Scholarlink

Purpose <p>Traditional methods struggle to accurately and efficiently monitor rice leaf area index (LAI) throughout the entire growth period, hindering precision agriculture management.</p> Methods <p>To address this problem, this study integrated unmanned aerial vehicle (UAV) multi-spectral remote sensing with machine learning to develop a novel model for dynamic rice LAI monitoring in Zhejiang province, China, leveraging data from critical stages including tillering, jointing, and heading. Then, the normalized difference vegetation index (NDVI), the normalized green-red difference index (NGRDI), the normalized difference red edge index (NDRE), the red-edge ratio vegetation index (RERVI) and the enhanced vegetation index (EVI) were calculated. On this basis, linear and nonlinear models using single vegetation index were constructed separately. Additionally, two empirical models including the partial least squares regression (PLSR) and the ridge regression (RR), and four machine learning models including the random forest (RF), the support vector machine (SVM), the k-nearest neighbor (kNN) and the improved feed forward neural network (FNN) were developed utilizing five vegetation indices.</p> Results <p>The results indicated that among the traditional empirical models, the nonlinear model LAI-EVI<sub>non-linear</sub> performed the best, with R<sup>2</sup> of 0.756 and RMSE of 0.680, which also suggested that incorporating more vegetation indices as input data did not necessarily effectively improve the inversion accuracy by the PLSRand RR models. However, the performances of four machine learning models exhibited significant improvements. The LAIFNN model achievedthe highest accuracy, with R<sup>2</sup> and RMSE by about 0.821 and 0.583, respectively. The contribution of each input feature was also quantified byusing the SHapley Additive exPlanations (SHAP) method. Based on the optimal model, this study mapped the multi-temporal spatial patterns ofrice LAI.</p> Conclusion <p>These findings implied that the UAV-based remote sensing combined with high-performance machine learning offers apractical approach for LAI monitoring of the rice agro-ecosystem.</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

Machine learning-enhanced UAV multispectral remote sensing for monitoring rice leaf area index across critical growth stages

  • 期刊论文

  • 发表日期：

    2026年5月8日

查看全文
  • Chao Su,
  • Zaiying Ling,
  • Liwang Han,
  • Yanmin Shuai,
  • Jianguang Wen,
  • Xingwen Lin,
  • Haifeng Tang,
  • Alexandre Maniçoba da Rosa Ferraz Jardim,
  • Hoi Leong Lee,
  • Xuguang Tang

摘要

Purpose

Traditional methods struggle to accurately and efficiently monitor rice leaf area index (LAI) throughout the entire growth period, hindering precision agriculture management.

Methods

To address this problem, this study integrated unmanned aerial vehicle (UAV) multi-spectral remote sensing with machine learning to develop a novel model for dynamic rice LAI monitoring in Zhejiang province, China, leveraging data from critical stages including tillering, jointing, and heading. Then, the normalized difference vegetation index (NDVI), the normalized green-red difference index (NGRDI), the normalized difference red edge index (NDRE), the red-edge ratio vegetation index (RERVI) and the enhanced vegetation index (EVI) were calculated. On this basis, linear and nonlinear models using single vegetation index were constructed separately. Additionally, two empirical models including the partial least squares regression (PLSR) and the ridge regression (RR), and four machine learning models including the random forest (RF), the support vector machine (SVM), the k-nearest neighbor (kNN) and the improved feed forward neural network (FNN) were developed utilizing five vegetation indices.

Results

The results indicated that among the traditional empirical models, the nonlinear model LAI-EVInon-linear performed the best, with R2 of 0.756 and RMSE of 0.680, which also suggested that incorporating more vegetation indices as input data did not necessarily effectively improve the inversion accuracy by the PLSRand RR models. However, the performances of four machine learning models exhibited significant improvements. The LAIFNN model achievedthe highest accuracy, with R2 and RMSE by about 0.821 and 0.583, respectively. The contribution of each input feature was also quantified byusing the SHapley Additive exPlanations (SHAP) method. Based on the optimal model, this study mapped the multi-temporal spatial patterns ofrice LAI.

Conclusion

These findings implied that the UAV-based remote sensing combined with high-performance machine learning offers apractical approach for LAI monitoring of the rice agro-ecosystem.