Hazardous events related to atmospheric precipitation depend not only on the intensity of surface precipitation, but also on its type. Uncertainty related to determination of the precipitation type (PT) leads to financial losses in many areas of human activity, such as the power industry, agriculture, transportation, and many more. In this study, we use machine learning (ML) algorithms with the data fusion approach to more accurately determine surface PT. Based on surface synoptic observations, ERA5 reanalysis, and radar data, we distinguish between liquid, mixed, and solid precipitation types. The study domain considers the entire area of Poland and a period from 2015 to 2017. The purpose of this work is to address the question: “How can ML techniques applied in observational and NWP data help to improve the recognition of the surface PT?” Despite testing 33 parameters, it was found that a combination of the near-surface air temperature and the depth of the warm layer in the 0–1000 m above ground level (AGL) layer contains most of the signal needed to determine surface PT. The accrued probability of detection for liquid, solid, and mixed PTs according to the developed Random Forest model is 98.0%, 98.8%, and 67.3%, respectively. The application of the ML technique and data fusion approach allows to significantly improve the robustness of PT prediction compared to commonly used baseline models and provides promising results for operational forecasters.

中文翻译：

---

基于数据融合方法的地表降水类型确定

与大气降水有关的危险事件不仅取决于地表降水的强度，还取决于其类型。与确定降水类型 (PT) 相关的不确定性会导致许多人类活动领域的经济损失，例如电力工业、农业、交通等。在这项研究中，我们使用机器学习 (ML) 算法和数据融合方法来更准确地确定表面 PT。根据地表天气观测、ERA5 再分析和雷达数据，我们区分了液体、混合和固体降水类型。研究领域考虑了整个波兰地区以及 2015 年至 2017 年的一段时间。这项工作的目的是解决以下问题：“应用于观测和 NWP 数据的 ML 技术如何有助于提高对表面 PT 的识别？” 尽管测试了 33 个参数，但发现近地表气温和地表以上 0-1000 m (AGL) 层中暖层深度的组合包含确定地表 PT 所需的大部分信号。根据开发的随机森林模型，液体、固体和混合 PT 的累积检测概率分别为 98.0%、98.8% 和 67.3%。与常用的基线模型相比，ML 技术和数据融合方法的应用可以显着提高 PT 预测的稳健性，并为业务预测人员提供有希望的结果。结果表明，近地表气温和地表以上 0-1000 米 (AGL) 层中暖层深度的组合包含确定地表 PT 所需的大部分信号。根据开发的随机森林模型，液体、固体和混合 PT 的累积检测概率分别为 98.0%、98.8% 和 67.3%。与常用的基线模型相比，ML 技术和数据融合方法的应用可以显着提高 PT 预测的稳健性，并为业务预测人员提供有希望的结果。结果表明，近地表气温和地平面以上 0-1000 m (AGL) 层中暖层深度的组合包含确定地表 PT 所需的大部分信号。根据开发的随机森林模型，液体、固体和混合 PT 的累积检测概率分别为 98.0%、98.8% 和 67.3%。与常用的基线模型相比，ML 技术和数据融合方法的应用可以显着提高 PT 预测的稳健性，并为业务预测人员提供有希望的结果。