The incidents of forest fires are escalating in frequency and intensity because of climate change and land use pressures, thus posing serious risks to ecosystems, biodiversity, and human safety. Detecting and segmenting forest fire regions accurately from aerial imagery remains a major challenge due to variations in fire size, smoke density, terrain complexity, and illumination. Traditional convolutional models often struggle with these heterogeneous conditions, leading to inaccurate or delayed responses. This study introduces DeepFire S3GA-Net, a novel deep learning-based segmentation framework designed for forest fire segmentation using unmanned aerial vehicle imagery. The architecture is built on a fully convolutional encoder-decoder network, enhanced with atrous spatial pyramid pooling in the encoder for capturing multiscale contextual features and group convolution modules in the decoder to improve spatial refinement and feature diversity. These components allow the model to distinguish both fine-grained fire edges and broad burned regions effectively. Extensive evaluation using standard metrics of mean intersection over union, accuracy, precision, recall, and Dice coefficient show that DeepFire S3GA-Net consistently outperforms existing models. Its robustness across complex visual environments highlights its potential as a scalable and reliable tool for real-time forest fire monitoring and ecological risk management.
Akhyar et al. (Thu,) studied this question.