Using long-term experiments to assess statistical models for determining critical P fertilisation thresholds

A - Papers appearing in refereed journals

Jordan-Meille, L., Mollier, A., Poulton, P. R., Glendining, M. J., Jouany, C., Gratecap, J. -., Hanocq, D., Montagnier, C. and Denoroy, P. 2024. Using long-term experiments to assess statistical models for determining critical P fertilisation thresholds. European Journal of Agronomy. 158 (August), p. 127220. https://doi.org/10.1016/j.eja.2024.127220

AuthorsJordan-Meille, L., Mollier, A., Poulton, P. R., Glendining, M. J., Jouany, C., Gratecap, J. -., Hanocq, D., Montagnier, C. and Denoroy, P.
Abstract

Phosphorus fertilisation rules particularly rely on a diagnosis step that consists of comparing the level of plant available P in the soil by analysis with a "critical P threshold value (Pcrit). Pcrit are generally calculated from data obtained on Long Term Fertilisation Experiments (LTFEs), providing yield vs available-P relationships. According to the statistical functions used to fit such relationships, the Pcrit may vary up to two-fold, therefore introducing a tremendous variability in recommendations between users of different methods. The purpose of this study is to provide objective arguments to help choose the "right" model that can handle data sets of varying quality. Our analysis is based on six LTFEs providing complete environmental description, regular Olsen P measurements of available P, and annual yields for wheat, maize, barley and durum wheat. The dataset is composed of 4500 point pairs "yield-Olsen P". Four models were tested: Linear Plateau (LP), Quadratic-Plateau (QP), Mitscherlich 95% (Mi95) and Cate-Nelson clustering algorithm (CN). The model assessment was mainly based on the three following criteria: quality of adjustment fitting (RMSE), risk of high interannual variability (Coefficient of Variation, CV), risk of providing out of range Pcrit, standing for accuracy, certainty and robustness, respectively. The Pcrit calculation of a given crop on a given trial was tested using either pooled data relying on relative yields, or by averaging on annual trials. Our results clearly show that the two first indicators (RMSE, CV) could not discriminate between models. However, the ability of models to cope with scarce data strongly favoured the LP function. We also proved that using pooled relative yields or averaging annual trails not only led to the same conclusions, but provided also the same orders of magnitude of Pcrit. As mentioned in the literature, we confirmed the strong impact of the method on the final Pcrit, with CN and QP functions exhibiting the lowest and highest values, respectively. Finally, we recommend the use of the LP method, which combines many advantages: simple algorithm, strong ability to provide sensible thresholds even on trials relying on few data, equivalent model fitting and interannual variations than more complex models, no need to rely on an arbitrary yield gap, and intermediate values of thresholds. The relative yield method should be preferred wherever annual data are too sparse for model fitting; alternatively, averaging annual data seemed simpler and not subject to the arbitrary choice of a "reference yield".

KeywordsP-threshold; Yield-response; Fertilisation; Linear-plateau; Assessment; Fitting models
Year of Publication2024
JournalEuropean Journal of Agronomy
Journal citation158 (August), p. 127220
Digital Object Identifier (DOI)https://doi.org/10.1016/j.eja.2024.127220
Open accessPublished as non-open access
FunderBiotechnology and Biological Sciences Research Council
Funder project or codeS2N - Soil to Nutrition [ISPG]
Output statusPublished
Publication dates
Online31 May 2024
Publication process dates
Accepted17 May 2024
ISSN1161-0301
PublisherElsevier

Permalink - https://repository.rothamsted.ac.uk/item/990ww/using-long-term-experiments-to-assess-statistical-models-for-determining-critical-p-fertilisation-thresholds

Restricted files

Publisher's version

Under embargo indefinitely

14 total views
0 total downloads
10 views this month
0 downloads this month