Scene-aware attention network and medium-scale drone dataset for diverse fire classification

Vision-based fire detection is a critical task for ensuring urban safety and protecting buildings against potential hazards. Although mainstream approaches practice conventional attention modules with deep models that are well-suited for static fire environments, they have limited outcomes in their diverse nature. These attention modules are mainly composed of conventional convolutional layers that generate redundant and overlapped spatial patterns in the case of complex fire scenarios, especially remote sensing views. To overcome these challenges, an intelligent vision model is desirable which has the capabilities to differentiate complex fire scenes with contextual information, particularly from remote sensing viewpoints. To address these challenges, an Adaptive Fire Network (AFN) is proposed that efficiently discriminates diverse fire scenes, geographical location and environmental conditions. To obtain more informative spatial features, a lightweight ConvNeXtBase is integrated with the Scene-Aware Attention Network (SAAN). Moreover, SAAN is mainly composed of the Adaptive Channel Module (ACM) and the Adaptive Spatial Module (ASM) to effectively capture global dependencies across channels, followed by local spatial patterns associated with fire. In the SAAN, depth-wise separable convolution is utilized which allows the model to gather more enhanced features from the input tensor. Additionally, the channel-wise processing leads to more understanding of the input features, potentially enhancing the model’s ability to make accurate predictions. Extensive experiments were conducted on three challenging fire datasets, Foggia, FireNet, and the newly developed Drone Fire Vision (DFV), achieving accuracies of 99.96%, 98.97%, and 90.78%, respectively. The results demonstrate that AFN significantly enhanced existing state-of-the-art methods, advancing fire detection and classification capabilities in dynamic and remote sensing environments.