A lack of consistency between analyses for predicting physical/ chemical parameters and colour raises the possibility that the method used to investigate the trained neural networks is not very effective for the identification and quantification of key relationships between colour and soil parameters. Such an approach is only likely
to work when the trained network to which it is applied is providing consistently accurate predictions of the outputs. In this case, the accuracy of the networks used for predicting soil physiochemical characteristics varies between parameters. Therefore, a relatively strong relationship identified for some parameter that is not predicted accurately (e.g. lead) could be an artefact of the training process, and is likely to be less accurate than a relationship identified for parameters that are predicted accurately.