Using Local Spectral Libraries to Create Prediction Models for Soil Physical Property Assessment in Washington State

Soil compaction is a major and extensive problem within agriculture, and we do not have standardized way to measure compaction or its effects. We can use tools such as penetrometers for rapid assessment of soil compaction; however, these measurements require data on soil organic carbon, clay content, and water content. This study was conducted to set the groundwork for visible and near infrared (VisNIR) as well as mid-infrared (MIR) spectrometers to measure multiple soil properties with one measurement. To evaluate MIR spectroscopy, I evaluated the impact of local vs. statewide spectral libraries for predicting organic carbon and clay content using lab analyzed soils from Washington State University and a larger statewide dataset derived from the Kellogg Soil Survey Laboratory. Additionally, partial least square models were compared with commercial Quant II software. To assess VisNIR spectroscopy, models were built using partial least squares regression and tested with and without external parameter orthogonalization, which is a method to correct for soil moisture. MIR prediction models are accurate at organic carbon prediction at all levels although the accuracy decreases from statewide to local soil samples. Bootstrapped R code models outperformed those developed with external parameter orthogonalization treatment improved prediction accuracy for wet-intact samples, although further testing is needed to address variability caused by moisture content. Ensuring each spectral library has a wide range of soil properties will improve prediction models. VisNIR prediction models are accurate for both clay content and organic carbon at a local level. The development of robust, regionally adaptable spectral libraries and standardized scanning protocols is essential for scaling these approaches to statewide applications.