Analysis and inversion of rice chlorophyll spectral characteristics using RNCA-PSO-ELM

Abstract: This study aims to explore the effective spectral feature selection and inversion modeling of rice chlorophyll content. Taking the Japonica rice as the research object, the Unmanned Aerial Vehicle (UAV) sensing was used to monitor the chlorophyll content at Shenyang Agricultural University Karima Rice Experiment Station in Northeast China in 2018-2020. The rice canopy hyperspectral data was also collected for the ground sample. A spectral feature selection was designed using the Regular Neighbor Component Analysis (RNCA), which was newly developed by the nearest NCA. A regularization term was then added to the loss function to reduce the risk of over-fitting. Different types of loss functions and regularization parameters were then tested to improve the feature extraction of RNCA. As such, the feature information was obtained from the hyperspectral to better predict the chlorophyll content of the rice plant. Subsequently, a chlorophyll inversion model was established using the input features. Extreme learning was also applied to build the model, due to its high accuracy and speed. Among them, the Extreme Learning Machine (ELM) was a feed-forward neural network with single or multiple hidden layers. Unlike the traditional back-propagation neural network, there was no need for reverse repetition, particularly after the parameters of the nodes were randomly set in the hidden layer of the ELM. Therefore, the amount of calculation was greatly reduced to learn and train the model faster than before. Furthermore, Particle Swarm Optimization (PSO) was selected to optimize the weights of the input layer and hidden layer deviations of ELM. The number of hidden layer nodes was automatically learned from the training data, instead of the randomness of the hidden layer node parameters of the ELM model. The generalization of the model was improved to reduce the number of hidden layer nodes for the target accuracy in practical applications. The results show that the RNCA presented the better feature selection. The best feature selection was achieved to initially select the 16 non-zero weight feature bands, particularly with the Mean Square Error (MSE) loss function and the regularization parameter of 0.306. According to the inversion accuracy of the chlorophyll ELM, the six characteristic bands with the highest weight were also selected: 710, 716, 508, 798, 532, and 708 nm, which were closely related to the chlorophyll content of rice. After that, the PSO was used to optimize the input weight and threshold deviation of the ELM model. An optimal combination was achieved, where the population size for the PSO orthogonal test, inertia weight, learning factors C1, C2, and velocity-position correlation coefficient were 50, 1.5, 1.3, 3.5, and 0.6, respectively. The RMSE and R2 of the chlorophyll inversion were 9.549 mg/L and 0.891 using RNCA-PSO-ELM, respectively. The findings can provide theoretical support to design the chlorophyll content sensor for the field-scale airborne rice