The fusion of machine olfactory data and UV–Vis-NIR-MIR spectra enabled accurate prediction of key soil nutrients

Abstract

Conventional approaches for evaluating soil nutrients typically involved lengthy and resource-intensive analytical procedures, rendering them inadequate for large-scale and high-throughput testing. To address these limitations, this study proposed an innovative solution based on sensor data fusion to predict the content of key soil nutrients. The proposed methodology entailed collecting olfactory data after soil pyrolysis using gas sensors and spectral data from soil samples utilizing ultraviolet–visible-near infrared (UV–Vis-NIR) and mid-infrared (MIR) techniques. Three fusion strategies including series and parallel modes were designed to effectively amalgamate the gathered data and supplemented with machine learning algorithms to predict the content of key soil nutrients. Tested a testing set consisting of 33 soil samples. The findings demonstrated that introducing a self-attention procedure into the series splicing fusion strategy significantly improved the predictive performance. This highlights the synergistic benefits of integrating information from olfactory and spectral data sources. Predicting multiple nutrient contents within the framework of the multi-layer perceptron combined with random forest (MLP-RF) fusion model showed superior performance, with the coefficient of determination (R2) ranging from 0.80 to 0.96. The predictive validity for the content of fundamental nutrients and available nutrients in the soil can benefit from the combination of biological and structural information captured by olfactory data and chemical information provided by spectroscopy.

Keywords

• Soil nutrient content
• Multi-source data
• Fusion strategy
• Machine learning