DAFDM: A Discerning Deep Learning Model for Active Fire Detection Based on Landsat-8 Imagery

Abstract

Monitoring active fire (AF) utilizing remote sensing imagery provides critical support for fire rescue and environmental protection. Traditional methods for detecting AFs rely on the statistical analysis of AF radiance and background features. However, these algorithms are resource-intensive to develop and exhibit limited adaptability, particularly in distinguishing AF from interference pixels. Deep learning (DL) technologies, which can extract deep features from images, offer a new solution for efficiently detecting AF. This article proposes an AF detection model based on convolutional neural networks, named DAFDM. By integrating multilayer features through an enhanced feature processing module, the model produces high-quality AF information, accurately detecting AF from the background. Given the presence of uncorrected false alarms in the training labels, it is challenging for DL models to distinguish interference pixels, we construct a Landsat-8 dataset encompassing various fire types and interference objects, with precise labels. Comparing several architectures, we find that only U-Net type models can discern the AF boundary pixels fully and accurately. The proposed method outperforms other AF detection algorithms, achieving IoU and F1-score of 87.28% and 93.21%, respectively. Experimental results demonstrate that DAFDM possesses robust generalization capability in distinguishing interference pixels. The incorporation of land surface temperature as auxiliary data further improves DAFDM's performance, with interpretability methods employed to elucidate the impact of input data on predictions. This method is anticipated to further contribute to AF monitoring and wildfire development pattern analysis.