Neural network treatement of 3 years long NO measurement in temperate and tropical climates
Résumé
The nitrogen monoxide (NO) is essential in atmospheric chemistry and especially in the troposphere, although this gas is in low concentration in the atmosphere (approximately 1 ppb). Indeed, it quickly reacts to obtain nitrogen dioxide (NO2) and to form tropospheric ozone, the major atmospheric oxidant. Nowadays, the increase of NOx concentration in the atmosphere is considered to be 0.25% per year (Davidson, 1997). This rise is essentially due to the anthropologic pressure, in which changing in land occupation takes place. Globally, soil NO emission could represent nearly 40% of total NOx emission, an amount comparable to fossil fuel combustions. However, uncertainties are large since results are ranging between 5.5 and 21 TgN per year. NO emission variation laws with soil humidity, soil temperature and other environmental parameters are well known, but those results present a high level of temporal and geographical variation. In order to generalise those complex phenomena evolution laws, a neural network have been developed. By introducing as inputs different meteorological parameters (temperature, humidity….) and soil characteristics (texture, N total, organic mater….), the neural network generates general temperate and tropical parameterisation laws that gives calculated NO fluxes in output. The goal of this study is to present 2 parameterization results on temperate and tropical forest soil NO emissions. The temperate parameterization of South-Eastern Germany forest (Hoëglwald, 1995-1997) use air and soil temperatures, soil humidity and humus pH as input data to generate soil NO fluxes in output. Fluxes are well represented by Neural Network in high and low frequencies (R2=0.76) and present good overall estimation of NO measurement (relative error < 1%). In tropical condition (Queensland, Australia), NO emission parameterization is performed from soil temperatures and humidity. It still leads to good high and low frequencies NO emission representation (R2=0.69) with a relative error < 1%. Then, temperate (1995-1997) parameterization is used to estimate (2002-2003) NO fluxes. This occurs well high and low frequency representations, but problems to evaluate magnitude of unknown meteorological situation fluxes. The final application of the neural network is not to be used as a prediction tool. Generated parameterizations are introduced on chemical model like Meso-NH to represent soil NO emissions and their impact on tropospheric chemistry, especially in tropospheric ozone production and destruction cycle.
