Construction of complex features for predicting soil total nitrogen content based on convolution operations

Abstract

On-demand fertilization does not only help to improve fertilizer use efficiency, but also to avoid over-use of chemical fertilizers. Rapid monitoring of the nutrient content constitutes the initial step of on-demand fertilization, and spectroscopy technology is considered as one of the most ideal nondestructive methods to detect the nutrient content. The typical soil spectra contain thousands of wavelengths, these spectral variables often contribute to collinearity and redundancies rather than relevant effective information. To resolve this issue, this paper selected soil total nitrogen (STN) content as the study subject, and the studies were carried out from two aspects, STN characteristic wavelengths screening and STN content prediction model building based on the limited number of independent variables. For the characteristic waveband screening process, four feature extraction methods (including F-test, mutual information, embedded method and deconvolution operation) were used and the results were compared and analyzed. The embedded method was selected as the benchmark method due to its advantages of simple, intuitive and reliable to screen the STN characteristic wavelengths. In the process of building the STN content prediction model, under-fitting problem is a major challenge which is driven by the limitation of soil nitrogen characteristic wavelengths. Towards this, this study proposed a method for constructing complex features for predicting STN content based on convolution operations which shows higher accuracy than those based on Multi-Layer Perceptron Neural Network and polynomial kernel functions. When the number of characteristic wavelengths is 16, the coefficient of determination (R2) is 0.69, root mean square error of prediction (RMSEP) is 4.34 g/kg, and the residual prediction deviation (RPD) is 0.86. After construction with convolution operations (256 features are calculated from these original 16 characteristic wavelengths), the R2 of the model reaches 0.86, the RMSEP is 1.98 g/kg, and the RPD is 1.89. This study provides a practical approach to extract target characteristics from soil spectra and enhance the relevant information for detecting STN based on the raw soil spectra.