Seasonal predictions of fire weather index: paving the way for their operational applicability in Mediterranean Europe

Managers of wildfire-prone landscapes in the Euro-Mediterranean region would greatly benefit from fire weather predictions a few months in advance, and particularly from the reliable prediction of extreme fire seasons. However, in some cases model biases prevent from a direct application of these predictions in an operational context. Fire risk management requires precise knowledge of the likely consequences of climate on fire risk, and the interest for decision-makers is focused on multi-variable fire danger indices, calculated through the combination of different model output variables. In this paper we consider whether the skill in dynamical seasonal predictions of one of the most widely applied of such indices (the Canadian Fire Weather Index, FWI) is sufficient to inform management decisions, and we examine various methodological aspects regarding the calibration of model outputs prior to its verification and operational applicability. We find that there is significant skill in predicting above average summer FWI in parts of SE Europe at 1 month lead time, but poor skill elsewhere. These results are largely linked to the predictability of relative humidity. Moreover, practical recommendations are given for the use of empirical quantile mapping in probabilistic seasonal FWI forecasts. Furthermore, we show how researchers, fire managers and other stakeholders can take advantage of a new open-source climate service in order to undertake all the necessary steps for data download, post-processing, analysis and verification in a straightforward and fully reproducible manner.We thank the European Union’s Seventh Framework Program [FP7/2007–2013] under Grant Agreement 308291 (EUPORIAS Project), in which this study was undertaken, and also for partially funding the ‘ECOMS User Data Gateway’ (ECOMS-UDG, http://meteo.unican.es/ecoms-udg), making available the System 4 hindcast, including derived variables from the raw model outputs required for this study (relative humidity, wind speed and deaccumulated precipitation). M. Iturbide thanks research funding from SODERCAN S.A. through the “Contrata” Programme (budget Ref. 12.04.461A.740.14). The first author thanks the FP7 Project SPECS (grant agreement 308378) for funding his current research contract and for supporting the development of the R package downscaleR for statistical downscaling and bias correction. Thanks to Jonas Bhend (MeteoSwiss), for the development of the verification routines and wrapper in R, and for fruitful discussions about the forecast verification methods. We are also grateful to two anonymous referees for their insightful comments