High precisive fine-scale temperature forecast serves as an important tool in quantifying human comfort, health threats with extreme temperature and energy consumption. In this study, Urban Microscale Temperature Forecast (UMTF) model was developed using machine learning techniques (k-means clustering and support vectors machine) with reference to global ensembles and geomorphometry datasets (sky view factor, daily sun trajectory and urban terrain model). It statistically downscales the daily temperature extremes of 10-km ensemble forecast into 50-m resolution across Hong Kong. It is observed that the downscaling process has contributed to a considerable reduction of mean absolute errors (up to 1.0 degree Celsius (45.1%)) with improved correlation (R² = 0.84) across 26 automatic weather stations. To further validate the result, an independent dataset from the Community Weather Information Network (CoWIN) and the ASTER heat island intensity map were added for the model evaluation. The comparison result illustrates clear and resemble spatiality in temperature extremes with sophisticated urban buildings, park greenery and coastal front features can be re-projected in details. At last, the study demonstrates the low-cost weather stations could be an effective way to achieve a detailed temperature forecast in a heterogeneous urban environment.