Abstract : This thesis is the first study to include hydrological reasoning already at the development step of a radar rainfall forecasting method. In this, we intend to be closer to the objectives of the final application in urban hydrology. First, the major rainfall forecasting projects of the last 20 years are reviewed, the ideas of the previous studies as well as new techniques have been analysed. There are several reasonable clustering algorithms of which the pros and cons are described. A special focus has been put on:- heavy rainfall regions (> 5 mm/h), - fast calculation time on the given computers (inferior to 60 sec) and - quality control of the forecast in real time. All these analyses led to a forecasting method that has been tested on 94 images of seven very different events. The summary of the test results states : - that the chosen scout II.0 method is the best together with the classical crosscorrelation method ; - that the forecasting results of scout II.0 encourage its operational testing ; - that a preliminary method also including a growth and decay forecast does not yet give satisfactory results ; - that it is possible to determine the reliability of a forecast by two imagery features before even issuing the forecast ; - that the matching step of scout II.0 can still be improved by a more optimal control procedure ; - that the future efforts on improving this method will be a better formalisation of the matching step, leading to a larger number of echoes recognized, and a closer look on to growth and decay of echoes. The scout II.0 method was then implemented and put into operational use on the Seine-Saint-Denis system.