Scholarlink

<p>Date palm cultivation faces annual economic losses from diseases caused by fungal pathogens (<i>Fusarium</i>), bacteria (<i>Erwinia</i>), viruses, and pests (<i>Rhynchophorus</i>), challenging visual diagnosis under field conditions. Existing deep learning approaches lack actionable outputs like localization and severity estimation required for precision agriculture. We present ViT-AdvancedDiseaseNet (ViT-ADN), a multi-task Vision Transformer framework delivering simultaneous classification, localization, and severity estimation. Evaluated on the rigorously annotated Al-Ahsa-Palm-Set dataset (6,665 field images, 8 classes, expert-validated), ViT-ADN achieves 96.83% accuracy, 89.12% IoU, and 0.15 MAE. The model integrates hierarchical feature extraction, masked autoencoding pretraining (50% masking), and adversarial training (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\epsilon \)</EquationSource> <EquationSource Format="MATHML"><math> <mi>ϵ</mi> </math></EquationSource> </InlineEquation>=0.03) yielding 93.41% robustness against field perturbations. We transparently document limitations: zero-shot generalization fails under domain shift (76.42% accuracy), requiring few-shot adaptation (20 samples/class for 92.87%); 12.5 GB memory necessitates cloud-assisted architecture; and the memory module functions primarily as parameter-efficient capacity expansion. ViT-ADN advances agricultural AI from academic benchmarks toward clinically validated, ethically grounded deployment where performance metrics translate to farmer livelihoods and environmental sustainability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-Task Vision Transformer for Date Palm Disease Analysis with Localization and Severity Estimation

  • 期刊论文

  • 发表日期：

    2026年5月15日

查看全文
  • Ben Othman Soufiane

摘要

Date palm cultivation faces annual economic losses from diseases caused by fungal pathogens (Fusarium), bacteria (Erwinia), viruses, and pests (Rhynchophorus), challenging visual diagnosis under field conditions. Existing deep learning approaches lack actionable outputs like localization and severity estimation required for precision agriculture. We present ViT-AdvancedDiseaseNet (ViT-ADN), a multi-task Vision Transformer framework delivering simultaneous classification, localization, and severity estimation. Evaluated on the rigorously annotated Al-Ahsa-Palm-Set dataset (6,665 field images, 8 classes, expert-validated), ViT-ADN achieves 96.83% accuracy, 89.12% IoU, and 0.15 MAE. The model integrates hierarchical feature extraction, masked autoencoding pretraining (50% masking), and adversarial training ( \(\epsilon \) ϵ =0.03) yielding 93.41% robustness against field perturbations. We transparently document limitations: zero-shot generalization fails under domain shift (76.42% accuracy), requiring few-shot adaptation (20 samples/class for 92.87%); 12.5 GB memory necessitates cloud-assisted architecture; and the memory module functions primarily as parameter-efficient capacity expansion. ViT-ADN advances agricultural AI from academic benchmarks toward clinically validated, ethically grounded deployment where performance metrics translate to farmer livelihoods and environmental sustainability.