In this work, parameters such as air temperature and air relative humidity were measured at and studied for meteorological stations across Samaria canyon (Greece), a narrow mountain passage with unusual morphological and climatological characteristics. The study of the cross-correlation function of the meteorological parameters among all station locations (Prinari, Samaria, Perdika and Christos measurement sites) has allowed the evaluation of the time range and the degree of dependency between sequential values of meteorological parameters at different sites. It also allowed the determination of a number of reference stations and the development of a multiple linear regression model (MLR) and an artificial neural network model (Multilayer Perceptron or MLP) for estimating meteorological values for the sites Perdika and Christos using data from Prinari and Samaria stations. The measured data were divided into two separate overall data sets for model development (training dataset) and model evaluation (testing dataset). To evaluate the quality of the results obtained by the MLR and the MLP model, the determination coefficient (R-2) and the mean absolute error (MAE) were used. Results showed that, for the Perdika station testing dataset, the was found to be 0.903 and 0.914 (MAE 0.897 degrees C and 0.821 degrees C) for the MLR and MLP model, respectively. For the Christos station dataset, (without the use of Perdika station data), the resulting R-2 was found to be 0.56 and 0.608 (MAE 1.810 degrees C and 1.718 degrees C) for the MLR and MLP model, respectively; with the inclusion of the Perdika station dataset, however, the values of the R-2 of the testing dataset were increased to 0.675 and 0.756 (MAE were decrease to 1.395 degrees C and 1.220 degrees C), for the MLR an MLP model, respectively. Based on these preliminary results, it was concluded that there has been achieved a very good estimation for Perdika station temperature and a satisfactory one for Christos station. In addition, results revealed that the ability of Artificial Neural Network models to take into account any nonlinear relationships among meteorological parameters may provide an advantage in the quality of temperature estimations in complex terrains.
