Soil phosphorus prediction by visNIR could be dependent on the conversional P determination method

Proximal soil spectroscopy can be useful to estimate relevant soil properties in real time and cheaply for agricultural decision support and soil health monitoring. However, prediction performance of plant available soil phosphorus by the visNIR has been unsatisfactory as it is considered among the least spectrally active soil properties. Hence, we compared prediction performance among plant available soil phosphorus (Olsen P), extractable soil phosphorus (ammonium-oxalate extract of P - AmOxP), total soil phosphorus (Aqua regia extract of P - TP) and phosphorus buffer index (PBI) using visNIR soil spectral sensing instrumentations (Neospectra and Fieldspec-4) using East African agricultural soils. The comparison was made by scanning 360 archived soil samples which were collected from 0-20cm soil depth in Ethiopia, Kenya and Tanzania. The spectra data was pre-treated with SavitskyGolay smoothing + first derivative and a PLSR was used to develop the predictive models from a 75% of the dataset (#270) subsampled by a conditioned Latin Hypercubic sampling (cLHS) method using the spectra space. The model performance was evaluated by an independent set of samples (#90) by calculating the concordance correlation coefficient (CCC), ratio of performance to interquartile range (RPIQ), bias and root mean square error of prediction (RMSEP).  The most important wavelengths for all soil P indicators in the NIR instrument ranged between 2150 -2400 nm whereas it included 500-570 nm for the visNIR instrument. PBI was predicted with higher CCC value of 0.94 and 0.89 for visNIR and NIR, respectively, however it has the least RPIQ (0.4 and 0.3, respectively) values when compared to other soil P prediction by both instruments. TP and AmOxP were predicted with higher accuracy and model consistency when compared to OlsenP and PBI. The visNIR range gave better prediction accuracy and model consistency for all soil P indicators than the NIR range. Hence, our findings indicated that TP and AmOxP could be preferred to predict soil phosphorus status for any agricultural and soil health monitoring using soil spectroscopic techniques.

Keywords: proximal soil spectroscopy, PLSR model, Savitzky-Golay smoothing filter, first derivative, Neospectra, Fieldspec-4, East Africa.