Achieving joint calibration of soil Vis-NIR spectra across instruments, soil types and properties by an attention-based spectra encoding-spectra/property decoding architecture

Abstract

Due to the differences in instruments and measurement methods, soil visible-near infrared (Vis-NIR) spectral data collected worldwide cannot be combined to provide consistent calibration models, reducing the benefit of producing soil spectra libraries. We propose an attention-based spectra encoding-spectra/property decoding model for soil Vis-NIR spectral analysis to solve this problem. By introducing the attention mechanism, the model can use the same parameters to achieve a joint calibration regardless of the length of the input spectral data and the type of soil samples. The model design pays attention to wavelengths combination, which accords to the essence of Vis-NIR analysis. Three datasets, including the LUCAS, Brazilian, and Kenyan datasets, were used to train and test the attention model. An alternative convolution neural network (CNN) encoding architecture that allows joint calibration was also introduced. Comparing with conventional partial least squares regression (PLSR) and CNN encoding model, the present attention model obtained the best results on all three datasets. This is the first time that an attention-based model has been applied to soil Vis-NIR spectra analysis. The model can process spectra data with different sizes and focus on the relation between different wavelengths, providing a joint calibration. It has the potential to become a “one model for all” in soil Vis-NIR spectra research.