AUTHOR=Long Teng , Che XiaoLiang , Guo Wenbin , Lan Yubin , Xie Ziran , Liu Wentao , Lv Jinsheng , Long Yongbing , Liu Tianyi , Zhao Jing TITLE=Visible-near-infrared hyperspectral imaging combined with ensemble learning for the nutrient content of Pinus elliottii × P. caribaea canopy needles detection JOURNAL=Frontiers in Forests and Global Change VOLUME=6 YEAR=2023 URL=https://www.frontiersin.org/journals/forests-and-global-change/articles/10.3389/ffgc.2023.1203626 DOI=10.3389/ffgc.2023.1203626 ISSN=2624-893X ABSTRACT=Introduction

Pinus elliottii × P. caribaea is one of the major tree species in commercial forest bases in developed countries. However, in the process of sapling cultivation, nutrients cannot be accurately detected and supplied to individual saplings, resulting in reduced yield and quality.

Methods

In this paper, visible-near-infrared (Vis-NIR) hyperspectral imaging (HSI) combined with ensemble learning (EL) was used to solve this problem. The content and distribution of nitrogen (N), phosphorus (P), and potassium (K) in the canopy needles of Pinus elliottii × P. caribaea saplings were obtained through HSI data analysis, and the nutritional needs of individual plants were reflected to provide a basis for nutritional supply decisions. The saplings were treated with deficient, sufficient, and excessive N, P, and K single-element fertilizers. After collecting the Vis-NIR hyperspectral images of these saplings, a variety of pre-processing, feature selection, and ensemble learning algorithms were used to establish predictive models. The R2 and RMSE were used to evaluate the performance of the prediction models.

Results

The results showed that the multiple scattering correction-competitive adaptive reweighted sampling-Stacking (MSC-CARS-Stacking) model had the best results among the three nutrient elements prediction models (Rp2-N = 0.833, RMSEP = 0.380; Rp2-P = 0.622, RMSEP = 0.101; Rp2-K = 0.697, RMSEP = 0.523). When studying the sensitive bands of N, P, and K, we found that the common characteristic wavelengths were 675.3 and 923.9 nm, while the non-common characteristic wavelengths were located at 550 nm (green peak), 680 nm (red valley), and 960 nm (water peak). In studying the generalization ability of the model, only the nitrogen group data were used to train the MSC-CARS-Stacking model for nitrogen prediction, which was then used to predict the nitrogen content in the phosphorus and potassium groups, obtaining good results (Rc2-N = 0.841, Rp2-P = 0.814, Rp2-K = 0.801). It showed a strong generalization ability for the prediction of nitrogen, and similarly, phosphorus and potassium.

Discussion

In conclusion, this study verifies that the Vis-NIR HSI combined with EL is indeed a reliable and stable method to predict the contents of N, P, and K in the needles of Pinus elliottii × P. caribaea sapling canopy.