Estimation of leaf nitrogen concentration in wheat using the MK-SVR algorithm and satellite remote sensing data

Abstract

The appropriate spectral vegetation indices can be used to rapidly and non-destructively estimate the leaf nitrogen concentration (LNC) in wheat for on-farm wheat management. However, the accuracy of estimation should be further improved. Previous studies focused on developing vegetation indices, but research about modeling algorithms were limited. In this study, multiple-kernel support vector regression (MK-SVR) was used to assess the LNC in wheat based on satellite remote sensing data. The objectives of this study were to (1) investigate the applicability of the MK-SVR algorithm for remotely estimating the LNC in wheat, (2) test the performance of the MK-SVR regression model, and (3) compare the performance of the MK-SVR algorithm with multiple linear regression (MLR), partial least squares (PLS), artificial neural networks (ANNs), and single-kernel SVR (SK-SVR) algorithms for wheat LNC estimation. In-situ LNC data over four years at different sites in Jiangsu Province of China were measured during the jointing, booting, and anthesis stages; one HJ-CCD image of wheat was obtained during each stage. Vegetation indices were calculated based on these images, and correlations between vegetation indices and LNC data were measured. Finally, a MK-SVR model whose inputs were vegetation indices was established to estimate the LNC during each stage. The results showed that the MK-SVR model performed well in estimating LNC. The coefficients of determination (R2) of the estimated-versus-measured LNC values for the three stages were respectively 0.73, 0.82, and 0.75, meanwhile, the corresponding root mean square errors (RMSE) and the relative RMSE were respectively 0.13 and 6.6%, 0.21 and 7.7%, and 0.20 and 6.5%. Thus, the MK-SVR algorithm provides an effective way to improve the prediction accuracy of LNC in wheat on a large scale.