Wildfire detection powered by involutional neural network and multitask learning with dark channel prior technique

Abstract

Wildfires are crucial to the ecosystem health, yet wildfires also pose a grave threat to human lives and the environment. Early detection remains challenging, especially in remote areas. Satellite data and autonomous aerial vehicles (AAVs) have become crucial tools for detecting and monitoring wildfires. Achieving proper identification is demanding due to the intricacy of wildfires, which are impacted by various environmental factors. Recent advances in remote sensing data analysis, especially the use of machine learning models, demonstrate captivating potential. Applying machine learning in remote sensing data analysis, this study examines state-of-the-art wildfire identification strategies based on the modified involutional neural network (Inv-Net). Inv-Net is used for wildfire segmentation and classification that enhances the U-Net architecture by incorporating the involutional neural network (INN) as its backbone. The INN layers dynamically perform kernel alterations, allowing for more efficient extraction of features and improved capture of spatial correlations. By utilizing the concept of multitask learning, Inv-Net achieves an accuracy of 98.1% in the classification task and exhibits a noteworthy Intersection over Union score of 97.6% in the segmentation task, notably when used for wildfire image analysis. To improve wildfire detection accuracy and resilience, Inv-Net uses the synergies between classification and segmentation. Furthermore, the dark channel prior (DCP) technique is used to address atmospheric disturbances such as haze and fog in static images. DCP reduces optical interference by estimating depth maps and transmission maps. The results highlight the importance of the DCP method, which achieves an excellent structural similarity index measure of 0.7831 and peak-signal-to-noise ratioof 17.4321, indicating greater image dehazing methods. In comprehensive trials and comparisons, we demonstrate the robustness of our approach in identifying and mapping wildfires in challenging conditions.