A hybrid neural network based Australian wildfire prediction: a novel approach using environmental and satellite imagery

Every year catastrophic wildfires during summer months have a major negative socio-economic impact on Australian life and fabric of the society. Early prediction of wildfire prone locations is yet a poorly understood environmental challenge. In this paper multiple neural networks based hybrid approach has been proposed to predict most probable wildfire locations on a monthly scale. Hybrid architecture constructed with five different neural networks, namely, Feed Forward Back Propagation, Multi-Layer Perceptron, Radial Basis Function, Elman, and Probabilistic networks were trained and tested using segmented image data. Main aspect of this study was to establish a methodology to predict wild fire prone locations as spots on the Australian map based on publicly available gridded maps of Australian weather variables and hydrological variables. Images were created from the gridded maps available from Australian Water Availability Project (AWAP) and Bureau of Meteorology (BOM). On the other hand, NASA-MODIS historical active fire image archives for Australia were used as training targets. It was interesting to explore the possibility to learn from the colour information on the various input maps and correlate that with the historical wild fire spots (represented by red dots) on the Australian maps, available from NASA-MODIS. These were used to train and test the neural networks. AWAP and BOM provided important and practical hydrological and environmental attributes whereas NASA-MODIS provided ground truth as training targets. The monthly average data of total evaporation, sensible heat flux, precipitation, incoming solar irradiance, maximum temperature and soil moisture from AWAP were used in the study. On top of these, BOM data consists of monthly average wind speed, vapour pressure and relative humidity. Three year period 2008-2010 were studied where all the monthly images were gathered to develop the complete data set, including 9 input images and 1 target image for every month of the three consecutive years. 70% of the data were used (2008-2009) were used to train all the neural networks where as the monthly image data from 2010 (30% of the data set) were used to test the networks. Independent training and testing were critical for this study to prove the generalization capability of the hybrid architecture based on the neural networks. 94% overall prediction accuracy was achieved from this approach, with 93% sensitivity and 95.3% specificity. Maximum false positive rate was 0.7% whereas overall precision was 96%. The results were very encouraging in predicting probable Australian wildfire locations and represent visually with high overall accuracy.