Optimizing soil carbon content prediction performance by multi-band feature fusion based on visible near-infrared spectroscopy.

Purpose: The spectral characteristic information of soil carbon is vulnerable to interference because of the complex soil composition. To avoid the interference of useless information and improve the prediction performance of soil carbon content, feature extraction by multi-band feature fusion based on visible near infrared spectroscopy is studied. Materials and methods: Ocean Optics QE65000 spectrometer was used to obtain visible near-infrared spectral data (200–1100 nm). The experimental samples were from Qingdao Inland, Aoshan Bay, and Jiaozhou Bay, which was to verify the feasibility of the proposed method. Voting mechanism multi-band feature fusion method and neighborhood adaptive multi-band feature fusion method were proposed for feature extraction. Generated adversarial network (GAN) was introduced into multi-band feature fusion to expand training samples. After feature extraction and training sample expansion, the estimation models of soil carbon content were established by partial least square regression (PLSR). The coefficient of determination (R_p²), root mean square error (RMSEP), and relative percent deviation (RPD) were used as the evaluation criteria to measure the quality of the model. Results and discussion: The results showed that the prediction accuracy of soil carbon content had been significantly improved by voting mechanism multi-band feature fusion method in Inland, Aoshan Bay, and Jiaozhou Bay (RPD = 3.959, RPD = 2.685, RPD = 2.108), and the RPD value increased by 70.50%, 38.33%, and 25.40%, respectively. Due to the wide distribution of carbon content in soil samples collected Inland, therefore the RPD was higher compared to others. Through the neighborhood adaptive multi-band feature fusion method, the prediction accuracy of Inland and Aoshan Bay was further improved (RPD = 4.099, RPD = 2.713). After training sample expansion by GAN, the three sample plots achieved the estimation of soil carbon content under a small number of samples. Conclusions: Multi-band feature fusion method fused multiple single feature band extraction algorithms to fully exploit the complementarity among multiple features and extract effective feature bands to improve the accuracy and stability of soil carbon content prediction. [ABSTRACT FROM AUTHOR]