Artificial Neural Network based Methodologies for the Spatial and Temporal Estimation of Air Temperature - Application in the Greater Area of Chania, Greece

Artificial Neural Networks (ANN) propose an alternative promising methodological approach to the problem of time series assessment as well as point spatial interpolation of irregularly and gridded data. ANNs can be used as function approximators to estimate both the time and spatial air temperature distributions based on observational data. After reviewing the theoretical background as well as the relative advantages and limitations of ANN methodologies applicable to the field of air temperature time series and spatial modelling, this work focuses on implementation issues and on evaluating the accuracy of the AAN methodologies using a set of metrics in the case of a specific region with complex terrain. A number of alternative feed forward ANN topologies have been applied in order to assess the spatial and time series air temperature prediction capabilities in different horizons. For the temporal forecasting of air temperature ANNs were trained using the Levenberg-Marquardt back propagation algorithm with the optimum architecture being the one that minimizes the Mean Absolute Error on the validation set. For the spatial estimation of air temperature the Radial Basis Function and Multilayer Perceptrons non-linear Feed Forward AANs schemes are compared. The underlying air temperature temporal and spatial variability is found to be modeled efficiently by the ANNs.