Spatio-temporal interpolation of climatic variables over large region of complex terrain using neural networks

Empirical models for seven climatic variables (monthly mean air temperature , monthly mean daily minimum and maximum air temperature, monthly mean rela tive humidity, monthly precipitation, monthly mean global solar irradiation and monthly potential evapotranspiration) were built using neural networks . Climatic data from 127 weather stations were used, comprising more than 3 0000 cases for each variable. Independent estimators were elevation, latitu de, longitude, month and time series of respective climatic variable observ ed at two weather stations (coastal and inland), which have long time-serie s of climatic variables (from mid last century). Goodness of fit by model w as very high for all climatic variables (R > 0.98), except for monthly mean relative humidity and monthly precipitation, for which it was somewhat low er (R = 0.84 and R = 0.80, respectively). Differences in residuals around m odel were insignificant between months, but significant between weather sta tions, both for all climatic variables. This was the reason for calculation of mean residuals for all stations, which were spatially interpolated by k riging and used as a model correction. Similarly interpolated standard devi ation and standard error of residuals are estimators of the model precision and model error, respectively. Goodness of fit after the averaging of mont hly values between years was very high for all climatic variables, which en ables construction of spatial distributions of average climate (climatic at las) for a given period. Presented interpolation models provide reliable, b oth spatial and temporal estimations of climatic variables, especially usef ul for dendroecological analysis. (C) 2001 Elsevier Science B.V. All rights reserved.