Extraction of reflectance spectra features for estimation of surface, subsurface, and profile soil properties

Abstract

• Accuracy of reflectance spectroscopy estimates of soil properties varied by depth. • Number and location of important wavebands varied by depth and by soil property. • Some calibrations could be applied successfully across multiple depth intervals. • Results could guide development of discrete-waveband soil property sensors. Diffuse reflectance spectroscopy in the visible and near-infrared wavelength ranges has potential to provide high-resolution, pollution-free, and nondestructive estimation of soil chemical and physical properties for use in precision agriculture. Practical implementation of this approach would be facilitated if soil property sensors using a limited number of reflectance bands could maintain accuracy similar to more expensive and complex full-spectrum sensors. Studies identifying such bands are limited, especially for subsurface soils. Thus, in this study, an existing spectral database of 697 soil samples was used to compare results for three soil categories (profile, surface, and subsurface) and multiple waveband selection methods. Soil cores were obtained to approximately 1.2 m depth from ten fields, two each in Missouri, Illinois, Michigan, South Dakota, and Iowa, USA, then sieved and air-dried. Laboratory soil spectra were obtained from 350 to 2500 nm using a commercial spectrometer and soil properties (total nitrogen, soil organic carbon, total carbon, magnesium, calcium, potassium, soil texture (clay, silt, and sand) fractions, cation exchange capacity, and pH) were measured using standard laboratory analyses. The ability of ten spectral preprocessing techniques to improve analysis results was investigated. Backward interval partial least squares was used to identify those spectral regions most predictive of soil properties. Alternatively, specific characteristic wavelengths were identified by a combination genetic algorithm (GA)-back propagation neural network (BPNN) approach. Results were compared for three soil property estimation methods: (1) partial least squares regression (PLSR) models based on the full spectrum, (2) PLSR models based on sensitive regions, and (3) BPNN models based on characteristic wavelengths. The best results for profile and subsurface soils were obtained with absorbance preprocessing, but for the surface soils, the standard normal variate transformation was best. For some soil properties, the prediction R 2 of the PLSR models based on sensitive regions was better than that of the PLSR models based on the full spectrum, demonstrating that retaining only sensitive wavebands could improve estimates. However, in some cases, the reduction in wavebands decreased accuracy. Differences in prediction accuracy across all calibration models over profile and subsurface soils were relatively small but were larger for surface soils. Furthermore, application of characteristic wavelength calibrations to other soil datasets resulted in a lower accuracy than with the full spectrum calibration developed for that dataset. In general, this study shows that there are measurable differences in prediction accuracy across all calibration models over the three soil depth categories. The experimental results of this study illustrate the potential for a set of wavelengths optimized for one depth category to still provide acceptable estimates for other depth categories. Overall, these results provide important guidance for the development of DRS soil sensors based on discrete wavebands to reduce cost and increase the speed of in-field data collection.