Aarhus University Seal / Aarhus Universitets segl

Mogens Humlekrog Greve

A data analysis workflow to enhance clay and organic carbon models using proximal Vis-NIR data

Publikation: KonferencebidragKonferenceabstrakt til konferenceForskning

Modelling proximal sensors data is becoming a norm in soil characterization and mapping. In many cases, these models still have low predictive capabilities and lack robustness due to the large amount of noise from several environmental factors. In this study we proposed a combination of extensive data preprocessing (preprocessing survey) and two variable selection methods to significantly increase visible near-infrared spectroscopy (Vis-NIRS) model performance and stability. Spectra of eight agricultural fields were measured in the range of 350-2200 nm using a mobile sensor platform (Veris Technologies, USA) towed by a tractor. A fuzzy c-means clustering was performed based on the first 3 principal components to select 15 representative sampling locations in each field. Clay and organic carbon (OC) were determined for all calibration samples using pipette and ignition methods, respectively. Spectral data were preprocessed using several thousands of combinations of methods/settings including Savistky-Golay smoothing/derivatives, multiplicative scatter correction, standard normal variate and generalized least squares weighting and the optimum Partial Least Squares (PLS) models for clay and for OC were chosen. Then several interval partial least squares (iPLS) regressions were performed and the most useful wavelengths were selected. Among the remaining wavelengths, using Martens Uncertainty Test, wavelengths with significant regression coefficient were selected for final analysis. For OC models, the results show 53% decrease in root mean square error of cross validation (RMSECV) and 26% increase in R2 compared to models with normal preprocessing and no variable selection methods. For clay models, there was 28% decrease in RMSECV and 5% increase in the R2. The results indicate the importance of following a well-designed workflow of preprocessing survey and variable selection in producing reliable and robust models from proximally sensed Vis-NIRS data.
Antal sider1
StatusUdgivet - 2017
Begivenhed2017 Canadian Society of Soil Science Annual Meeting - Trent University, Petersborough, Canada
Varighed: 10 jun. 201714 jun. 2017


Konference2017 Canadian Society of Soil Science Annual Meeting
LokationTrent University

Se relationer på Aarhus Universitet Citationsformater

ID: 114480781