Scholarlink

<p>Fire risk detection is critical for timely interventions and effective management strategies in mitigating wildfire impacts. This study examines the efficacy of many advanced models, emphasizing the Swin Transformer architecture for efficient fire detection. We assessed RGB input evaluations, highlighting the Swin_S model, which attained a test accuracy of 62.1%, and the Swin_T model at 61%. Comparative analysis with current models demonstrated that Swin_T_Edge surpassed its competitors, achieving the maximum accuracy of 66% and an F1 score of 0.587, confirming its efficacy in classification tasks while maintaining a balance in model complexity. Cross-dataset tests further illustrated the models’ durability across various fire conditions, underscoring the necessity for solid generalization capabilities in real-world applications. Statistical evaluations utilizing t-tests confirmed the substantial performance enhancements of the suggested models. The findings highlight the Swin_T_Edge model’s promise as a premier option for fire risk detection systems, recommending future improvements via ensemble learning and the incorporation of temporal data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Investigating the performance vs. computational complexity tradeoff in cross-domain fire risk detection

  • 期刊论文

  • 发表日期：

    2025年6月9日

查看全文
  • Abhishek Verma,
  • Virender Ranga,
  • Dinesh Kumar Vishwakarma

摘要

Fire risk detection is critical for timely interventions and effective management strategies in mitigating wildfire impacts. This study examines the efficacy of many advanced models, emphasizing the Swin Transformer architecture for efficient fire detection. We assessed RGB input evaluations, highlighting the Swin_S model, which attained a test accuracy of 62.1%, and the Swin_T model at 61%. Comparative analysis with current models demonstrated that Swin_T_Edge surpassed its competitors, achieving the maximum accuracy of 66% and an F1 score of 0.587, confirming its efficacy in classification tasks while maintaining a balance in model complexity. Cross-dataset tests further illustrated the models’ durability across various fire conditions, underscoring the necessity for solid generalization capabilities in real-world applications. Statistical evaluations utilizing t-tests confirmed the substantial performance enhancements of the suggested models. The findings highlight the Swin_T_Edge model’s promise as a premier option for fire risk detection systems, recommending future improvements via ensemble learning and the incorporation of temporal data.