Biological Systems Engineering

 

First Advisor

Yufeng Ge

Date of this Version

Summer 6-2016

Citation

Wijewardane, N. K. (2016). Using a VNIR Spectral Library to Model Soil Carbon and Total Nitrogen Content. (Agricultural and Biological Systems Engineering MS), University of Nebraska-Lincoln, NE, USA.

Comments

A THESIS Presented to the Faculty of The Graduate College at the University of Nebraska In Partial Fulfilment of Requirements For the Degree of Master of Science, Major: Agricultural and Biological Systems Engineering, Under the Supervision of Professor Yufeng Ge. Lincoln, Nebraska: June, 2016

Copyright (c) 2016 Nuwan K. Wijewardane

Abstract

n-situ soil sensor systems based on visible and near infrared spectroscopy is not yet been effectively used due to inadequate studies to utilize legacy spectral libraries under the field conditions. The performance of such systems is significantly affected by spectral discrepancies created by sample intactness and library differences. In this study, four objectives were devised to obtain directives to address these issues. The first objective was to calibrate and evaluate VNIR models statistically and computationally (i.e. computing resource requirement), using four modeling techniques namely: Partial least squares regression (PLS), Artificial neural networks (ANN), Random forests (RF) and Support vector regression (SVR), to predict soil carbon and nitrogen contents for the Rapid Carbon Assessment (RaCA) project. The second objective was to investigate whether VNIR modeling accuracy can be improved by sample stratification. The third objective was to evaluate the usefulness of these calibrated models to predict external soil samples. The final objective was devised to compare four calibration transfer techniques: Direct Standardization (DS), Piecewise Direct Standardization (PDS), External Parameter Orthogonalization (EPO) and spiking, to transfer field sample scans to laboratory scans of dry ground samples. Results showed that non-linear modeling techniques (ANN, RF and SVR) significantly outperform linear modeling technique (PLS) for all soil properties investigated (accuracy of PLS < RF < SVR ≤ ANN). Local models developed using the four auxiliary variables (Region, land use/land cover class, master horizon and textural class) improved the prediction for all properties (especially for PLS models) compared to the global models (in terms of Root Mean Squared Error of Prediction) with master horizon models outperforming other local models. From the calibration transfer study, it was evident that all the calibration transfer techniques (except for DS) can correct for spectral influences caused by sample intactness. EPO and spiking coupled with ANN model calibration showed the highest performance in accounting for the intactness of samples. These findings will be helpful for future efforts in linking legacy spectra to field spectra for successful implementation of the VNIR sensor systems for vertical or horizontal soil characterization.

Advisor Yufeng Ge

Share

COinS