Moisture content classification of soil and stalk residue samples from spectral data using machine learning

Remotely sensed spectral data are commonly used to quantify material properties in agricultural applications. Typically, a few distinct bands are selected and formulated into a reflectance index. Machine learning presents an alternative approach for quantifying material properties from spectral data due to the ease at which it can be used to process large datasets. This study aimed to test several commercially available machine learning algorithms using spectral data collected from moisture-controlled silt-loam soil and wheat stalk residue samples. The spectral data used in this analysis were previously used to develop a normalized difference water index (NDWI) for remotely quantifying the moisture content of agricultural background materials by selecting a pair of narrow-band wavelengths. However, results showed mixed performance for index-based processing. In this study, raw spectral data were preprocessed using partial least squares (PLS) regression to optimize the number of input components. The PLS components were fed into 20 different machine learning algorithms available in MATLAB, and the best two performing methods were compared to the index-based method. Cubic support vector machine (SVM) and ensemble bagged trees methods produced the highest composite prediction accuracies of 96% and 93%, respectively, for silt-loam soil samples and 86% and 93%, respectively, for wheat stalk residue samples. Prediction accuracy using the index-based method was 86% for silt-loam soil and 30% for wheat stalk residue. A potential limitation of both machine learning methods was the discrete classification of moisture content rather than the continuous output of the index-based method. However, the substantial improvement in prediction accuracy of individual samples likely outweighs concerns with limited precision.