Meta-modeling methods for estimating ammonia volatilization from nitrogen fertilizer and manure applications

JOURNAL OF ENVIRONMENTAL MANAGEMENT

卷:236

Article

Ramanantenasoa, Maharavo Marie Julie A.^1,2  Genermont, Sophie¹  Gilliot, Jean-Marc¹  Bedos, Carole¹  Makowski, David^3,4 

[1] Univ Paris Saclay, UMR ECOSYS, INRA, AgroParisTech, Route Ferme, F-78850 Thiverval Grignon, France

[2] ADEME, 20 Ave Gresille,BP 90406, F-49004 Angers 01, France

[3] Univ Paris Saclay, UMR Agron, INRA, AgroParisTech, Ave Lucien Bretignieres, F-78850 Thiverval Grignon, France

[4] CIRED, 45bis Ave Belle Gabrielle, F-94130 Nogent Sur Marne, France

关键词: Meta-modeling;    Volt'Air;    Ammonia emission;    Fertilization;    Process-based model;   

DOI  :  10.1016/j.jenvman.2019.01.066

来源: Elsevier

【 摘 要 】

Accurate estimations of ammonia (NH3) emissions due to nitrogen (N) fertilization are required to identify efficient mitigation techniques and improve agricultural practices. Process-based models such as Volt'Air can be used for this purpose because they incorporate the effects of several key factors influencing NH3 volatilization at fine spatio-temporal resolutions. However, these models require a large number of input variables and their implementation on a large scale requires long computation times that may restrict their use by public environmental agencies. In this study, we assess the capabilities of various types of meta-models to emulate the complex process-based Volt'Air for estimating NH3 emission rates from N fertilizer and manure applications. Meta-models were developed for three types of fertilizer (N solution, cattle farmyard manure, and pig slurry) for four major agricultural French regions (Bretagne, Champagne-Ardenne, Ile-de-France, and Rhone-Alpes) and at the national (France) scale. The meta-models were developed from 106,092 NH3 emissions simulated by Volt'Air in France. Their performances were evaluated by cross-validation, and the meta-models providing the best approximation of the original model were selected. The results showed that random forest and ordinary linear regression models were more accurate than generalized additive models, partial least squares regressions, and least absolute shrinkage and selection operator regressions. Better approximations of Volt'Air simulations were obtained for cattle farmyard manure (3% < relative root mean square error of prediction (RRMSEP) < 8%) than for pig slurry (17% < RRMSEP < 19%) and N solution (21% < RRMSEP < 40%). The selected meta models included between 6 and 15 input variables related to weather conditions, soil properties and cultural practices. Because of their simplicity and their short computation time, our meta-models offer a promising alternative to process-based models for NH3 emission inventories at both regional and national scales. Our approach could be implemented to emulate other process-based models in other countries.

【 授权许可】

Free