Building a near-infrared (NIR) soil spectral dataset and predictive machine learning models using a handheld NIR spectrophotometer

Abstract

This near-infrared spectral dataset consists of 2,106 diverse mineral soil samples scanned, on average, on six different units of the same low-cost commercially available handheld spectrophotometer. Most soil samples were selected from the USDA NRCS National Soil Survey Center-Kellogg Soil Survey Laboratory (NSSC-KSSL) soil archives to represent the diversity of mineral soils (0–30 cm) found in the United States, while 90 samples were selected from Ghana, Kenya, and Nigeria to represent available African soils in the same archive. All scanning was performed on dried and sieved (<2 mm) soil samples. Machine learning predictive models were developed for soil organic carbon (SOC), pH, bulk density (BD), carbonate (CaCO3), exchangeable potassium (Ex. K), sand, silt, and clay content from their spectra in the R programming language using most of this dataset (1,976 US soils) and are included in this data release. Two model types, Cubist and partial least squares regression (PLSR) were developed using two strategies: (1) using an average of the spectral scans across devices for each sample and, (2) using the replicate spectral scans across devices for each sample. We present the internal performance of these models here. The dry spectra and Cubist models for these soil properties are available for download from 10.5281/zenodo.7586621. An example of detailed code used to produce these models is hosted at the Open Soil Spectral Library, a free service of the Soil Spectroscopy for the Global Good Network (soilspectroscopy.org), enabling broad use of these data for multiple soil monitoring applications.