Enhancing Bushfire Classification in Satellite Imagery through Gradient Based Channel Selection

Bushfire detection is a critical task for environmental monitoring, early warning systems and disaster management. State-of-the-art methods utilize deep learning, particularly Convolutional Neural Networks (CNN), for detecting bushfires or mapping burned areas from satellite imagery. Satellite images often contain more than 10 channels, making the selection of appropriate channels a significant challenge. Traditional methods for channel selection in remote sensing often depend on intuition, domain-specific knowledge or trial and error. In contrast, we argue that Deep Learning models can identify previously unseen patterns in the data that may not align with human intuition. These conventional methods tend to use more than three channels, which can significantly increase the computational resources required. There is a clear need for a mathematical automated approach to select the most relevant channels for bushfire classification. This study proposes a novel gradient-based method for identifying the most appropriate channels for bushfire classification using satellite imagery. Unlike traditional approaches, the proposed method leverages gradient information to automatically rank channels based on their contribution to classification accuracy. The method was applied to a combined EuroSAT and Landsat 8 Active Fire Dataset. Based on the results, the NIR, Red, and Blue channels were selected, achieving an accuracy of 96.63%. This approach reduced the number of parameters trained and minimized computational resource usage. Furthermore, the selected channels and resulting classification performance were compared with other state-of-the-art methods. The results demonstrate that the gradient-based method enables optimal channel selection, significantly improving efficiency while maintaining high accuracy. This approach advances the field of remote sensing by offering a scalable and interpretable solution for bushfire monitoring and decision-making.