Creation of mid-infrared spectroscopy calibration algorithms for soil property predictions

Abstract

Precision agriculture (PA) has been named as a cultivation method which could help alleviate food shortages in the future. However, PA relies on knowing the spatial distribution of soil properties, which requires repeated, quick, accurate, and cost-effective soil analysis. Conventional methods of soil analysis are slow and costly, reducing the application of PA in South Africa. Soil spectroscopy can fulfil the need for quick cost-effective soil analysis, but depend on robust calibration curves, of which none exist openly for South Africa. It is expected that mid-infrared soil spectroscopy can be used to analyse soil samples for soil properties pH, Effective Cation Exchange Capacity and Phosphorus within the Western Highveld summer grain area. Data for this study was provided by Noord Wes Kooperasie and Griekwaland Wes Kooperasie which resulted in a selection pool of ± 5 180 samples already analysed for soil properties. Conditioned Latin Hypercube sampling (cLHS) was then used to select 1 000 samples based on the pH, Effective CEC and P values of the dataset. The samples were then prepared and scanned with Mid Infrared spectroscopy (4 000 to 400 cm−1) using a Bruker Alpha II with DRIFTS module attached to create a spectral library. The soil property database and spectral library was combined and the R programming language was used to create calibration models using Cubist, Partial Least Square regression (PLSR) and Random Forest (RF). Each calibration algorithm was also applied onto two types of spectral datasets which includes spectra with pre-processing and spectra with minimal to no pre-processing applied. These models were validated using statistical performance measures including root mean square error (RMSE), squared correlation coefficient ( r2), standard deviation, bias and ratio of performance to deviation (RPD). Results show that models created for pH with Cubist and pre-processed spectral data had the best performance (R2=0.86,RMSE=0.3,RPD=2.66) along with ECEC with Cubist (R2=0.86,RMSE=0.3,RPD=2.66) and RF (R2=0.85,RMSE=0.72) and then P with some success using RF (R2=0.57,RMSE=13.48,RPD=1.51). Overall performance increase was observed with Cubist, PLSR and RF models using pre-processed spectral data compared to spectra with no pre-processing and produced acceptable prediction models to be able to predict pH and ECEC from soil spectra. Findings are consistent with other studies conducted worldwide but with little to no data to compare from South Africa more research and data is needed to create models that include all soil properties used for PA and that is representative of the whole of South Africa.