Scholarlink

<p>Accurate rice disease detection from thermal imagery faces challenges including limited training data, image artifacts, and low thermal contrast hindering precision agriculture deployment. This study introduces AVC-GAN, a novel generative adversarial framework integrating Adaptive Variant Connectivity (AVC) with dynamic multi-component loss functions for enhanced thermal image augmentation in rice disease detection. The AVC module dynamically modulates pixel connectivity based on local thermal gradients in rice leaf structures, improving structural coherence, while the adaptive meta-heuristic loss function incorporates temporal coherence and spatial consistency through performance-based weight adjustment tailored to rice thermal signatures. Evaluation on rice thermal datasets demonstrates state-of-the-art performance with superior metrics (FID: 6.21, IS: 8.97, SSIM: 0.912, LPIPS: 0.087) compared to 15 baseline models. Integration into downstream rice disease detection models consistently enhances classification performance across architectures, with accuracy improvements of 3.3–6.2% (MaxViT: 95.6–98.9%, EfficientNetV2-S: 88.7–94.1%). Ablation studies confirm individual component contributions, with AVC providing 20.7% FID improvement. Cross-domain validation on wheat, tomato, and potato datasets achieves 90.4% average transfer accuracy with robust geographic (91.6%) and temporal stability (coefficient of variation &lt; 5%). Statistical validation confirms significant rice disease detection improvements (<i>p</i> &lt; 0.001), establishing AVC-GAN as an effective rice-focused solution with demonstrated broader agricultural applicability.</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

Cross-domain thermal image augmentation for agricultural disease detection: a multi-crop validated framework with adaptive connectivity and optimized loss function

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Heba M. Khalil,
  • Rahmah N. AlQthanin,
  • Mostafa Elbaz,
  • Mohamed Loey

摘要

Accurate rice disease detection from thermal imagery faces challenges including limited training data, image artifacts, and low thermal contrast hindering precision agriculture deployment. This study introduces AVC-GAN, a novel generative adversarial framework integrating Adaptive Variant Connectivity (AVC) with dynamic multi-component loss functions for enhanced thermal image augmentation in rice disease detection. The AVC module dynamically modulates pixel connectivity based on local thermal gradients in rice leaf structures, improving structural coherence, while the adaptive meta-heuristic loss function incorporates temporal coherence and spatial consistency through performance-based weight adjustment tailored to rice thermal signatures. Evaluation on rice thermal datasets demonstrates state-of-the-art performance with superior metrics (FID: 6.21, IS: 8.97, SSIM: 0.912, LPIPS: 0.087) compared to 15 baseline models. Integration into downstream rice disease detection models consistently enhances classification performance across architectures, with accuracy improvements of 3.3–6.2% (MaxViT: 95.6–98.9%, EfficientNetV2-S: 88.7–94.1%). Ablation studies confirm individual component contributions, with AVC providing 20.7% FID improvement. Cross-domain validation on wheat, tomato, and potato datasets achieves 90.4% average transfer accuracy with robust geographic (91.6%) and temporal stability (coefficient of variation < 5%). Statistical validation confirms significant rice disease detection improvements (p < 0.001), establishing AVC-GAN as an effective rice-focused solution with demonstrated broader agricultural applicability.