GUARD-net: A geostationary satellite-based network for near-real-time active fire detection under class imbalance

Deep learning has substantially advanced satellite-based wildfire monitoring; however, near-real-time detection using geostationary satellite sensors remains challenged by severe class imbalance and domain shift across regions and satellite platforms. This study presents GUARD-Net, a geostationary satellite-based network for near-real-time active fire detection under class imbalance that integrates spectral, spatial, and topographic information within a convolutional framework. The model employs a bottleneck residual (BR) architecture together with focal loss (FL) to enhance sensitivity to rare fire pixels under highly imbalanced conditions. To enhance transferability, we introduce a heterogeneous-features weighted domain adaptation (HWDA) strategy, which aligns deep mixed features and shallow physical attributes (spectral/spatial/topographic) with a dynamic self-adaptive loss weighting scheme. Ablation experiments using Himawari-8/9 and Geostationary Operational Environmental Satellite-16 (GOES-16) data under controlled imbalance scenarios show consistent performance gains over baseline models. In baseline tests with a fire:non-fire ratio greater than 1:20, GUARD-Net attains F1 score > 0.94, Area Under the Precision-Recall Curve (PR-AUC) > 0.98, and Map-level Image Classification Efficacy (MICE) > 0.93; in a controlled sensitivity analysis with a prescribed 0.1% false-positive rate among non-fire samples, the 1:1000 tests yielded F1 score > 0.77, PR-AUC > 0.86, and MICE >0.71. Under the oracle checkpoint-selection protocol used for the six cross-domain transfer tasks, HWDA achieved unsupervised F1 score up to 0.955 (maintaining F1 = 0.863 with 10% unlabeled target data) and supervised F1 score up to 0.989 (reaching F1 score = 0.943 with 20% labeled target samples), with relative performance varying by transfer task and metric. Across 30 documented events from 2023, GUARD-Net detected 29 events; processing of the evaluated regional scenes required 2.5–3.9 min on the tested workstation, while comparisons with operational fire products varied by region. The results indicate that topographic inputs, focal-loss training, and heterogeneous-feature alignment can contribute to robustness across the evaluated terrain, class-imbalance, and domain-shift settings, supporting near-real-time wildfire detection from geostationary satellite observations.