Increasing the Statistical Significance for MODIS Active Fire Hotspots in Portugal Using One-Class Support Vector Machines

Abstract

Forest fires are an important part of the environment as they cause ecological and economical damage, apart from claiming numerous lives. Support Vector Machines (SVMs) coupled with satellite images and sensor information have been proven to aid in the prediction of forest fires and the extent of burnt areas. The National Aeronautics and Space Administration's (NASA) moderate resolution imaging spectroradiometer (MODIS) provides an active fire product in the form of spatial hotspots using an enhanced Collection 6 algorithm since 2015, which basically consists of active fires. This study proposes a one-class SVM based approach to identify the hotspots that are more likely to be a part of large forest fires. For this, Hotspots in Portugal in the time period 2000-2016 were analysed. 38 known large forest fire events across Portugal were used to train the VM model. As expected, it was observed that these forest fires generally occurred in spatio-temporal clusters when compared to the overall hotspot population. The input parameters for the model were based on spatial and temporal clustering behaviour of the hotspots, external factors such as land cover, temperature, elevation etc. and attributes from the Collection 6 data such as latitude, fire detection confidence, fire radiative power and brightness temperatures. Leave-one-out cross-validation and hold-out validation techniques were used to validate the model. As a result, 79.8 % of the overall hotspots from 2000 to 2017 were classified as "true" forest fires. These "true" detections were analysed over time. This led to the identification of other forest fire events that were not included in the training or the test dataset for the model. The results also showed that the model was more sensitive to land cover, temperature, FRP and cluster parameters when compared to other parameters. This proves that the model is responsive to active fires occurring on forest areas and also to spatio- emporal clustering. It is assumed that this approach could be successfully adapted to other study areas as there are no study area dependent input parameters used, except for the latitude, which needs to be taken into consideration. The model can be extended by using other input parameters such as humidity, slope etc. In addition, assigning higher weightage to the more influential parameters such as land cover and spatio-temporal clustering could lead to better results.