• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Tuesday, November 17, 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

IGE-SegNeXt Model Segments Cabbage Leaf Damage but Shows Inconsistent Field Gains

by
November 17, 2026
in Agriculture, Technology
Reading Time: 6 mins read
0
IGE-SegNeXt Model Segments Cabbage Leaf Damage but Shows Inconsistent Field Gains

IGE-SegNeXt Model Segments Cabbage Leaf Damage but Shows Inconsistent Field Gains

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Agricultural researchers have developed an integrated deep learning framework named IGE-SegNeXt to delineate and quantify disease symptoms and insect damage across field-grown cabbage seedlings at pixel resolution, though subsequent validation indicates its performance advantages over standard models vary across data partitions. Published in Smart Agricultural Technology, the investigation led by Hongda Li and colleagues analyzed Brassica oleracea var. capitata, specifically cultivar Zhonggan 21, during its post-transplant establishment phase at an experimental farm in Beijing, China. The post-transplant seedling stage represents an especially sensitive window in vegetable cultivation because juvenile plants confront concurrent biotic and abiotic pressures, where damage across a few early leaves can compromise a substantial fraction of the canopy. Conventional scouting procedures depend heavily on labor-intensive visual inspection that proves difficult to scale across commercial plots, while standard computer vision tools often rely on coarse image classification or bounding boxes that fail to capture the irregular boundaries typical of foliar lesions and feeding damage.

The biological rationale behind monitoring transplanted seedlings centers on stage-specific vulnerability to foliar pathogens and chewing pests that can undermine early vegetative vigor. In cabbage cultivation, prior research demonstrates that bacterial infection caused by Xanthomonas campestris pv. campestris significantly impairs shoot biomass accumulation and photosynthetic capacity in young seedlings, whereas mature plants exhibit far greater physiological tolerance. Concurrently, open-field seedlings frequently endure severe insect pressure from major pest species, notably diamondback moth and cabbage aphid, which chew margins, create holes, and deplete leaf area. While artificial intelligence tools like YOLO-based detection networks have advanced automated monitoring in greenhouse crops, bounding boxes fundamentally lack spatial precision. Semantic segmentation offers a solution by assigning categorical labels to every visible pixel, but implementing it under open-field conditions introduces severe hurdles, including fluctuating natural illumination, mutual leaf occlusion, background clutter from soil, and spectral similarity between damaged tissue and surroundings.

To generate empirical training data reflecting authentic agronomic conditions, the researchers constructed a dedicated dataset at the National Precision Agriculture Research Base in Xiaotangshan, Changping District, Beijing. Image collection occurred on clear days in late September 2024 between 10:00 and 14:00 under natural sunlight to capture realistic illumination angles and shadow patterns. The team recorded top-down photographs of Zhonggan 21 cabbage seedlings using an iQOO Neo8 Pro smartphone positioned at working distances between 0.2 and 0.5 meters, mimicking the clearance of field plant protection machinery. Each high-resolution capture originally measured 4096 by 3072 pixels with a 6-millimeter focal length before being resized to 800 by 600 pixels to optimize computational throughput. The investigators screened the repository to discard blurred, defocused, or incorrectly exposed frames, ultimately isolating 350 high-quality images from 350 distinct individual seedlings cultivated across four contiguous planting ridges. Recording specific ridge identifiers for every sampled cabbage allowed the authors to retain relative layout coordinates.

Manual annotation of the 350 field images was executed using AnyLabeling software under a leaf-level assignment protocol verified by professional agricultural experts. The authors established four mutually exclusive semantic categories: Background, Healthy, Disease-symptom, and Pest-damage, categorizing each visible leaf according to its dominant visual phenotype. Under these criteria, healthy foliage exhibited uniform green coloration and intact margins without visible spotting, chlorosis, necrosis, or perforations. Leaves categorized as disease-symptom displayed chlorosis, localized spots, wilting, or necrotic patches while maintaining an unbroken perimeter, whereas pest damage was identified by structural discontinuities such as feeding notches, perforated laminae, and chewed margins. Across the 350 unaugmented images comprising 168 million annotated pixels, the dataset demonstrated severe class imbalance: Background encompassed 71.8064 percent, Disease-symptom comprised 13.8967 percent, Healthy accounted for 12.8671 percent, and Pest-damage represented merely 1.4299 percent. At the leaf-instance level in the original training set, healthy leaves accounted for 57.09 percent (1244 leaves), disease symptoms represented 37.08 percent (808 leaves), and pest damage totaled just 5.83 percent (127 leaves).

To prevent data leakage, the authors partitioned the 350 seedlings into mutually exclusive sets using an 8-to-1-to-1 ratio, allocating 280 seedlings for training, 35 for validation, and 35 as a held-out test set. Data augmentation was applied exclusively to the training subset to represent orientation, optical sharpness, and illumination shifts without altering semantic classes. Operations included random horizontal flipping with probability 0.5, uniform rotations between minus 15 and plus 15 degrees, Gaussian blurring using a 13-by-13 kernel with standard deviation sigma of 2.3, and brightness and contrast scaling factors spanning 0.7 to 1.2. Generating five augmented variants per original training sample expanded the training pool to 1680 images containing 13,074 annotated leaf instances. Computational experiments were executed on an Ubuntu 24.04 LTS workstation featuring an NVIDIA A100 GPU with 80 gigabytes of memory, an Intel Xeon Platinum 8260 CPU, PyTorch, and Python 3.8.20. Training utilized the AdamW optimizer with weight decay of 0.01, batch size of 4, and 40,000 iterations including a 1500-iteration warmup.

Selecting an appropriate loss function proved critical for managing the numerical disparity between dominant background pixels and scarce pest-damaged tissues. Using SegNeXt as an initial baseline, the researchers evaluated four loss formulations: Lovasz-Softmax, cross-entropy, Dice loss, and multi-class focal loss. Experimental results demonstrated that while all functions effectively segmented background pixels, standard formulations struggled with minority categories. Dice loss achieved 82.82 percent intersection over union on healthy tissue and 80.33 percent on disease symptoms, yet yielded an unusable 0.01 percent on pest damage, reducing its mean intersection over union to 65.11 percent. Cross-entropy produced 35.19 percent on pest damage with a 63.80 percent mean intersection over union, while Lovasz-Softmax registered 6.50 percent on pest damage and a 57.50 percent overall mean. In contrast, multi-class focal loss with alpha of 1.0 and gamma of 2.0 lifted pest-damage intersection over union to 63.33 percent and achieved the highest baseline mean intersection over union at 78.23 percent, making it the standard objective for downstream evaluations.

To establish an optimized baseline, the authors explored model scales and decoder dimensions within the convolutional SegNeXt family. Assessing configurations across SegNeXt-T, SegNeXt-S, SegNeXt-B, and SegNeXt-L demonstrated diminishing performance returns alongside escalating parameter footprints. SegNeXt-S balanced computational cost and accuracy, and pairing it with a 512-dimensional decoder established an optimized baseline achieving 78.65 percent mean intersection over union with 14.61 million parameters and 72.90 gigaFLOPs. Building upon this foundation, the authors engineered IGE-SegNeXt by introducing the Inception Depthwise Convolutional Attention Network (IDCAN), Global Context Feature Enhancement (GCFE), and Channel-Efficient Multi-Scale Attention (CEMA). IDCAN replaces the standard MSCAN encoder while preserving the four-stage hierarchy with 2, 2, 4, and 2 blocks producing feature maps downsampled by factors of 4, 8, 16, and 32. Within each block, the IDCA module partitions input channels across four parallel paths: a 3-by-3 depthwise convolution, two orthogonal strip convolutions measuring 1-by-11 and 11-by-1, and an identity mapping branch. Concatenating these pathways followed by a 1-by-1 convolution and element-wise attention multiplication allows the network to capture directional margins and localized discoloration.

Decoder enhancements were integrated to suppress background interference and sharpen class discrimination across overlapping foliage. The Global Context Feature Enhancement module processes aggregated decoder features by applying parallel global average pooling and global maximum pooling across spatial dimensions, summing the descriptors, and projecting them through a two-layer bottleneck reducing channel depth to one-sixteenth before restoring it, followed by sigmoid activation and learnable channel scaling. This sequence reinforces cabbage-specific channels while attenuating spectral interference from soil, weeds, and shadows. Concurrently, the Channel-Efficient Multi-Scale Attention module refines leaf-status differentiation by applying channel attention to decoder features, then fusing them through element-wise addition with aligned high-level semantic representations from the fourth encoder stage. The fused representation undergoes cross-dimensional multi-scale refinement using Efficient Multi-Scale Attention. In ablation experiments on the original fixed split, adding IDCA alone boosted mean intersection over union to 80.81 percent. GCFE alone achieved 79.42 percent, and CEMA alone attained 80.56 percent. Combining all three components delivered 82.01 percent mean intersection over union, 97.07 percent pixel accuracy, and 89.44 percent mean Dice score.

When benchmarked against established segmentation models on the fixed test partition, IGE-SegNeXt surpassed standard convolutional and transformer alternatives, including UNet at 64.93 percent mean intersection over union, PSPNet at 62.51 percent, FCN at 73.07 percent, OcrNet at 65.85 percent, and Segmenter at 66.29 percent. It also outperformed DeepLabv3+ with ResNet-50, which attained 79.74 percent mean intersection over union but required 456.63 gigaFLOPs and 26.79 million parameters, and SegFormer with MiT-B2, which reached 80.04 percent. However, class-specific evaluation of IGE-SegNeXt revealed pronounced performance disparities. While Background achieved an intersection over union of 99.21 percent, Disease-symptom reached 86.41 percent, and Pest-damage recorded 81.95 percent, the Healthy class exhibited the lowest accuracy, logging an intersection over union of only 60.47 percent and an F1-score of 75.37 percent. Pixel-level confusion matrix analysis revealed that Healthy foliage achieved a recall of only 68.90 percent, with 23.64 percent of healthy pixels misclassified as disease symptoms and 6.23 percent assigned to background, while 8.21 percent of pest-damage pixels were misclassified as healthy.

Despite strong single-split metrics, the authors performed a paired ten-fold cross-validation that demonstrated the model does not achieve a statistically significant average improvement over the baseline architecture. In this ten-fold analysis across randomized training seeds, the optimized SegNeXt baseline averaged a mean intersection over union of 78.263 plus or minus 5.070 percent, whereas IGE-SegNeXt averaged 77.553 plus or minus 4.146 percent, representing a paired deficit of minus 0.710 percentage points with an unadjusted 95 percent confidence interval spanning minus 2.569 to plus 1.149 percentage points. Exploratory hypothesis testing via a two-sided paired t-test yielded t of minus 0.864 (p equals 0.410), while a two-sided Wilcoxon signed-rank test produced W of 26.0 (p equals 0.922), confirming no statistically detectable performance distinction. Beyond pure segmentation, the authors established continuous canopy metrics by calculating Disease-Symptom Canopy Proportion and Pest-Damage Canopy Proportion as percentages of total visible seedling pixels, plotting single-date spatial heatmaps across four experimental ridges. Nevertheless, data collection was restricted to a single cultivar, developmental stage, geographic site, and imaging date, with leaf-level dominant annotations obscuring coexisting damage on individual leaves. Furthermore, canopy proportion indices lack biological validation against standardized disease scales, and proposed edge-device execution on NVIDIA Jetson hardware remains untested.

Subject of Research: Semantic segmentation of foliar disease symptoms and pest damage in field cabbage seedlings

Article Title: IGE-SegNeXt: semantic segmentation of disease symptoms and pest damage for precision monitoring of field cabbage

Article References: Li, H., Si, C., Su, K., Wang, X., Wu, H., & Zhao, C. (2026). IGE-SegNeXt: semantic segmentation of disease symptoms and pest damage for precision monitoring of field cabbage. Smart Agricultural Technology, 15, Article 102725. https://doi.org/10.1016/j.atech.2026.102725

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102725

Keywords: IGE-SegNeXt, Brassica oleracea var. capitata, semantic segmentation, cabbage seedlings, plant disease detection, pest damage monitoring, IDCAN encoder, global context feature enhancement, channel-efficient multi-scale attention, multi-class focal loss, disease-symptom canopy proportion, pest-damage canopy proportion, SegNeXt baseline, agricultural computer vision, precision vegetable farming, paired cross-validation, leaf status classification, AnyLabeling dataset

News Source: Alan Morgan. (November 17, 2026). IGE-SegNeXt Model Segments Cabbage Leaf Damage but Shows Inconsistent Field Gains. Scienmag.

Tags: agricultural computer visionAnyLabeling datasetBrassica oleracea var. capitatacabbage seedlingschannel-efficient multi-scale attentiondisease-symptom canopy proportionglobal context feature enhancementIDCAN encoderIGE-SegNeXtleaf status classificationmulti-class focal losspaired cross-validationpest damage monitoringpest-damage canopy proportionplant disease detectionprecision vegetable farmingSegNeXt baselinesemantic segmentation
Share12Tweet7Share2ShareShareShare1

Related Posts

Brain Imaging and AI Support Staged Care for Treatment-Resistant Depression

Brain Imaging and AI Support Staged Care for Treatment-Resistant Depression

November 17, 2026
LexiCAL Framework Demonstrates Strong Encryption Distortion and Near-Perfect Image Recovery

LexiCAL Framework Demonstrates Strong Encryption Distortion and Near-Perfect Image Recovery

November 17, 2026

Onboard YOLO11n Model Completes Aircraft Walk-Around Video Processing Within Recording Duration

November 17, 2026

Machine Learning Predicts Heat Transfer in Magnetized Hybrid Nanofluid Flow

November 17, 2026

POPULAR NEWS

  • Sesbania Hay Replacement Modifies Lamb Muscle Lipids and Meat Flavor Precursors

    Sesbania Hay Replacement Modifies Lamb Muscle Lipids and Meat Flavor Precursors

    29 shares
    Share 12 Tweet 7
  • Human Guided Knowledge Fusion Enhances Industrial Fault Diagnosis Graphs and Retrieval

    29 shares
    Share 12 Tweet 7
  • Geostatistical Kriging Maps Regional Ionospheric Storm Anomalies Across Türkiye

    29 shares
    Share 12 Tweet 7
  • Pediatric Mononucleosis Study Finds Exploratory Links Between Viral Load and Laboratory Measures

    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

Intraoperative Hemoadsorption Linked to Reduced Mortality in High-Risk Endocarditis Surgery

Brain Imaging and AI Support Staged Care for Treatment-Resistant Depression

Wild Hedgehogs in Northwestern Italy Harbor Genetically Diverse mecC Methicillin-Resistant Bacteria

Subscribe to Blog via Email

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

Join 88 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.