Please find my comments in the attached pdf file.

The paper is well-written and addresses a topical issue i.e. the scale of the data sets for optimizing spectral models. However, the authors rely on quality performance parameters that are not independent: RMSE, R² and RPD. Minasny et al. already demonstrated some years ago that there is an overlap between these parameters and that in recent literature the RPD is mostly used for global comparison between studies. Bellon Maurel proposed to use the PIQ which is better suited for non-normally distributed data sets.

Section 2.7. Regressions between RPD and CV are spurious, as the standard deviation appears in both terms. These relations should not be used (e.g. Fig. 5 and 7). The same applies to the relation between PRMSE and CV. Such relations always give a positive trend.

Finally, the authors conclude that local models perform better than general models, but do not really give a recommendation on a sampling strategy for fields/regions with no prior knowledge of the soil properties.

Throughout the manuscript

I am not a native speaker, but I would avoid using ‘high’ and ‘low’, e.g. small and large errors. Please check with a native speaker.

Generally, ‘concentration’ refers to solutions, while ‘content’ is a more general term that can also refer to a solid in another solid. You use both terms, but I would advise to systematically use ‘content’.