Answers are grounded in this paper and cite their sources.
Forest fires pose severe threats to ecosystems, human life and socioeconomic development, rendering effective risk assessment essential for proactive prevention and timely management. Existing approaches generally rely on fixed-scale feature extraction, limiting their ability to simultaneously capture fine-scale spatial details and broad contextual information in complex mountainous environments. This study proposes a forest fire multi-scale adaptive attention network (FMAAN) for forest fire risk assessment. The core multi-scale adaptive module determines appropriate scale combinations and to fuse the corresponding features adaptively, adjusting the effective receptive field for each input sample to capture fire-driving information across spatial scales. Deformable convolutions and a two-stage attention mechanism are further incorporated to enhance geometric adaptability and emphasise critical fire-risk features. The model is evaluated in the Anning River Valley, southwestern Sichuan Province, China. Results demonstrate that FMAAN outperforms all baseline models across multiple evaluation metrics, achieving 90.48% accuracy and an F1-score of 90.99%. Risk maps reveal pronounced seasonal variability and spatial heterogeneity, with high-risk areas concentrated in Xichang, Dechang and Mianning. SHAP-based interpretation further identifies NDII7 and relative humidity as the dominant factors influencing model assessments. FMAAN supports targeted fire prevention and resource allocation and informs forest fire risk assessment in mountainous regions.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: