Hyperspectral Sensing of Soil Pedons for Soil Classification and Survey

Abstract

Proximal sensing using visible-near infrared (VNIR) diffuse reflectance spectroscopy (DRS) has demonstrated substantial potential for rapid, accurate estimation of key soil properties. Many of these soil properties are diagnostic for the purpose of soil classification and survey. As many of the soil surveys in the United States approach 40-50 y in age, there is need to enhance and supplement the current methodologies employed in the soil survey update process. This study was guided by the following objectives: (1) characterize the hyperspectral response of selected soil chemical and physical properties that are important to soil survey, (2) assess the quality and usefulness of hyperspectral data collected using laboratory- and field-based methods for soil classification, (3) develop an accurate method for estimating the spatial distribution of selected soil chemical and physical properties within a pedon sample, and (4) conduct a comparative assessment of soil characterization data from National Soil Survey Center (NSSC) and Cornell Nutrient Analysis Laboratory (CNAL) for paired analyses of selected soil chemical and physical properties. The hyperspectral responses of selected soil samples from pedons in Allegany and Ontario Counties in western New York were correlated with CNAL and NSSC soil property values using Partial Least Squares 1 (PLS1) regression. In addition, the degree of spectral discrimination among soil samples from known horizons was examined using principal components analysis (PCA). Also, an unaligned sampling grid (1.2 m ? 0.2 m) with a sampling interval of 0.05 m (n = 125 sample points) was used to assess the feasibility of predicting the spatial distribution of selected soil properties throughout the profile of one pedon in Allegany County. Conventional laboratory-based methods employed in this study yielded results consistent with previous studies, with most R2 values in excess of 0.85. Field-based methods implemented in situ yielded results which performed similarly to laboratory-based methods. Where model performance measures were less than expected, likely sources of error included small sample sizes and variable soil moisture in the field setting, which has been shown to influence soil spectral response. An efficient data processing flow was also developed for accurately estimating the spatial distribution of selected soil physical and chemical properties using PLS1 regression and ordinary kriging. While model predictions were generally within acceptable error tolerances, a larger sample size and a more robust prediction model would reduce error rates. Finally, a comparative assessment of soil characterization data produced by CNAL and NSSC demonstrated substantial agreement between the two laboratories. The results can help local practitioners obtain data more rapidly through local university-based laboratories, while still allowing comparison and statistical analyses with data provided by the national soil survey laboratory. This study clearly demonstrated the potential of proximal sensing using VNIR DRS for estimating selected soil physical and chemical properties of importance to soil survey. The results of this study suggest that advanced proximal sensing techniques, both laboratory-based and in situ, are worthy of additional study to assess the full potential of this technology for soil survey. For VNIR DRS to be deployed as a complement to field-based soil survey operations, potential users must have relevant information in terms of system capabilities and limitations and an appropriate level of training for them to use this technology in an efficient and effective manner.