Improving the transferability of potato nitrogen concentration estimation models based on hybrid feature selection and Gaussian process regression

Abstract

Feature selection methods are widely used to improve the performance of plant nitrogen concentration (PNC) estimation models. However, the performance of individual feature selection methods can vary across different environments due to various uncertainties. This study aimed to propose a hybrid feature selection method to accurately identify the sensitive bands for the PNC estimation. Field experiments with different potato cultivars and N treatments were carried out in the Inner Mongolia during 2018, 2019, and 2021. The results showed that the hybrid feature selection method can effectively identify the sensitive bands for PNC. When combined with variational heteroscedastic Gaussian process regression (VHGPR), the hybrid method significantly improves the prediction accuracy of potato PNC. Validation using an independent dataset demonstrated that the hybrid feature selection method achieved the highest prediction accuracy compared to traditional feature selection methods, with the mean coefficient of determination (R²) increasing by 16.27 %. Additionally, the performance of VHGPR was benchmarked against partial least squares regression (PLSR). The results indicated that the VHGPR model outperforms the PLSR model across various data types, with a mean R² improvement of 8.92 %. In conclusion, combining the hybrid feature selection method with VHGPR can facilitate real-time PNC estimation in the field, thereby assisting farmers in accurately applying nitrogen fertilization strategies.