Forest fire susceptibility mapping with sensitivity and uncertainty analysis using machine learning and deep learning algorithms

Abstract

In the hilly region of the Western Himalayas, forest fires play a crucial role in forest destruction and biodiversity loss. Therefore, addressing the problem of forest fires is an immediate priority to mitigate the impact of forest fires. Therefore, the objective of this research is to assess and map the susceptibility zones of forest fire in the hilly state of Uttarakhand using geospatial techniques and machine learning algorithms (MLA) and deep Learning based sensitivity and uncertainty analysis of the forest fire ignition parameters. Thirteen forest fire ignition parameters were used to map forest fire susceptibility. Further, we used five distinct MLAs, including artificial neural networks (ANN), random forest (RF), logistic regression (LR), support vector machine (SVM) and an ensemble machine learning model. The generated models were validated using the receiver operating characteristic (ROC). The result shows that the ''very high'' and ''high'' zones for forest fires cover almost 55% of the total area of Uttarakhand. Validation of the models used indicated that all models produced most significant results. The ensemble machine learning had the highest accuracy (AUC = 0.977), followed by RF (0.973) and ANN (0.972). In addition, a Deep Neural Networks (DNN) based sensitivity approach is proposed to perform sensitivity and uncertainty analysis of the ignition parameters. For this, we have selected the best predictive model (Ensemble Machine Learning) to evaluate the sensitivity and uncertainty analysis. The sensitivity analysis shows that evapotranspiration has the highly influencing parameter to forest fires, followed by annual rainfall, wind speed, distance to agriculture, distance to tourist spots, distance to built-up and distance to road. However, aspect, temperature and distance to religious centers have less influence on susceptibility to forest fires in this study area. This study is unique because it provides new insights into the construction of an ensemble machine learning model and the integration of DNN-based sensitivity and uncertainty for robust and scientific modelling of forest fire susceptibility. The contribution of parameters demonstrates a novel approach in understanding the impact of a single parameter sample on forest fire susceptibility, which will help in developing prevention strategies. A similar approach may be used in other parts of the world with similar natural and anthropogenic conditions for mapping forest fire susceptibility.