• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Sunday, November 22, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Hybrid Deep Learning Models Accurately Classify Sugarcane Leaf Diseases Across Six Classes

by
November 21, 2026
in Technology
Reading Time: 6 mins read
0
Hybrid Deep Learning Models Accurately Classify Sugarcane Leaf Diseases Across Six Classes

Hybrid Deep Learning Models Accurately Classify Sugarcane Leaf Diseases Across Six Classes

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A comprehensive comparative investigation has evaluated whether combining convolutional neural networks and vision transformers into unified ensemble frameworks enhances automated diagnostic accuracy for foliar sugarcane infections. Evaluating four convolutional backbones alongside two transformer architectures under identical experimental conditions on an extensively audited dataset of 19,926 leaf images, researchers observed that a six-model weighted probability fusion attained 95.00 percent test accuracy and a macro-F1 score of 94.52 percent. Concurrently, a meta-learning ensemble driven by CatBoost reached 95.10 percent accuracy and a macro-F1 score of 94.62 percent. Among standalone architectures, the convolutional model DenseNet121 registered the highest individual classification capability, achieving 94.13 percent accuracy and a 93.58 percent macro-F1 score. While the improvement delivered by the six-model weighted ensemble over standalone DenseNet121 proved statistically significant across paired testing protocols, the minor performance margin separating the full hybrid ensemble from an ensemble composed exclusively of convolutional networks was not statistically detectable, illuminating critical trade-offs between architectural complexity, predictive diversity, and computational deployment overhead.

Sugarcane serves as an indispensable commercial crop across developing agricultural economies, functioning not only as the primary raw commodity for global sugar production but also anchoring bioethanol manufacturing, localized agro-processing enterprises, and rural employment structures. Nevertheless, commercial cultivation remains vulnerable to disruptive foliar pathogens, notably Red Rot, Rust, Mosaic, Yellow Leaf Disease, and Bacterial Blight, all of which compromise overall stalk tonnage, cane vitality, and sucrose yields when diagnosis is delayed. Conventional diagnostic practices in field settings continue to rely predominantly on physical visual appraisal by agricultural extension personnel or farmers. This manual diagnostic paradigm is inherently labor-intensive, slow, subjective, and prone to significant observer bias, particularly across vast plantation acreages where symptom variations occur under fluctuating sunlight, natural occlusion, and dynamic environmental degradation. These operational bottlenecks have accelerated research into computational computer vision frameworks capable of delivering consistent, high-throughput disease screening from digital imagery.

Computer vision applications within digital agriculture have historically favored convolutional neural networks, including ResNet, DenseNet, EfficientNet, and VGGNet, due to their specialized capacity for extracting localized spatial representations such as distinct lesion margins, chlorotic mottling, color deviations, and microscopic foliar surface textures. However, traditional convolution mechanisms rely primarily on local receptive fields, which inherently constrain their ability to represent extended spatial relationships spanning an entire leaf blade. To address this contextual limitation, vision transformers have emerged, applying global self-attention across segmented image patches to map broader structural associations. Swin Transformer variants refine this approach further by integrating shifted-window self-attention, creating hierarchical multi-scale feature hierarchies while permitting information exchange across local window boundaries. Despite these conceptual strengths, empirical evidence establishing whether vision transformers meaningfully enhance diagnostic precision or justify their substantial parameter footprints when integrated with convolutional architectures for sugarcane pathology has remained sparse.

To guarantee equitable experimental conditions and prevent artificial diagnostic inflation, the investigative team deployed a meticulous data-auditing protocol across an experimental corpus of 19,926 digital images spanning six diagnostic categories: Bacterial Blight, Healthy, Mosaic, Red Rot, Rust, and Yellow Leaf Disease. Initial cryptographic audits established that the raw collection contained 19,086 unique image hashes alongside 736 duplicate groupings, while filename examinations delineated 3,321 distinct source families representing image variations. The authors applied perceptual hashing complemented by Scale-Invariant Feature Transform adjudication to screen for candidate near-duplicate pairings, ultimately resolving the repository into 291 source relationships and 3,076 unified source groups. These grouped records were subsequently allocated across training, validation, and testing partitions following an approximate 70:15:15 stratified ratio, ensuring no single source group spanned multiple partitions. This auditing structure yielded 13,968 training images, 2,976 validation samples, and 2,982 held-out test images, establishing an audited baseline that minimizes image leakage while acknowledging the absence of plant-level metadata.

All input images were standardized to uniform dimensions of 224 by 224 pixels with three RGB channels, applying architecture-specific ImageNet preprocessing regimes across bilinear interpolation pipelines. The convolutional cohort encompassed EfficientNetB0, utilizing compound depth, width, and resolution scaling; DenseNet121, leveraging dense iterative layer connections for cumulative feature reuse; ResNet50, incorporating identity shortcut paths to prevent training gradient degradation; and a customized VGG16 backbone incorporating a modified classification head comprising global average pooling, batch normalization, fifty-percent dropout, and a final dense projection. Transformer evaluations incorporated ViT-B/16 and Swin-B/7, whose standardized pixel intensities were normalized against ImageNet statistical means and variances. Optimization for all convolutional networks utilized the Adam optimizer initialized with a learning rate of 0.0001, batch sizes of 16, and early stopping triggers across a 30-epoch ceiling, whereas transformer models underwent targeted full fine-tuning under gradient clipping caps of 1.0.

Evaluation of standalone model performance across the 2,982 untouched test images revealed distinct performance hierarchies among the investigated architectural paradigms. DenseNet121 secured the foremost standalone result, delivering 94.13 percent test accuracy, a 93.58 percent macro-F1 score, and an exceptional macro-averaged receiver operating characteristic area under the curve of 0.9957. ResNet50 followed with 93.33 percent test accuracy, while EfficientNetB0 achieved 92.39 percent accuracy, demonstrating exceptional parameter economy by requiring merely 4.057 million parameters and 0.800 GFLOPs per image. The modified VGG16 network attained 91.45 percent accuracy. Conversely, standalone transformer performance trailed the convolutional baselines; Swin-B/7 achieved 90.04 percent accuracy and an 89.54 percent macro-F1 score alongside a 0.9916 macro-AUC, whereas ViT-B/16 recorded 87.89 percent accuracy and an 86.87 percent macro-F1 score, illustrating that expanded parameter capacity does not inherently translate into superior diagnostic accuracy within this domain.

To explore potential operational complementarities across these divergent feature-learning paradigms, the researchers engineered a probability-level fusion framework that concatenated predicted six-class probability distributions from all six base learners into a 36-dimensional meta-feature representation for every image. Using the 2,976 validation instances, a random simplex search across 200,000 candidate combinations determined optimal multiclass log-loss fusion weights, yielding fixed allocations of 0.4076 for DenseNet121, 0.2575 for ResNet50, 0.1246 for EfficientNetB0, 0.1005 for Swin-B/7, 0.0945 for VGG16, and 0.0154 for ViT-B/16. Simultaneously, meta-learning candidates including CatBoost, XGBoost, LightGBM, and Support Vector Machines were benchmarked using five-fold stratified cross-validation on this fixed validation matrix. Evaluated on the held-out test partition, the resulting CatBoost meta-learner achieved 95.10 percent accuracy, a 94.62 percent macro-F1 score, and a 0.9972 macro-AUC, whereas the log-loss-weighted fusion recorded 95.00 percent accuracy and 94.52 percent macro-F1.

Statistical validation utilizing 5,000 paired bootstrap resamples and two-sided exact McNemar testing with Holm adjustments established the boundaries of these performance gains. The 0.872 percentage point accuracy improvement demonstrated by the six-model weighted ensemble over standalone DenseNet121 proved statistically significant, supported by a 95 percent confidence interval spanning 0.402 to 1.341 percentage points and an adjusted p-value of 0.004095. Across discordant test samples, the ensemble corrected 39 instances misclassified by DenseNet121, while DenseNet121 was accurate on 13 samples failed by the ensemble. In contrast, comparing the six-model weighted ensemble against an ensemble restricted entirely to convolutional models revealed an accuracy disparity of merely 0.067 percentage points, which was not statistically significant with an adjusted p-value of 1.000 and a confidence interval spanning negative 0.335 to positive 0.201 percentage points. Similarly, the 0.101 percentage point descriptive gain of CatBoost over weighted fusion was statistically indistinguishable.

Architectural ablation and sensitivity assessments underscored the disproportionate influence exerted by specific convolutional backbones within the ensemble framework. Sequentially removing individual models demonstrated that omitting DenseNet121 triggered an accuracy drop of 1.040 percentage points down to 93.96 percent, an outcome confirmed statistically significant with an adjusted p-value of 0.0068. Omitting ResNet50 similarly induced a significant performance loss of 0.604 percentage points to 94.40 percent with an adjusted p-value of 0.0315. Conversely, excising EfficientNetB0, VGG16, ViT-B/16, or Swin-B/7 generated marginal accuracy shifts whose bootstrap confidence intervals spanned zero. An ensemble composed exclusively of transformer models demonstrated pronounced performance degradation, declining to 91.48 percent accuracy, which was 3.521 percentage points lower than the full six-model hybrid ensemble, confirming that transformer streams provided limited additive diagnostic utility despite exhibiting quantifiable prediction diversity.

Investigation of predictive diversity and probabilistic calibration highlighted notable differences across individual models and ensemble configurations. Pairwise evaluations revealed that ViT-B/16 and Swin-B/7 manifested the highest mutual disagreement rate at 0.1321 alongside an output probability correlation of 0.9082, whereas EfficientNetB0 paired with ViT-B/16 demonstrated the highest complementary correctness metric at 0.1100. Calibration analysis indicated that the six-model weighted ensemble achieved an expected calibration error of 0.0154, a maximum calibration error of 0.1909, a multiclass Brier score of 0.0755, and a log loss of 0.1488. The CatBoost meta-learner further optimized probabilistic calibration, yielding an improved Brier score of 0.0695 and log loss of 0.1403. Confusion matrix inspections showed the lowest error rates for Bacterial Blight at 99.58 percent recall, whereas diagnostic confusions concentrated around subtle phenotypic transitions between Mosaic, Yellow Leaf Disease, and Healthy foliage.

Deployment feasibility was rigorously quantified through computational complexity benchmarks executed on an NVIDIA L4 GPU utilizing 16-sample mini-batches under float32 precision. The six-model ensemble accumulated 221.986 million parameters and consumed 109.541 GFLOPs per image, producing an amortized per-image latency of 76.679 milliseconds corresponding to a throughput of 13.04 images per second. In comparison, standalone DenseNet121 achieved its 94.13 percent accuracy using only 7.048 million parameters, 5.700 GFLOPs, and 29.578 milliseconds of amortized latency, while EfficientNetB0 required only 4.057 million parameters and 0.800 GFLOPs. Standalone ViT-B/16 and Swin-B/7 demanded 85.803 million and 86.749 million parameters while recording per-image latencies of 5.088 and 6.454 milliseconds, respectively. Consequently, adopting the hybrid ensemble imposes a 31.5-fold increase in parameter burden and a 19.2-fold expansion in floating-point operations over DenseNet121 to secure a modest 0.872 percentage point increase in accuracy.

The authors contextualized their empirical findings by articulating critical methodological constraints and delineating future research trajectories. The reported evaluations were derived from a single experimental dataset that, despite rigorous source-group hashing, lacked plant-level and temporal acquisition metadata, preventing definitive claims regarding model generalizability across independent geographical plantations or varying illumination regimes. Furthermore, all experimental figures reflect a single unified training execution, meaning observed metrics do not capture variance associated with alternate stochastic seed initializations. Future efforts must prioritize cross-dataset validation against field-acquired imagery, explore self-supervised representation pretraining to mitigate label scarcity, and investigate model compression strategies including neural pruning, post-training quantization, and knowledge distillation to translate high-performing hybrid ensembles into computationally viable edge-diagnostic tools for precision agricultural machinery.

Subject of Research: Deep learning architectures and hybrid ensemble methods for multi-class sugarcane leaf disease classification

Article Title: CNN and vision transformer-based hybrid ensemble learning for sugarcane leaf disease classification

Article References: Chand, H. V., Islam, M. N. S., Jiang, W., & Cengiz, K. (2026). CNN and vision transformer-based hybrid ensemble learning for sugarcane leaf disease classification. Discover Computing, 29(1), Article 918. https://doi.org/10.1007/s10791-026-10805-7

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10805-7

Keywords: sugarcane leaf diseases, convolutional neural networks, vision transformers, DenseNet121, ResNet50, EfficientNetB0, VGG16, Swin Transformer, Vision Transformer, hybrid ensemble learning, CatBoost meta-learner, precision agriculture, probability-level fusion, computational complexity, expected calibration error, ablation study, image classification, plant pathology

News Source: Blake Davidson. (November 21, 2026). Hybrid Deep Learning Models Accurately Classify Sugarcane Leaf Diseases Across Six Classes. Scienmag.

Tags: ablation studyCatBoost meta-learnercomputational complexityconvolutional neural networksDenseNet121EfficientNetB0expected calibration errorhybrid ensemble learningimage classificationplant pathologyprecision agricultureprobability-level fusionResNet50sugarcane leaf diseasesSwin TransformerVGG16vision transformervision transformers
Share12Tweet7Share2ShareShareShare1

Related Posts

Strut-and-Tie Model Accurately Predicts Flexural Capacity of Prefabricated Bridge Wet Joints

Strut-and-Tie Model Accurately Predicts Flexural Capacity of Prefabricated Bridge Wet Joints

November 22, 2026
Precast Concrete Shear Walls with UHPC Boundary Connections Improve Seismic Performance

Precast Concrete Shear Walls with UHPC Boundary Connections Improve Seismic Performance

November 21, 2026

Portable Raman Spectroscopy Screens Microbial Carotenoids for Solar Cell Fabrication

November 21, 2026

New Unified Contraction Framework Establishes Fixed Point Results in Metric Spaces

November 21, 2026

POPULAR NEWS

  • West African Study Finds No Detectable Link Between Carbon Intensity and Mortality

    West African Study Finds No Detectable Link Between Carbon Intensity and Mortality

    29 shares
    Share 12 Tweet 7
  • GEN1 Resolvase Protects Genome Integrity and Self-Renewal in Mouse Embryonic Stem Cells

    29 shares
    Share 12 Tweet 7
  • Malawi Simulation Shows Health System Strengthening Outperforms Disease-Specific Healthcare Funding

    29 shares
    Share 12 Tweet 7
  • Leadership Support and Adapted Workflows Shape Rural Wisconsin Police Deflection Programs

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Screening Human Fat Cells Uncovers PLCXD1 Enzyme as Insulin-Regulated Lipid Storage Driver

Disinhibited Prefrontal Somatostatin Neurons Mediate Stress Resilience in Male Mice

Open-Source Network Tool Quantifies Mammary Ductal Branching in Developing Mice

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.