming, motivating semi-supervised learning that uses unlabeled images alongside limited labeled samples.The introduction frames robustness and uncertainty handling as remaining research problems.Source: §1 · p. 3
• The proposed Mean-Teacher approach uses student–teacher consistency and unlabeled data, with sugarcane disease datasets serving as a representative case study.The method also applies wavelet preprocessing, KL-divergence pseudo-label correction, and supervised contrastive learning.Source: §1 · p. 3
• Discrete wavelet transform is introduced to enhance visual-feature semantic context and improve the model’s representation ability.The study’s method description says wavelet processing focuses on disease patterns.Source: §1 · p. 3
• Supervised contrastive learning is used across leaf-disease classes to improve class separability and classification performance.The approach targets intra-class variation and inter-class similarity in the sugarcane case study.Source: §1 · p. 3
• KL-divergence-based pseudo-label correction is proposed to reduce pseudo-label noise and improve model stability.The framework uses it to reduce confirmation bias and increase prediction accuracy.Source: §1 · p. 3

2.1. Traditional Methods

• Handcrafted color, shape, and texture features are sensitive to leaf position, background, lighting, and inter-class variability, limiting generalization and scalability.Traditional classifiers include SVMs, Naive Bayes, Decision Trees, and Random Forests.Source: §2.1 · pp. 3–4

2.2. Deep Learning for Plant Disease Detection

• CNNs provide strong feature learning, but performance depends on dataset size and diversity and can degrade under real-world field conditions.Prior work includes deeper architectures, transfer learning, and lightweight models for resource-constrained settings.Source: §2.2 · p. 4

2.3. Semi-Supervised Learning in Agriculture

• Agricultural SSL exploits unlabeled data, but pseudo-label errors can accumulate and consistency methods may assume reliable predictions without modeling uncertainty.Ambiguous disease classes and environmental variation make robust handling of noisy labels and uncertain samples an open problem.Source: §2.3 · p. 4

2.4. Mean-Teacher and Uncertainty-Aware Learning

• Mean-Teacher methods enforce student–teacher prediction consistency on unlabeled data, yet typically do not explicitly account for prediction uncertainty or distinguish unlabeled samples.Uncertainty-aware Mean-Teacher frameworks have rarely been explored in agricultural disease detection.Source: §2.4 · pp. 4–5

3. Problem Formulation

• The task uses both labeled and unlabeled data for semi-supervised image classification.The labeled dataset pairs each input image with a class label; a separate equation defines the unlabeled dataset.Source: §3 · p. 5
• The student is trained with supervised loss on labeled data and uncertainty-guided consistency loss to align student and teacher predictions on labeled and unlabeled data.The combined loss is intended to use unlabeled data and improve generalization under limited annotations.Source: §3 · p. 5

4. Methodology

• The framework’s submodules cover preprocessing and augmentation, supervised contrastive inter-class embeddings, and Mean-Teacher semi-supervised global representations.The framework overview also identifies KL-divergence-based uncertainty estimation for utilizing unlabeled data.Source: §4 · p. 5, Figure 1 · p. 6

4.1. Sugarcane Leaf Preprocessing and Augmentation:

• Images are resized and augmented, using flips and rotations for spatial variation and stronger color jittering, blur, and erasing for labeled data.These transformations are intended to increase data diversity and address varied orientations and environmental conditions.Source: §4.1 · p. 5

4.1.1. Color Jittering

• Color jittering modifies RGB brightness and contrast, then varies hue, saturation, and brightness in HSV space.The RGB transformation uses random contrast scaling and brightness adjustment; HSV parameters separately control hue shift and saturation and brightness scaling.Source: §4.1.1 · p. 5

4.1.2. Random Erasing

• Random Erasing replaces a randomly positioned image region with neutral gray to simulate occlusion while avoiding obscuring critical lesion information.The region covers 0.02–0.33 of image area, uses aspect ratio 0.3–3.3, and is applied with 50% probability.Source: §4.1.2 · p. 6

4.2. Supervised Contrastive Learning for Inter-Class Discriminative Embedding

• SCL addresses overlapping disease features by making same-class representations compact and different-class representations more distinct.The challenge is compounded by complex backgrounds, uneven lighting, and limited labeled data.Source: §4.2 · p. 6
• Each labeled sample is an anchor; other same-class samples in its mini-batch are positives, and different-class samples are negatives.The loss compares each anchor with the other labeled samples in the batch.Source: §4.2 · p. 7
• SCL uses a Siamese network with shared weights, while unlabeled samples are excluded from its loss and handled by Mean-Teacher consistency and pseudo-label correction.The loss uses L2-normalized student features and cosine similarity; temperature τ controls separation sharpness.Source: §4.2 · p. 7

4.3. Mean-Teacher based Semi-Supervised Learning for Global Feature Representation

• The framework combines feature enhancement, student–teacher consistency, pseudo-label rectification, and a unified loss balancing labeled and unlabeled objectives.The student–teacher consistency is enforced under input, and rectification refines noisy or uncertain teacher predictions before they supervise learning.Source: §4.3 · p. 7

4.3.1. Feature Enhancement

• DWT separates global leaf structure from directional detail bands, whose fine textures and edges are associated with disease symptoms.The components are flattened and concatenated; high-frequency bands respond to small spots and streaks, while the low-frequency band preserves morphology.Source: §4.3.1 · p. 8, Figure 2 · p. 8

4.3.2. Mean-Teacher Model

• The student learns from labeled and unlabeled images, while an EMA-updated teacher generates unlabeled pseudo-labels and guides the student's predictions.Both models use EfficientNetB4; the teacher is updated using an EMA of student weights and receives no gradient updates.Source: §4.3.2 · p. 9, Figure 4 · p. 10

4.3.3. Pseudo Label Rectification

• The consistency loss aligns student and teacher predictions on unlabeled samples using pointwise L2 difference and KL divergence between probability distributions.KL divergence estimates prediction uncertainty to guide consistency learning rather than acting as an independent loss term.Source: §4.3.3 · p. 10

4.3.4. Total Loss

• The student’s total objective combines labeled supervised classification, unlabeled Mean-Teacher consistency, and supervised contrastive loss for inter-class discrimination.Weights α and β control the contributions of consistency and contrastive losses, respectively.Source: §4.3.4 · pp. 10–11

5. Results

• The section covers the experimental setup, datasets, evaluation metrics, qualitative results, and ablation studies.Source: §5 · p. 11

5.1.1. SLD-1

• SLD-1 comprises 2,521 smartphone images from farms in Maharashtra, balanced across five classes and varied in resolution, lighting, and viewing angle.The classes are Healthy, Mosaic, Redrot, Rust, and Yellow.Source: §5.1.1 · p. 11

5.1.2. SLD-2

• SLD-2 contains 6,748 high-resolution images across 12 categories, collected from multiple angles and disease stages.It includes 10 disease types, a healthy class, and a dried-leaf class.Source: §5.1.2 · p. 12

5.1.3. SLD-3

• SLD-3 dataset contains 224 RGB images across Healthy, Redrot, and Redrust classes, all at a fixed original resolution of 224 × 224 pixels.Images were collected from sugarcane fields using the Google Teachable Machine platform.Source: §5.1.3 · p. 12

5.2. Experimental Setup

• Each dataset is split into 90% training data and a strictly unseen 10% test set reserved for final evaluation.The paper describes this fixed split as an assessment of generalization ability.Source: §5.2 · p. 12

5.2.1. Comparison Methods

• The comparison includes MAIAnet, SoyaTrans, ResTS, TSwin-F, ConvNext, and CorePlant.These methods span supervised and semi-supervised approaches.Source: §5.2.1 · p. 13

5.2.2. Evaluation Metrics for Leaf Disease Classification

• Classification performance is evaluated using Accuracy, Precision, Recall, and F1-score.The paper abbreviates these as Acc, Pre, Rec, and F1.Source: §5.2.2 · p. 13

5.3.1. Comparative Analysis with State-of-the-art Methods

• U-SSDNet leads supervised and semi-supervised methods across labeled-data ratios on all three datasets, reaching 97.83% accuracy and 97.42% F1-score on SLD-1, 95.07% accuracy and 94.73% F1-score on SLD-2, and 98.55% accuracy and 98.91% F1-score on SLD-3.The reported comparisons cover 10–50% labeled data on SLD-1 and 10–50% on SLD-2; the SLD-3 results are described as stable across labeled proportions.Source: §5.3.1 · pp. 13–14, Table 2 · p. 14, Table 3 · p. 14

5.3.2. Comparative Analysis with the Baseline Models

• At 50% labeled data on SLD-1, U-SSDNet’s F1-score is 97.42%, above ResNet-50 (93.67%), DenseNet-121 (93.22%), VGG-16 (70.14%), ViT (93.42%), and Swin Transformer (85.66%).On SLD-2, its F1-scores are 94.73%, 92.95%, and 89.88% at 50%, 20%, and 10% labeled data, respectively.Source: §5.3.2 · pp. 14–15, Table 5 · p. 15, Table 6 · p. 15

5.3.3. Abalation Study

• Wavelet enhancement performs best across accuracy, precision, recall, and F1; Deep plus Cut&Delete augmentation reports 94.80% accuracy and 94.20% F1-score.HOG and LPB trail Wavelet in accuracy by 3.52% and 2.28%, respectively; the combined augmentation reports 94.82% precision and 94.88% recall.Source: §5.3.3 · p. 16, Table 7 · p. 16, Table 8 · p. 16
• EfficientNet is the best-performing tested feature extractor, while the KL-Divergence uncertainty-aware method outperforms the no-uncertainty and MC-based alternatives.For EfficientNet, the reported metrics are 94.80% accuracy, 94.82% precision, 94.88% recall, and 94.20% F1-score.Source: §5.3.3 · pp. 16–17, Table 9 · p. 16, Table 10 · p. 17

5.3.4. Qualitative Analysis

• Qualitative predictions are mostly confident: SLD-1 and SLD-2 samples generally exceed 0.9 confidence, while SLD-2 examples mostly exceed 0.95.One SLD-1 Yellow leaf is misclassified as Redrot at 0.82 confidence; similar SLD-2 diseases retain confidence above 0.75.Source: §5.3.4 · p. 17

5.3.5. Qualitative Comparison with Existing Methods

• Grad-CAM comparisons show U-SSDNet focusing more consistently on disease-relevant regions than baselines across SLD-1, SLD-2, and SLD-3.Baseline maps are described as diffuse, extending beyond symptoms, or fragmented depending on the dataset.Source: §5.3.5 · pp. 17–18

5.3.6. Class-wise Performance Assessment

• The class-wise assessment uses confusion matrices and ROC curves for SLD-1 and SLD-2, with AUC as its main metric.The plots indicate correct classifications, errors, and class-level differences in discrimination.Source: §5.3.6 · p. 18

6.1. CAM Visualization

• Grad-CAM, Grad-CAM++, Score-CAM, and LIME maps generally focus on lesion regions; Grad-CAM++ localizes sharper boundaries, while Rust pustules are sometimes only partly highlighted.The highlighted regions align with visible lesion patterns, including subtle symptoms.Source: §6.1 · pp. 18, 21

6.2. t-SNE Visualization

• As the labeled-data proportion rises from 10% to 50%, t-SNE plots show increasingly compact, separated clusters, with clearer class boundaries at 30%–50%.At 10%, clusters begin to form; at 20%, class separation becomes more apparent.Source: §6.2 · p. 21, Figure 17 · p. 22

6.3. Limitation

• Generalization to other crops or environmental conditions may be limited; unlabeled-data noise and difficulty with disease levels and visually similar classes remain concerns.Noise is noted particularly for distorted images or images with overlapping symptoms, and classification performance is affected by the remaining detection challenges.Source: §6.3 · pp. 22–23

• The model achieves performance comparable to fully supervised approaches with fewer labeled samples and outperforms existing methods.The reported evaluation uses sugarcane leaf disease data.Source: §7 · p. 23
• U-SSDNet combines the Mean-Teacher framework, wavelet-based preprocessing, KL-divergence-driven pseudo-label correction, and supervised contrastive learning.Source: §7 · p. 23
• Wavelet preprocessing improves lesion visibility and within-class consistency, KL correction reduces confirmation bias, and supervised contrastive learning sharpens class separation.Source: §7 · p. 23
• The conclusion describes the approach as practical for early disease detection and precise agricultural monitoring.Source: §7 · p. 23