Scholarlink

Purpose <p>Predicting grain yield and protein content of rice using multi-source remote sensing data is crucial for optimizing agricultural management and guiding harvest decisions. Although multi-source remote sensing data can characterize rice growth by integrating multi-dimensional remote sensing information, its feasibility and reliability for the prediction of grain yield and protein content remain unclear. This study aimed to investigate the potential and feasibility of using hyperspectral fusion images from hyperspectral images and RGB images to enhance grain yield and protein content prediction.</p> Methods <p>An image fusion algorithm was employed to pan-sharpen hyperspectral and RGB images, generating spatially resolved hyperspectral fusion images of 1–4&#xa0;cm pixel<sup>− 1</sup>. RGB images were used to enhance the spatial detail and textural information of hyperspectral images. Leaf nitrogen accumulation (LNA) was used as a bridge to develop a multi-source remote sensing data-driven prediction framework for grain yield and protein content. Machine learning regression algorithms were used to predict LNA by integrating the vegetation indices (VIs) and texture indices (TIs). The relationship between the LNA and grain yield and protein content was determined at the tillering, jointing, and full-heading stages. Multiple linear regression (MLR) models were then developed to predict grain yield and protein using multi-temporal LNAs. The most accurate LNA prediction model was embedded with the MLR models of grain yield and protein content.</p> Results <p>The hyperspectral fusion images of 1&#xa0;cm pixel<sup>− 1</sup> were the most stable and reliable for predicting rice LNA, exhibiting 21.3%, 12.3%, and 26.8% lower RMSE than the raw hyperspectral images at the tillering, jointing, and full-heading stages, respectively. Grain yield and protein content exhibited a quadratic relationship with LNA, and the MLR models of grain yield and protein content achieved the best prediction accuracy. Hyperspectral fusion images of 1&#xa0;cm pixel<sup>− 1</sup> obtained more accurate predictions of grain yield (RMSE = 0.279 t ha<sup>− 1</sup>, rRMSE = 2.996%, MAPE = 2.623%) and protein content (RMSE = 0.047 t ha<sup>− 1</sup>, rRMSE = 7.611%, MAPE = 6.676%) than the other hyperspectral fusion images.</p> Conclusion <p>Therefore, coupling multi-temporal LNAs with high-spatial-resolution hyperspectral fusion images could improve the prediction accuracy of grain yield and protein content.</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

Improving spatial resolutions of unmanned aerial vehicle hyperspectral images to predict grain yield and protein content in rice

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Zhonglin Wang,
  • Rong Hu,
  • Mingming Hu,
  • Huilai Yin,
  • Tao Liu,
  • Yufei Xie,
  • Yangming Ma,
  • Feng Yang,
  • Feijie Li,
  • Zhiyuan Yang,
  • Yongjian Sun,
  • Jun Ma

摘要

Purpose

Predicting grain yield and protein content of rice using multi-source remote sensing data is crucial for optimizing agricultural management and guiding harvest decisions. Although multi-source remote sensing data can characterize rice growth by integrating multi-dimensional remote sensing information, its feasibility and reliability for the prediction of grain yield and protein content remain unclear. This study aimed to investigate the potential and feasibility of using hyperspectral fusion images from hyperspectral images and RGB images to enhance grain yield and protein content prediction.

Methods

An image fusion algorithm was employed to pan-sharpen hyperspectral and RGB images, generating spatially resolved hyperspectral fusion images of 1–4 cm pixel− 1. RGB images were used to enhance the spatial detail and textural information of hyperspectral images. Leaf nitrogen accumulation (LNA) was used as a bridge to develop a multi-source remote sensing data-driven prediction framework for grain yield and protein content. Machine learning regression algorithms were used to predict LNA by integrating the vegetation indices (VIs) and texture indices (TIs). The relationship between the LNA and grain yield and protein content was determined at the tillering, jointing, and full-heading stages. Multiple linear regression (MLR) models were then developed to predict grain yield and protein using multi-temporal LNAs. The most accurate LNA prediction model was embedded with the MLR models of grain yield and protein content.

Results

The hyperspectral fusion images of 1 cm pixel− 1 were the most stable and reliable for predicting rice LNA, exhibiting 21.3%, 12.3%, and 26.8% lower RMSE than the raw hyperspectral images at the tillering, jointing, and full-heading stages, respectively. Grain yield and protein content exhibited a quadratic relationship with LNA, and the MLR models of grain yield and protein content achieved the best prediction accuracy. Hyperspectral fusion images of 1 cm pixel− 1 obtained more accurate predictions of grain yield (RMSE = 0.279 t ha− 1, rRMSE = 2.996%, MAPE = 2.623%) and protein content (RMSE = 0.047 t ha− 1, rRMSE = 7.611%, MAPE = 6.676%) than the other hyperspectral fusion images.

Conclusion

Therefore, coupling multi-temporal LNAs with high-spatial-resolution hyperspectral fusion images could improve the prediction accuracy of grain yield and protein content.