The distribution of ambient air temperature in a city and the urban heat island intensity are investigated during the summer period in the major Athens region where ambient air temperature data are recorded at twenty stations. A neural network approach, based on predicted or recorded hourly values, is designed for modeling, predicting and estimating the air temperature at each station. Various feedforward, multiple layered, neural network architectures based on backpropagation algorithm are designed and trained for the stations' temperature prediction and estimation. The results were tested using extensive sets of measurements and it was found that they correspond well with the actual values. Furthermore, each one of the estimated stations is used as one input for the estimation of the next station's temperatures. The results were compared with the measured data and the neural network method was found able to simulate with sufficient accuracy the urban temperature field at several locations in a large urban region during the summer. ¿ 1999 American Geophysical Union