Estimating the total nitrogen content of Aquilaria sinensis leaves based on a hybrid feature selection algorithm and image data from a modified digital camera

Abstract

With the development of imaging devices and image processing algorithms, numerous features have come to be used for the estimation of total nitrogen content (TNC) in plants. However, higher-dimensional inputs contain more correlated variables that can detrimentally affect model performance. In this study, a hybrid feature selection approach was developed for TNC estimation in Aquilaria sinensis. A low-cost modified digital camera with external filters was used to capture canopy images. Three feature selection methods, namely, random forest (RF), Pearson correlation coefficient (PCC)-based feature selection, and sequential backward selection (SBS), were combined into two hybrid feature selection algorithms (RF_SBS and PCC_SBS). In addition, three regression algorithms were used in hybrid feature selection process: random forest regression (RFR), support vector regression (SVR), and partial least squares regression (PLSR). The hybrid feature selection process consists of two steps. First, the lowest number of dimensions is sought based on the feature ranking. Then, SBS is used to find the best feature combinations. Compared with the original models, the R2 values of the RF-SBS-based models are improved by 0.094 (RF_SBS_RFR), 0.190 (RF_SBS_SVR), and 0.116 (RF_SBS_PLSR), while the R2 values of the PCC-SBS-based models are improved by 0.055 (PCC_SBS_RFR), 0.092 (PCC_SBS_SVR) and 0.128 (PCC_SBS_PLSR). Finally, the two best TNC estimation models are found to be PCC_SBS_PLSR, with an R2 of 0.863, and RF_SBS_SVR, with an R2 of 0.872. The proposed hybrid feature selection approach not only has great capacity to improve estimation accuracy but also can reduce model complexity by choosing the best feature subset.