Scholarlink

<p>Crop diseases and insect pests cause annual yield losses of 20–40&#xa0;% worldwide, yet conventional field scouting is subjective, labour-intensive, and detects stress only after symptoms appear. We propose a unified UAV hyperspectral–wavelet pipeline for early disease and pest detection that fuses three complementary feature streams: vegetation indices (VIs), one-dimensional continuous wavelet transform (1D&#xa0;CWT) coefficients of the spectral signature at seven scales over 400–900&#xa0;nm, and two-dimensional discrete wavelet transform (2D&#xa0;DWT) texture descriptors. The framework is validated on three public benchmarks—UMN wheat hyperspectral (1,021 plots), New Plant Diseases (87,848 RGB images), and PlantDoc (2,598 field images)—and yields consistent gains of 12–15 percentage points (pp) in overall accuracy over VI-only baselines. Best configurations reach 88.2&#xa0;% on wheat stress classification, 96.7&#xa0;% on controlled-environment leaf disease, and 91.8&#xa0;% on the more challenging field benchmark. Ablation analysis shows wavelet features contribute 9.9–12.5&#xa0;pp of cumulative gain, with red-edge CWT coefficients (720&#xa0;nm, scales 16–32) being the most discriminative descriptors. End-to-end processing of 95&#xa0;min/ha on commodity hardware (USD 3,000–5,000) supports same-day operational analysis. The framework offers a generalisable and scalable solution for precision agriculture across diverse crops and conditions.</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

Integration of UAV hyperspectral remote sensing and wavelet algorithms for crop disease and pest detection

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Dufen Gan,
  • Zhongwen Zhang

摘要

Crop diseases and insect pests cause annual yield losses of 20–40 % worldwide, yet conventional field scouting is subjective, labour-intensive, and detects stress only after symptoms appear. We propose a unified UAV hyperspectral–wavelet pipeline for early disease and pest detection that fuses three complementary feature streams: vegetation indices (VIs), one-dimensional continuous wavelet transform (1D CWT) coefficients of the spectral signature at seven scales over 400–900 nm, and two-dimensional discrete wavelet transform (2D DWT) texture descriptors. The framework is validated on three public benchmarks—UMN wheat hyperspectral (1,021 plots), New Plant Diseases (87,848 RGB images), and PlantDoc (2,598 field images)—and yields consistent gains of 12–15 percentage points (pp) in overall accuracy over VI-only baselines. Best configurations reach 88.2 % on wheat stress classification, 96.7 % on controlled-environment leaf disease, and 91.8 % on the more challenging field benchmark. Ablation analysis shows wavelet features contribute 9.9–12.5 pp of cumulative gain, with red-edge CWT coefficients (720 nm, scales 16–32) being the most discriminative descriptors. End-to-end processing of 95 min/ha on commodity hardware (USD 3,000–5,000) supports same-day operational analysis. The framework offers a generalisable and scalable solution for precision agriculture across diverse crops and conditions.