March 12, 2020
Institute of Electrical and Electronics Engineers

Mining spectral libraries with machine learning for soil sensing in low resource settings

Wenbo Wang, Liming Hu, James W. Stafford, Marie Connett, Benjamin K. Wilson, Matthew D. Keller

Smallholder farmers and extension workers in sub-Saharan Africa have limited access to the soil information needed to improve crop yields, and therefore boost incomes and improve food security. A promising approach to provide this information at farm level is a low-cost optical spectroscopy device and an accompanying decision support application. This paper describes our efforts to explore spectral libraries, both publicly accessible and acquired through our own efforts, to define the instrument specifications and data handling procedures for such an approach. Spectral ranges throughout the mid-infrared (MIR) and near-infrared (NIR) regions and spectral resolution pertinent to available hardware technologies were evaluated. Given the complex physical-chemical properties of soil samples, numerous preprocessing strategies were applied prior to regression with partial least squares, support vector machine, and multilayer perceptron neural network (MLPNN) models for predicting relevant soil properties. The experiments with existing spectral datasets demonstrated that the MLPNN approach outperformed other approaches most consistently for soil property prediction, and that the MIR region offers better performance than NIR, even with truncated spectral ranges and degraded spectral resolutions. At this time, however, there are hardware limitations for MIR systems that prevent a low cost device from collecting acceptable signals, thus making an NIR solution more appealing overall.

View Journal