Document Type

Honors Project On-Campus Access Only

Abstract

Visible to near-infrared (Vis/NIR) reflectance spectroscopy offers a rapid, non-destructive alternative to soil analysis, prompting the creation of globally-trained machine learning models for spectral prediction. However, the accuracy of globally-trained models in underrepresented regions has not been widely researched. This study evaluated the performance of a locally calibrated Cubist regression model trained on Vis/NIR spectra from 775 soil samples collected  across South Africa and Kenya, and compared its predictive accuracy against the Open Soil Spectral Library (OSSL) estimation engine. The locally trained model was built using the same preprocessing, training, and testing procedures as the OSSL, but due to its training on spatially representative samples, it substantially outperformed the global model in predictions of several key soil properties, including soil organic carbon (C), pH, nitrogen (N), texture (silt, sand, and clay), and δ13C and δ15N. The local model achieved R2 values of up to 0.93 for clay, while the OSSL model had much lower accuracy for all traits due to its inability to generalize when predicting samples from unseen locations. Notably, the local model also produced strong predictions for δ13C and δ15N isotope ratios (0.75 and 0.81 respectively), and successfully predicted environmental variables including vegetation indices and climatic variables (R2 up to 0.93). Leave-one-spatial-cluster-out cross-validation, however, revealed poor transferability across geographically distinct locations. These results demonstrate that while locally calibrated models show strong predictive power, meaningful geographic generalizability remains limited without adequate spatial representation within training data. Therefore, expanding spectral libraries to capture regional specificities, and standardizing operating procedures will be critical to making Vis/NIR soil spectroscopy a globally reliable and accessible tool.

Download

Share

COinS
 
 

© Copyright is owned by author of this document