Mapping Plant Nitrogen Concentration and Aboveground Biomass of Potato Crops from Sentinel-2 Data Using Ensemble Learning Models

Abstract

Excessive nitrogen (N) fertilization poses environmental risks at regional and global levels. Satellite remote sensing provides a novel approach for large-scale N monitoring. In this study, we evaluated the performance of different types of spectral bands and indices (SIs) coupled with ensemble learning models (ELMs) at retrieving the plant N concentration (PNC) and plant aboveground biomass (AGB) of potato from Sentinel-2 images. Cloud-free Sentinel-2 imagery was acquired during the tuber-formation to starch-accumulation stages from 2020 to 2021. Fourteen optimal SIs were selected using the successive projections algorithm (SPA) and principal component analysis (PCA). The PNC and AGB estimation models were then built using an ELMs. The results showed that the SIs based on chlorophyll absorption bands were strongly related to potato PNC and AGB. Also, the N-correlated bands were mainly concentrated in the red-edge (705 nm) and short-wave infrared (1610 and 2190 nm) regions. The ELMs successfully predicted PNC and AGB (R2PNC = 0.74; R2AGB = 0.82). Compared with the other five base models (k-nearest neighbor (KNN), partial least squares regression (PLSR), support vector regression (SVR), random forest (RF), and Gaussian process regression (GPR)), the ELMs provided higher PNC and AGB estimation accuracy and effectively reduced overfitting to training data. This study demonstrated that the promising solution of using SPA-PCA coupled with an ensemble learning model improves the estimation accuracy of potato PNC and AGB based on Sentinel-2 imagery data.