Artificial Neural Network Modeling of Relative Humidity and Air Temperature Spatial and Temporal Distributions Over Complex Terrains

Abstract

In this work we present a methodological approach of applying Artificial Neural Networks (ANN) for modeling of both the air temperature (AT) and relative humidity (RH) spatial and temporal distributions over complex terrains. A number of implementation issues are discussed, along with their relative advantages and limitations. Moreover, after the introduction of a set of metrics, the accuracy of the evaluation of ANN based spatial and time series AT and RH modeling in the case of a specific region is examined, by applying a number of alternative feed forward ANN topologies. The Levenberg-Marquardt back propagation algorithm was used for the ANNs training in the temporal forecasting of AT and RH, with the optimum architecture being the one that minimizes the Mean Absolute Error on the validation set. The Radial Basis Function and the Multilayer Perceptrons non-linear Feed Forward ANNs schemes are compared for the spatial estimation of AT and RH. We found that the spatial and temporal AT and RH variability over complex terrains can be modeled efficiently by ANNs.