【摘要】 强对流等灾害性天气给人民生活和社会经济发展造成了严重影响，准确理解强对流发生的机理及提高其预报效果仍然是具有挑战性的工作。综合利用我国自主研发的新一代地球静止轨道气象卫星风云四号高时空分辨率观测数据和中国气象局全球数值预报(China Meteorological Administration-Global Forecast System, CMA-GFS)格点化产品，研究局地对流发生前大气环境场的特征和关键影响因子的变化。分析表明：卫星观测得到的云顶冻结信息以及表征大气的不稳定性、水汽含量等数值模式变量是预测局地对流发生的重要因子。利用面积重叠法和光流法对云团进行连续追踪，采用机器学习技术建立了中国区域局地对流发生和强度分级(弱、中、强)预警模型2.0版本(Storm Warning In Pre-convective Environment Version 2.0, SWIPE-V2.0),实现了局地对流的智能化预警。独立检验结果表明：模型对6个不同分区的雨季8 mm/h以下强度降水相关的对流判识准确率在0.5～0.85,对8 mm/h以上强度降水相关的对流判识准确率在0.69～0.91之间，具有较好的提前预警效果和实际应用价值。目前，SWIPE-V2.0已投入实时应用。更多还原

【Abstract】 Local severe convective storms significantly impact people’s lives and socio-economic development. Understanding the mechanism of severe convective storms and predicting the occurrence and development of local severe storms remains challenging. We investigate local severe convective storms’ environmental and thermodynamic characteristics in pre-convection environments by combining observations from China’s new-generation geostationary satellites(FY-4 series), which offer high spatial-temporal resolution, with numerical weather forecast products from the China Meteorological Administration(CMA) global forecast system(CMA-GFS). Furthermore, we explore how their changes impact the future development intensity of local convection. Results show a close association between cloud top cooling information from satellite observations and numerical prediction model variables, such as atmospheric instability and water vapour content, with convection storm occurrence and intensity. Changes in these factors closely relate to the future development intensity of local convection. The Storm Warning In Pre-convective Environment Version 2.0(SWIPE-V2.0) system aims to predict the occurrence and intensity(weak, medium, and strong) of local severe storms in China. Established using machine-learning techniques, SWIPE-V2.0 employs the brightness temperature threshold method and area threshold method to identify the local convective cloud. Meanwhile, it uses the overlapping and optical flow methods to track the movement of the local convective cloud. The machine-learning model uses the CLDAS(Global Land Data Assistance System) data of the multi-source precipitation fusion dataset, obtained half an hour to one hour after the cloud cluster, as the training tag. Independent validation results reveal SWIPE-V2.0’s strong performance in early warning for local convective storms, with recognition rates of 0.5-0.85 for cases with precipitation below 8 mm/h and 0.69-0.91 for cases with precipitation above 8 mm/h in the rainy season across six different regions. In the non-rainy seasons, across the same regional spread, recognition rates are 0.53-0.98 for cases with precipitation below 8 mm/h and 0.77-0.99 for cases with precipitation above 8 mm/h. Early warning results from SWIPE-V2.0 on real-time local convection systems demonstrate its potential for near real-time applications, while also indicating its useful role in understanding the environmental factors associated with local severe storms across various weather regimes. Currently, we are utilising SWIPE-V2.0 in real-time applications.更多还原