中文

打破极端对流检测中的统计相似陷阱

机器学习 2025-09-12 v1 计算机视觉与模式识别

摘要

当前深度学习天气模型的评估指标创造了 ``统计相似陷阱'', rewards 模糊的预测同时忽略稀有且高影响事件。我们提供了量化证据, 显示 sophisticated baselines achieve 97.9% correlation yet 0.00 CSI for dangerous convection detection。我们引入 DART (Dual Architecture for Regression Tasks), 一个解决将粗糙大气预报转化为针对极端对流检测优化的高分辨率卫星亮度温度场的框架。DART 采用 dual-decoder architecture with explicit background/extreme decomposition, physically motivated oversampling, 和 task-specific loss functions。我们呈现四项关键发现: (1) 在多个 sophisticated baselines 上对统计相似陷阱进行经验验证; (2) ``IVT 矛盾'', 移除 Integrated Water Vapor Transport, 广泛认为是大气河流分析的 essential 因素, 可将极端对流检测提高 270%; (3) 通过 operational flexibility (DART achieves CSI = 0.273 with bias = 2.52 vs. 6.72 for baselines at equivalent CSI) 说明 architecture 的必要性; (4) 以 2023 年 8 月孟加拉河灾难 as a case study 进行 real-world validation。 To our knowledge, this is the first work to systematically address this hybrid conversion-segmentation-downscaling task, with no direct prior benchmarks identified in existing literature。Our validation against diverse statistical and deep learning baselines sufficiently demonstrates DART's specialized design。该框架通过 beta-tuning 启用 precise operational calibration, 在 standard hardware 上训练时间 < 10 分钟, 并 seamless integration with existing meteorological workflows, 显示出面向可信赖 AI 的极端天气准备的通往道路。

关键词

引用

@article{arxiv.2509.09195,
  title  = {Breaking the Statistical Similarity Trap in Extreme Convection Detection},
  author = {Md Tanveer Hossain Munim},
  journal= {arXiv preprint arXiv:2509.09195},
  year   = {2025}
}

备注

43 pages, 7 figures