I believe your reasoning is incorrect. If we model the number of fires by a poisson distribution (very reasonable), a single sample would correspond to the mode of the distribution. What you would like is an estimate of the mean, i.e. lambda. It so happens that the mode of a poisson distribution is floor(lambda) or ceiling(lambda)-1, and thus it is an underestimate of the mean. In other words, fires are likely to be less frequent than what a single sample suggests.