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

Shalini Dangi; Divya Teja; Omkar Kabde; Shahid Shafi Dar; Nagendra Kumar

cs.CVOct 13, 2026

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

Authors: Shalini Dangi, Divya Teja, Omkar Kabde, Shahid Shafi Dar, Nagendra Kumar

Organizations: Indian Institute of Technology Indore, Madhya Pradesh, India · Chaitanya Bharathi Institute of Technology, Telangana, India

Abstract

Sugarcane leaf diseases severely impact crop yield and quality, making accurate and timely detection critical for sustainable agriculture. Conventional deep learning models for plant disease detection depend on large annotated datasets, which are expensive to obtain and challenging to prepare due to challenges commonly observed across crop disease detection tasks, including (i) high intra-class variation caused by diverse symptom expressions across growth stages, (ii) subtle inter-class similarities between diseased and healthy tissues, and (iii) the risk of noisy labels degrading model reliability. To handle these challenges, we presented an Uncertainty-aware Mean-Teacher based Semi-Supervised Learning Disease Detection Network (U-SSDNet) that leverages unlabeled data while maintaining robustness. First, discrete wavelet transform preprocessing is applied to enhance lesion visibility and reduce intra- class variation. Subsequently, the KL divergence is employed to refine the pseudo-labels and reduce the impact of confirmation bias. At the same time, supervised contrastive learning is applied to make the features more distin- guishable and improve the separation among different classes. Experimental evaluations on sugarcane leaf disease datasets, considered a representative case study, demonstrate that the proposed approach achieves performance close to that of fully supervised models while requiring far fewer annotated samples and consistently outperforms state-of-the-art supervised and semi-supervised approaches.

Figures & tables

Figure 1: The proposed U-SSDNet framework contains four key elements: (i) Sugarcane Leaf Preprocessing and Augmentation; (ii) A Mean-Teacher model for Consistency Learning; (iii) Supervised Contrastive Learning for discriminative representation; and (iv) KL-divergence–based Uncertainty Estimation to guide the utilization of unlabeled data.

Figure 2: DWT decomposition of a sugarcane leaf. Low-frequency (LL) captures the global structure; high-frequency sub-bands (LH, HL, HH) highlight textures and edges associated with lesions.

Figure 3: Basic Mean-Teacher model for sugarcane leaf disease detection.

Figure 4: Architecture of EfficientNetB4 model.

Panel of Figure 5 — Figure 5: Distribution of samples across various categories in SLD-1, SLD-2, and SLD-3.

Figure 6: Illustration of sugarcane leaf disease types on the SLD-1 dataset.

Figure 7: Overview of the SLD-2 dataset illustrating various sugarcane leaf disease types.

Figure 8: Overview of sugarcane leaf disease types on the SLD-3 dataset.

Hyperparameter	Value
Image Size	224 $\times$ 224
Batch Size	8
Optimizer	AdamW
Learning Rate	0.001
Number of Epochs	35
Weight Decay	0.0001

Table 1: This table shows the final hyperparameter details of U-SSDNet.

Model	50% Labeled				20% Labeled				10% Labeled
Model	Acc	Pre	Rec	F1	Acc	Pre	Rec	F1	Acc	Pre	Rec	F1
ResNet-50 ( He et al., 2016 )	93.67	93.77	93.67	93.67	87.72	87.87	87.72	87.52	82.97	84.31	82.97	83.03
VGG-16 ( Liu and Deng, 2015 )	70.50	70.45	70.50	70.14	53.07	54.27	53.07	52.88	46.93	48.82	46.93	47.03
DenseNet-121 ( Huang et al., 2017 )	93.27	93.33	93.27	93.22	90.30	90.22	90.30	90.21	70.20	70.22	71.02	70.22
ViT ( Dosovitskiy et al., 2020 )	93.42	94.22	93.42	93.42	70.22	71.82	70.71	70.81	50.02	49.05	50.04	50.55
Swin Transformer ( Liu et al., 2021 )	85.66	85.45	85.66	85.66	75.52	77.45	75.52	75.66	45.45	45.67	45.45	45.49

Table 5: Performance comparison of SOTA models and the proposed U-SSDNet framework on varying labeled data proportions on the SLD-1 dataset.

Model	50% Labeled				20% Labeled				10% Labeled
Model	Acc	Pre	Rec	F1	Acc	Pre	Rec	F1	Acc	Pre	Rec	F1
ResNet-50 ( He et al., 2016 )	91.48	91.73	91.48	91.51	92.07	92.03	92.07	91.96	85.48	85.18	85.48	84.93
VGG-16 ( Liu and Deng, 2015 )	93.93	94.12	93.93	93.95	91.48	91.51	91.48	91.46	80.44	82.73	80.44	79.42
DenseNet-121 ( Huang et al., 2017 )	91.93	92.20	91.93	91.89	88.67	89.23	88.67	88.67	86.81	88.24	86.81	86.71
ViT ( Dosovitskiy et al., 2020 )	84.59	85.59	84.59	84.67	74.81	77.53	74.81	75.01	65.19	65.72	65.19	63.92
Swin Transformer ( Liu et al., 2021 )	92.67	92.69	92.67	92.63	91.41	91.42	91.41	91.36	85.56	85.80	85.56	85.27

Table 6: Performance comparison of SOTA models and the proposed U-SSDNet framework under different labeled data proportions on the SLD-2 dataset.

Model	Acc	diff	Pre	diff	Rec	diff	F1	diff
U-SSDNet (w/o)	92.21	2.59	92.38	2.44	92.98	1.90	92.51	1.69
U-SSDNet (HOG)	91.28	3.52	91.53	3.29	92.11	2.77	91.45	2.75
U-SSDNet (LPB)	92.52	2.28	92.78	2.04	93.19	1.69	92.81	1.39
U-SSDNet (Wavelet)	94.80	0.00	94.82	0.00	94.88	0.00	94.20	0.00

Table 7: Comparative analysis of U-SSDNet model performance across different feature enhancement methods. The “diff” columns represent the performance gap relative to the best-performing configuration (Wavelet).

Model	Acc	diff	Pre	diff	Rec	diff	F1	diff
U-SSDNet (w/o Aug)	92.52	2.28	92.61	2.21	93.02	1.86	92.62	1.58
U-SSDNet (Deep Aug)	92.83	1.97	93.09	1.73	93.31	1.57	93.14	1.06
U-SSDNet (Cut&Delete Aug)	93.77	1.03	93.84	0.98	94.15	0.73	93.96	0.24
U-SSDNet (Deep+Cut&Delete)	94.80	0.00	94.82	0.00	94.88	0.00	94.20	0.00

Table 8: Comparative analysis of U-SSDNet model performance across different augmentation techniques. The “diff” columns indicate the performance gap relative to the best-performing configuration (Deep+Cut&Delete).

Model	Acc	diff	Pre	diff	Rec	diff	F1	diff
U-SSDNet (ResNet)	71.34	23.46	59.56	35.26	73.04	21.84	64.88	29.32
U-SSDNet (DenseNet)	84.42	10.38	86.21	8.61	85.64	9.24	84.12	10.08
U-SSDNet (EfficientNet)	94.80	0.00	94.82	0.00	94.88	0.00	94.20	0.00

Table 9: Comparative analysis of U-SSDNet model performance across different feature extraction methods. The “diff” columns represent the performance gap relative to the best-performing configuration (EfficientNet).

Model	Acc	diff	Pre	diff	Rec	diff	F1	diff
U-SSDNet (w/o UA)	91.28	3.52	91.54	3.28	92.05	2.83	91.55	2.65
U-SSDNet (MC)	92.83	1.97	93.45	1.37	93.06	1.82	93.17	1.03
U-SSDNet (KL-Divergence)	94.80	0.00	94.82	0.00	94.88	0.00	94.20	0.00

Table 10: Comparative analysis of U-SSDNet performance across different uncertainty-aware methods. The “diff” columns represent the performance gap relative to the best-performing configuration (KL-Divergence).

Panel of Figure 9 — Figure 9: Ablation study results on the Sugarcane Leaf Disease (SLD-1) dataset on various components such as: (a) Augmentation module; (b) Feature Encoding module; (c) Feature Extraction module; and (d) Uncertainty-Aware module.

Figure 10: Prediction results of sugarcane leaf diseases on the SLD-1 dataset.

Figure 11: Prediction results of sugarcane leaf diseases on the SLD-2 dataset.

Figure 12: Comparison of Grad-CAM visualizations for different methods on the SLD-1 dataset (10% labeled data).

Figure 13: Comparison of Grad-CAM visualizations for different methods on the SLD-2 dataset (10% labeled data).

Figure 14: Comparison of Grad-CAM visualizations for different methods on the SLD-3 dataset (10% labeled data).

Figure 15: Performance across various classes in the SLD-1 and SLD-2 datasets. Figures (a) and (b) show the confusion matrices of U-SSDNet for SLD-1 and SLD-2, respectively, while Figures (c) and (d) present the corresponding ROC curves for both datasets.

Figure 16: This figure presents a comparison of CAM and LIME visualizations on the SLD-1 dataset for different diseases and shows the effectiveness of the proposed method in identifying disease-affected regions.

Panel of Figure 17 — Figure 17: t-SNE representation of SLD-1 data with different labeling ratios.

Explore similar work

DenseSwinV2: Channel Attentive Dual Branch CNN Transformer Learning for Cassava Leaf Disease Classification
Mar 26, 2026Shah Saood, Saddam Hussain KhanPlant Disease Classification Hybrid CNN-Transformer Architectures
Artificial Intelligence Lab, Department of Computer Systems Engineering, University of Engineering and Applied Sciences · Artificial Intelligence Lab, Department of Computer Systems Engineering, University of Engineering and Applied Sciences (UEAS), Swat 19060, Pakistan · Interdisciplinary Research Center for Smart Mobility and Logistics +1
LightTeaNet: A Weakly Supervised Lightweight CNN for Multi-Label Tea Leaf Disease Detection and Localization
Aug 14, 2026Naif Haider Chowdhury, Md Rahim, Syed Farhan Hasan +2Plant Disease Classification Multi-Label Classification
Department of Computer Science & Engineering, Leading University, Kamal Bazar, Sylhet, 3112, Sylhet, Bangladesh.
CottonLeafVision: An Explainable and Robust Deep Learning Framework for Cotton Leaf Disease Classification
Jun 12, 2026Rafi Ahamed, Md. Abir Rahman, Tasnia Tarannum Roza +3Plant Disease Classification Explainable Image Classification
Dept. of CSE East West University Dhaka, Bangladesh

Mar 26, 2026J/K move · Enter open · S save

DenseSwinV2: Channel Attentive Dual Branch CNN Transformer Learning for Cassava Leaf Disease Classification

Shah Saood, Saddam Hussain Khan

Artificial Intelligence Lab, Department of Computer Systems Engineering, University of Engineering and Applied Sciences · Artificial Intelligence Lab, Department of Computer Systems Engineering, University of Engineering and Applied Sciences (UEAS), Swat 19060, Pakistan · Interdisciplinary Research Center for Smart Mobility and Logistics+1

Plant Disease Classification Hybrid CNN-Transformer Architectures Vision Transformer

Open paper AskPDF arXiv

This work presents a new Hybrid Dense SwinV2, a two-branch framework that jointly leverages densely connected convolutional features and hierarchical customized Swin Transformer V2 (SwinV2) representations for cassava disease classification. The proposed framework captures high resolution local features through its DenseNet branch, preserving the fine structural cues and also allowing for effective gradient flow. Concurrently, the customized SwinV2 models global contextual dependencies through the idea of shifted-window self attention, which enables the capture of long range interactions critical in distinguishing between visually similar lesions. Moreover, an attention channel-squeeze module is employed for each CNN Transformer stream independently to emphasize discriminative disease related responses and suppress redundant or background driven activations. Finally, these discriminative channels are fused to achieve refined representations from the dense local and SwinV2 global correlated strengthened feature maps, respectively. The proposed Dense SwinV2 utilized a public cassava leaf disease dataset of 31000 images, comprised of five diseases, including brown streak, mosaic, green mottle, bacterial blight, and normal leaf conditions. The proposed Dense SwinV2 demonstrates a significant classification accuracy of 98.02 percent with an F1 score of 97.81 percent, outperforming well-established convolutional and transformer models. These results underline the fact that Hybrid Dense SwinV2 offers robustness and practicality in the field level diagnosis of cassava disease and real world challenges related to occlusion, noise, and complex backgrounds.

Abstract

Explore similar work

DenseSwinV2: Channel Attentive Dual Branch CNN Transformer Learning for Cassava Leaf Disease Classification

LightTeaNet: A Weakly Supervised Lightweight CNN for Multi-Label Tea Leaf Disease Detection and Localization

CottonLeafVision: An Explainable and Robust Deep Learning Framework for Cotton Leaf Disease Classification