Forest Fire Classification and Detection in Aerial Images using Inception-V3 and SSD Models

Abstract

Wildfires contribute to a lot of damage than what is visible. A significant population of a wide variety of species can easily disappear in a single wildfire if unnoticed. The human population including tribes, forest department staff, and rescuers too often lose their lives in wildfires. The unnoticed damage includes high carbon emission, heat waves, the amount of time that takes to rebuild such a green environment, etc. The impact of such wildfires can be highly reduced if detected at an earlier stage. This study has used a Convolutional Neural Network (CNN) based Inception-V3 model, which classifies a given aerial image based on the presence of fire or smoke, and a Single Shot Detector model that detects the fire or smoke areas in the image. These models were trained on aerial imagery using transfer learning which led to an overall accuracy of 88% for classification and 91% for detection. These models can be used to detect fires in live images captured by aerial vehicles helping the disaster management entities to react and respond immediately and accordingly.