Wildfire Impact Analysis and Spread Dynamics Estimation on Satellite Images Using Deep Learning

Abstract

Wildfires are a natural disaster that results in significant harm and catastrophic destruction. Forest areas tend to be more prone to the devastating effects of wildfires. Global warming causes wildfires to occur more frequently and with severe effects, forcing them to spread across wide amount of land areas, causing unimaginable harm and even claiming lives. In this paper, we propose a novel methodology to analyze the effects of wildfire and estimating its probability to spread using satellite data. The severity of wildfire is determined through fire and smoke detection via deep learning approach Modified-Residual Unet. To categorize areas based on their susceptibility to wildfires, NDVI imagery is given to the ZFNet classifier which determines the region's risk of being prone to wildfire. It achieves an impressive accuracy of 98.3% proving its ability in classifying wildfire risk. A novel Deep Probabilistic (P) Learning along with Cellular Automaton and Diffusion Limited Aggregation Algorithm is used to simulate the spread of wildfires and estimates are made by Anisotropic Generalized Regression Neural Network for the impacted areas. Thus, the efficiency of this novel approach has been tested with various datasets and our approach proves to have notable merits with greater accuracy and substantially lesser time when compared to other methods.