Faster Machine Learning for MIR Soil Spectroscopy with Discrete Haar Wavelet Transform

Although soil data is critical for research and monitoring efforts, wet chemistry methods are often labor intensive and expensive. Mid-infrared (MIR) spectroscopy is a cost effective method of inferring some soil properties using predictive machine learning. Predictions require instrument specific calibration libraries where wet chemistry data are paired with MIR spectra. Spectral data are generally large and thus training the machine learning models can be computationally intensive. Therefore, reducing the spectral signal size allows for more efficient model development, enabling exploration of new methods.

In this study we investigated multiresolution Haar (discrete) wavelet compression as a pre-processing step to reduce signal size in the context of a calibration library and machine learning training.

We predicted soil organic carbon fraction (OC) using the original and Haar transformed MIR spectra (n > 60,000; from the USDA NRCS National Soil Survey Center - Kellogg Soil Survey Laboratory library) with a partial least squares regression (PLSR) machine learning method. As expected, wavelet compression reduced the size of the spectral data and resulting runtime, and we saw a trade-off between linear fit of predicted and measured OC. The R2 value (linear fit between predicted and measured OC fraction) did not decrease significantly until a 4-fold compression ratio (6% of the original signal) was imposed. Unexpectedly, we also saw a reduction in root mean squared error as compression increased.

The discrete Haar wavelet transform is a promising preconditioner to increase computational efficiency of PLSR or other machine learning methods, such as regression tree based algorithms. While more study is needed to determine the exact compression cut-off, discrete wavelet transformations could allow soil scientists to study properties faster at larger spatiotemporal scales, potentially improving soil health and soil carbon monitoring efforts.