Abstract

Reflectance spectrometry is an emerging and non-destructive detection technique bearing fast, cheap, and accurate results compared with conventional assessments. Most field and laboratory-based spectrometers are restricted to VNIR (visible–near-infrared). However, soils fail to show well-defined narrow absorption bands in this region. This obstructs the use of curve feature as a diagnostic criterion for soil nutrient predictions.

In this paper artificial neural network (ANN) is implemented to estimate soil organic matter, phosphorous, and potassium from the VNIR spectrum (400–1100 nm). Macronutrients were modelled from 41 bare soil reflectances of Lop Buri province, Thailand. Neurons were trained from 7 bandwidth categories derived from laboratory-based StellarNet spectroradiometer and in situ photometer.

Satisfactory results were attained and compared across different synthesised bandwidths. Models exhibited slightly better estimates from the laboratory than in situ spectra, and from narrower than broader bandwidths. Widening bandwidth corresponds with attenuated predictive powers, coupled with rising errors. Cross validation of models yielded acceptable correlations. The strength of models confirmed the capability of ANN to estimate macronutrients by solving difficulties incurred from high cross-channel correlations prevailing in conventional statistical techniques.