A new method to produce probability forecast about behavior of an on-going forest fire is proposed. Data used consist of telemetered infrared photographs of the on-goinq fire; terrain and fuel measurements; past and current weather data; and weather forecasts. Startup procedures based on Rothermel's model and quadratic programming are used to obtain measurements or directional rates of spread from successive photographs. Response surface tecnniques are used to formulate prediction equations from data on fire environment. Dynamic programming (Dijkstra algorithm) and predicted directional rates of spread are used to obtaln fire size and configuration from a point process analog of the fire phenomenon. Monte-Carlo procedures are used to obtain probability limits to forecasted arrival times of a given set of points. ©Society of American Foresters. Abstract reproduced by permission.