#1Unil UNIL | Université de Lausanne Faculté des géosciences et de l'environnement CC 0 BY cima RESEARCH FOUNDATION Wildfire susceptibility mapping via machine learning: the case study of Liguria Region, Italy PAOLO FIORUCCI (1), MIRKO D'ANDREA(1), ANDREA TRUCCHIAI (1), MARJ TONINI (2) (1) CIMA RESEARCH FOUNDATION, ITALY (2) INSTITUTE OF EARTH SURFACE DYNAMICS, FACULTY OF GEOSCIENCES AND ENVIRONMENT, UNIVERSITY OF LAUSANNE, SWITZERLAND EGU European Geosciences Union#2Motivation & Objectives CC 0 BY The identification of areas most vulnerable to fire risk is a key tool in wildfire management, particularly in view of the limited availability of fire risk management resources. Wildfire susceptibility mapping allows to identify these areas, assessed defining a rank from low to high. These are elaborated taking into account two aspects: where the wildfires occur and which are the predisposing factors. Objectives: Elaborate wildfire susceptibility mapping for Liguria region (Italy) with a Machine Learning approach (Random Forest) ➤Identify the most influential factors in determining high susceptibility areas WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). 2#3. Input Data • DEM (digital elevation model) and derivatives (slope, northness, • and eastness); • • • • Distance from anthropogenic features (urban areas, roads, pathways, crops); Vegetation type (categorical variable with 37 classes) % of neighboring vegetation type surrounding each pixel; Protected areas. CC 0 BY N km 0 25 50 Aree non vegetate Altri coltivi Praterie Oliveti Pinete Castagneti Altre latifoglie potenzialmente soggette al fuoco Vegetazione arbustiva. Altri boschi poco soggetti al fuoco WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). Ligurian Land Cover 3#4Input variables Observations + YES (68%) Slope 13%? Observations Veg-37? NO (32%) Temp > 28 °C? Machine Learning approach: Random Forest YES (30%) Prec 120mm? NO (38%) YES (32%) NO (6%) YES (10%) NO (20%) Observations Veg-37? Observations Veg-37? YES (68%) NO (32%) Slope 13% ? Temp > 28 °C? YES (30%) Prec 120mm? NO (38%) YES (32%) NO (6%) YES (68%) NO (32%) Slope 13% ? Temp > 28 °C? YES (30%) NO (38%) YES (32%) NO (6%) YES (10%) NO (20%) Prec 120mm ? YES (10%) NO (20%) Observations YES (68%) Slope 13% ? Veg-37? NO (32%) Temp > 28 °C? YES (30%) NO (38%) YES (32%) NO (6%) Prec 120mm? YES (10%) NO (20%) Ensemble of decision trees Output (susceptibility map) High probability Low probability CC 0 WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). 4 BY#5Variable-class importance RF allows to measure the relative importance of each variable on the prediction. This is obtained by looking at how much the tree nodes, which use that variable, reduce the mean square errors across all the trees in the forest. The vegetation type variable was a categorical one, composed of 37 different classes. The importance of each class is validated through Partial dependence plots, which give a graphical depiction of the marginal effect of a single class on the class on the variable importance on overall susceptibility. Such marginal effect can be positive (class augmenting fire susceptibility) or negative (class decreasing fire susceptibility) CC 0 WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). 5 BY#6CC 8 BY VEG TYPE 40 DEM. % Moore-Bushes- % Shrubs 50 Tracks Dist - Roads Dist - % PASTURES- Importance % URBAN Dist- Variable WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). Slope North crops Dist- % Mediterr. Pine - % Chestunt - % Conif. Woods. % Mixed Woods 80 Variable importance#7CC 0 BY Insights on veg-type: class importance › Classes of vegetation enhancing susceptibility: bushes, Coniferous woods, pastures, Sclerophyll, mixed woods. ➤ Classes of vegetation dampening susceptibility: chestnut, Fagus, Quercus Cerris, Orno-ostrietum 0.75- Importance 0.50- 0.25- 0.00 - -0.25- -0.50 - Moore - bushes- Conif. woods- Pastures- Sclerophyllous- Mixed Woods- Class WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). Orno-Ostryenion- Quercus Cerris- Fagus(Beech)- Chestnut- 7#8CC Susceptibility maps: Summer season 0 25 50 km 0 BY WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). nv_fivefolds_perc_s Very Low Low Middle High Extreme 8#9Susceptibility maps: Winter season 0 25 25 50 km 0 WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). CC BY nv_fivefolds perc_w Very Low Low Middle High Extreme 9#10Validation For RF the prediction error is assessed by evaluating predictions on the “out-of-bag" data, which were not used in the training subset. The testing and the OOB mean squared error are in our case both equal to about 0.2 for winter season and to 0.17 for summer season, attesting the robustness of the model. ►To assess the methods prediction capabilities, an indipendent dataset (BA 2016-2017) was employed. Summer Winter % 2016 2017 2016 2017 0.30 0.01 0.04 0.02 0.02 16/01/2017 770 ha 0.20 0.02 0.12 0.02 0.04 0.30 0.11 0.45 0.12 0.16 0.15 0.25 0.22 0.18 0.30 0.05 0.53 0.10 0.39 0.46 CC 0 BY 23/08/2016 322 ha 8 km WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). 17/01/2017 315 ha Wildfire Very low Low Medium High Extreme 10#11CC Conclusions Higher class match with the observed burned areas, which are closer to the coast in summer and develop along the interior in winter. • The RF algorithm seems to achieve good performances and has the advantage of being able to extract knowledge and insights from data, and developing promising ranking of variables and categorical classes. ⚫ In summary, RF seems to be a promising alternative to deterministic or statistical expert-based method for wildfire susceptibility mapping. BIBLIOGRAPHY Tonini, M.; D'Andrea, M.; Biondi, G.; Degli Esposti, S.; Trucchia, A.; Fiorucci, P. A Machine Learning-Based Approach for Wildfire Susceptibility Mapping. The Case Study of the Liguria Region in Italy. Geosciences 2020, 10, 105. 0 BY WILDFIRE SUSCEPTIBILITY MAPPING IN LIGURIA (ITALY). 11