Prediction of soil characteristics and colour using data from the National Soils Inventory of Scotland

Abstract

A neural network (NN) approach was used to determine relationships between soil colour and a range of physical and chemical characteristics, using a dataset derived from the NSIS (National Soil Inventory of Scotland) database. It was found that several soil characteristics could be predicted accurately from colour, using only Red, Green and Blue (RGB) values from the RGB system or L, a and b from the CIELab system. These characteristics included organic matter content (measured by Loss On Ignition), nitrogen content and several elements including Ca, Ti and Mo. It was found that some parameters, such as potassium and phosphorus, were not predicted accurately, however. Prediction of soil colour from available physiochemical parameters was found to give high levels of accuracy, with the strongest influence of prediction coming from LOI, nitrogen, mineral texture and a few metals including V, Cr, Mn and Zn. Sensitivity analysis of the trained neural network models was carried out, but did not provide much useful information. Potential applications of the NN modelling approach are discussed, including rapid field assessment of soil nutrient status, and potential improvements to soil horizon classifications.