The ability to predict and retrodict heat island intensity has several practical applications in relation to forecasts for energy and water demand, heat waves, wind chill, fog occurrence, icing, atmospheric stability, etc. of urban areas. Three years (January 1, 1997-December 31, 1999) of hourly records of screen-level air temperature from three stations (rural outskirts, city core and suburban) in and around Lodz, Poland (population 0.85M) and a parallel hourly record of other meteorological variables (including humidity, cloud amount and type, wind speed and direction, solar radiation and daily values of rainfall and snow depth at the rural site) are exploited in order to assess the efficacy of four different approaches to obtain the hourly UHI. The approaches tested are traditional linear regression, artificial neural network, the objective algorithmic scheme of Oke (1998) and the Town Energy Balance (TEB) numerical model of Masson (2000). Whilst all of these methods use standard weather information from a rural station as input, they have very different needs with respect to either city/site description or city-specific statistical coefficients based on pre-existing heat island data from the location. The relative performance of the approaches are compared.