A deep learning-based method for screening soil total nitrogen characteristic wavelengths

Abstract

Spectroscopy technology is one of the most important methods for non-destructive and rapid detection of soil total nitrogen (STN) content. It would be convenient and practical for developing a low-cost portable detector of STN content based on spectroscopy if limited wavelengths could be singled out as soil nitrogen characteristic wavelengths. This paper proposed a deep learning-based method for STN characteristic wavelength screening by using a public dataset called LUCAS Soil. First, a multi-channel convolutional network incorporated Inception model was selected as the benchmark model for STN prediction because of its better performance than other models. Then, the absorbance of each wavelength of the full spectra (4200 wavelengths) was sequentially all set to zeros, and 4200 new spectra datasets were formed accordingly. Each new dataset still consisted of 4200 wavelengths, but the absorbance at one of these wavelengths was all set to zeros. The benchmark model was used to carry out STN predictions through each new dataset. In order to assess the effectiveness of STN prediction, an evaluation standard was established. Finally, according to the evaluation scores, the top highest 8–50 STN characteristic wavelengths were tentatively extracted from 4200 wavelengths (400–2499.5 nm). In the process of determining the benchmark model, three other predictive methods were carefully investigated besides the above deep learning-based model, which included ordinary least square estimation (OLSE), random forest (RF) and extreme learning machine (ELM). The results indicated that the deep learning-based model for STN prediction worked the best on large-scale and high-dimensional spectroscopic data (coefficients of determination (R2) = 0.93, root mean square error of prediction (RMSEP) = 0.97 g/kg, and residual prediction deviation (RPD) = 3.85), and accordingly it was chosen as the benchmark model. In addition, in order to verify the effectiveness of selected feature wavelengths, STN prediction models were established by using the top 8 to 50 of the STN characteristic wavelengths based on the above four different predictive methods. It was found that the ELM model with 21 characteristic wavelengths (R2 = 0.85, RMSEP = 2.05 g/kg, and RPD = 1.83) performed better than that with full-wavebands (R2 = 0.82, RMSEP = 2.49 g/kg, and RPD = 1.50). The accuracies of other models based on characteristic wavelengths were somewhat lower than that of the full-spectrum model, but it still was practical (R2 >= 0.57), which showed that the soil nitrogen characteristic wavelength screening method proposed in this paper was indeed effective, and the STN prediction model with a limited number of wavelengths could be established based on these STN characteristic wavelengths.