Frontiers in Plant Science (Dec 2024)

Mapping rapeseed (Brassica napus L.) aboveground biomass in different periods using optical and phenotypic metrics derived from UAV hyperspectral and RGB imagery

  • Chuanliang Sun,
  • Chuanliang Sun,
  • Weixin Zhang,
  • Genping Zhao,
  • Qian Wu,
  • Wanjie Liang,
  • Ni Ren,
  • Hongxin Cao,
  • Lidong Zou,
  • Lidong Zou

DOI
https://doi.org/10.3389/fpls.2024.1504119
Journal volume & issue
Vol. 15

Abstract

Read online

Aboveground biomass (AGB) is a key indicator of crop nutrition and growth status. Accurately and timely obtaining biomass information is essential for crop yield prediction in precision management systems. Remote sensing methods play a key role in monitoring crop biomass. However, the saturation effect makes it challenging for spectral indices to accurately reflect crop changes at higher biomass levels. It is well established that rapeseed biomass during different growth stages is closely related to phenotypic traits. This study aims to explore the potential of using optical and phenotypic metrics to estimate rapeseed AGB. Vegetation indices (VI), texture features (TF), and structural features (SF) were extracted from UAV hyperspectral and ultra-high-resolution RGB images to assess their correlation with rapeseed biomass at different growth stages. Deep neural network (DNN), random forest (RF), and support vector regression (SVR) were employed to estimate rapeseed AGB. We compared the accuracy of various feature combinations and evaluated model performance at different growth stages. The results indicated strong correlations between rapeseed AGB at the three growth stages and the corresponding indices. The estimation model incorporating VI, TF, and SF showed higher accuracy in estimating rapeseed AGB compared to models using individual feature sets. Furthermore, the DNN model (R2 = 0.878, RMSE = 447.02 kg/ha) with the combined features outperformed both the RF (R2 = 0.812, RMSE = 530.15 kg/ha) and SVR (R2 = 0.781, RMSE = 563.24 kg/ha) models. Among the growth stages, the bolting stage yielded slightly higher estimation accuracy than the seedling and early blossoming stages. The optimal model combined DNN with VI, TF, and SF features. These findings demonstrate that integrating hyperspectral and RGB data with advanced artificial intelligence models, particularly DNN, provides an effective approach for estimating rapeseed AGB.

Keywords