Prediction of in situ soil nitrogen mineralisation using artificial neural networks: a promising way to improve model accuracy
Résumé
The ability of artificial nural networ (ANN) to predict soil N mineralisation in fiels conditions using simple soil characteristics was tested on 74 unfertilised arable crop fields distributed all over France. In situ N mineralisation of soil mineral N and water contents. The ANN method used was the multilayer feed forward neuronal network trained by back propagation algorithm. A set of 56 sites was selected for ANN imput variables (clay content, log transformed CaCO3 content, organic N stock on 30 cm depht and soil pH) explained 62% of in situ N mineralisation rate of the training dataset without bias. However the prediction of N mineralisation for the validation dataset as biased with an error of ca 80 kg ha-1 for an average year. The use of cropping system information as input variables significantly improved both explpanation of dataset (R2=0.83, no bias) and rediction of independant dataset (prediction error of 30 kg ha-1). Thus ANN seemed well adapted to model N mineralisation using a limited number of easily measurable soil variables.
Origine : Fichiers éditeurs autorisés sur une archive ouverte
Loading...