Advanced Deep Learning Framework for Improved Wildfire Detection and Aerosol Identification Using Active Satellite Imagery

Wildfires rank among the most prevalent natural disasters globally and have emerged as a significant factor in climate change over the past decade. Early detection of wildfires and smoke plumes through satellite imagery is crucial since they are not easily extinguishable which may lead to catastrophic consequences for both wildlife and forest ecosystems. Classic deep-learning models for wildfire and aerosol identification have shown significant progress, but high false-positive rates remain a key limitation. This paper proposes a custom-designed Convolutional Neural Network (CNN) model that aims to improve the identification of wildfires and aerosols, leveraging satellite imagery categorized into cloud, dust, haze, land, seaside, and smoke. Moreover, we considered popular deep learning and transfer learning models, specifically EfficientNet, MobileNetV3, and Inception V3, to identify and distinguish smoke plumes. Hyper-parameter tuning has been performed to achieve more accurate results. The metrics evaluated in this research are accuracy, precision, recall, and f1-score. A comprehensive analysis was performed that aimed to identify the best transfer learning model and the model that closely aligns with the performance of the CNN built. This paper provides valuable insights into the potential use of transfer learning models and of the proposed custom-designed CNN model in early wildfire detection by identifying the smoke plumes and helping in reducing false fire alarms.