A neural network experiment on the simulation of daily nitrate-nitrogen and suspended sediment fluxes from a small agricultural catchment
François Anctil,
Mélanie Filion and
Julien Tournebize
Ecological Modelling, 2009, vol. 220, issue 6, 879-887
Abstract:
In this paper, we report an application of neural networks to simulate daily nitrate-nitrogen and suspended sediment fluxes from a small 7.1km2 agricultural catchment (Melarchez), 70km east of Paris, France. Nitrate-nitrogen and sediment losses are only a few possible consequences of soil erosion and biochemical applications associated to human activities such as intensive agriculture. Stacked multilayer perceptrons models (MLPs) like the ones explored here are based on commonly available inputs and yet are reasonably accurate considering their simplicity and ease of implementation. Note that the simulation does not resort on water quality flux observations at previous time steps as model inputs, which would be appropriate, for example, to predict the water chemistry of a drinking water plant a few time steps ahead. The water quality fluxes are strictly mapped to historical mean flux values and to hydro-climatic variables such as stream flow, rainfall, and soil moisture index (12 model input candidates in total), allowing its usage even when no flux observations are available. Self-organizing feature maps based on the network structure established by Kohonen were employed first to produce the training and the testing data sets, with the intent to produce statistically close subsets so that any difference in model performance between validation and testing has to be associated to the model and not to the data subsets. The stacked MLPs reached different levels of performance simulating the nitrate-nitrogen flux and the suspended sediment flux. In the first instance, 2-input stacked MLP nitrate-nitrogen simulations, based on the same-day stream flow and on the 80-cm soil moisture index, have a performance of almost 90% according to the efficiency index. On the other hand, the performance of 3-input stacked MLPs (same-day stream flow, same-day historical flux, and same-day stream flow increment) reached a little more than 75% according to the same criterion. The results presented here are deemed already promising enough, and should encourage water resources managers to implement simple models whenever appropriate.
Keywords: Neural networks; Surface water quality; Nitrate-nitrogen; Suspended sediment (search for similar items in EconPapers)
Date: 2009
References: View references in EconPapers View complete reference list from CitEc
Citations: View citations in EconPapers (3)
Downloads: (external link)
http://www.sciencedirect.com/science/article/pii/S0304380008005978
Full text for ScienceDirect subscribers only
Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.
Export reference: BibTeX
RIS (EndNote, ProCite, RefMan)
HTML/Text
Persistent link: https://EconPapers.repec.org/RePEc:eee:ecomod:v:220:y:2009:i:6:p:879-887
DOI: 10.1016/j.ecolmodel.2008.12.021
Access Statistics for this article
Ecological Modelling is currently edited by Brian D. Fath
More articles in Ecological Modelling from Elsevier
Bibliographic data for series maintained by Catherine Liu ().