Modeling forest-shrubland fire susceptibility based on machine learning and geospatial approaches in mountains of Kurdistan Region, Iraq

Abstract

Forest-shrubland fires cause remarkable ecological and socio-economic damages in different parts of the world. The damages could increase in the wake of global climate change, particularly in arid and semi-arid regions where recurrent fires have already been reported. In Iraq, the majority of natural forests (including shrubland) are prone to seasonal fire events, situated in the mountainous regions of the northeast of the country. In Iraq, information on the spatial pattern of fire hot spots, the potential spatial distribution of fire events, and the factors that determine the spatial distribution of fire across the mountain ecosystem is limited or non-existent. To fill in this gap, this study presents an integrated approach of the Maximum entropy (Maxent) modeling, Getis-Ord Gi* hot spot analysis, and Moran’s I index within the GIS environment to achieve the following aims: (i) modeling forest-shrubland fire susceptible areas based on satellite-based time-series (2001–2020) fire records and relevant environmental predictors; (ii) mapping forest-shrubland fire hot spots (fire susceptible areas); (iii) integrating the generated maps from (i) and (ii) to showcase the fire susceptible zones; and (iv) identifying the most important determinants of the spatial distribution of fire events. Modeling demonstrated that around 15.6% (8215.15 km²) of the study area is susceptible to fire. The susceptible areas were further categorized into moderate (2.62% (1376.57 km²)), high (1.98% (1039.22 km²)), very high (6.82% (3575.10 km²)), and extremely high (4.24% (2224.25 km²)) fire susceptible classes. In addition, hot spot analysis demonstrated the availability of spatial clustering of the fire events and statistically significant hot spots. Land cover, temperature, precipitation, and distance to the road were among the factors that contributed relatively to the occurrence and distribution of fire events across the study area. Modeling and geospatial techniques provide useful means for identifying and predicting potential fire susceptible zones based on which preventive-precautionary measures could be taken before the fire season starts. Preventive measures should mainly focus on ‘extremely high’ susceptible areas and significant hot spots. This study contributes by providing useful information on the fire susceptible zones in the mountain regions of Iraq, based on which accurate management actions can be achieved.

Appendix

Fig. 6

The Jackknife test (AUC metric) for the fire events for predictor gains

Fig. 7

The Jackknife test of regularizing training gain for the fire events in respect to predictors used in modeling

Fig. 8

Mean AUC and standard deviations (model performance)

Table 4 Environmental predictors prepared and considered in forest-shrubland fire model building after collinearity analysis