摘要
在对流单体的成熟阶段,冷性下沉气流作为一股冷空气,在近地面的底层向外扩展,与单体运动前方的暖湿气流交汇而形成飑锋,又称阵风锋。在获得CINRAD雷达的雷达数据基础之上,利用计算机视觉领域的技术可以进一步在雷达图像上进行自动识别和检测。阵风锋过境往往伴随着严重的灾害性天气,对提高此类灾害性天气的短临预警能力显得尤为重要,也对气象防灾减灾具有重要意义。阵风锋的自动识别任务目前还存在以下弊端:传统数理方法的识别率有待提高,由于阵风锋具有地域差异性,在雷达图像上目标小,特征不明显,要想精准识别还是有不小的难度,我国目前也没有一种真正推广到业务化的阵风锋自动识别算法。本文采用近年大火的深度学习方法,以Mask RCNN模型和resnet101主干网络为基础,在检测出有效阵风锋后,结合出现的问题以及阵风锋的特征,在结合通道和空间两个维度的注意力机制,提取目标细化特征,进行深层次的小目标检测。采用2013~2016四年里河南省各市的CINRAD天气雷达数据,利用Data Aug for Object Segmentation中的Data Augment for Label Me方法扩充了数据集的数量,增强了数据的多样性,考虑到阵风锋通常伴随着强雷暴天气,在雷达强度图上表现为一条窄带回波这一特征,设计使用Mask RCNN模型完成在雷达图上进行阵风锋自动识别,在检测出具体位置后分割出阵风锋的具体形态,可以进一步提高识别准确率。
During the mature stage of the convective monomer, the cold sinking air, as a cold air, extends out-ward at the bottom layer near the ground, meeting with the warm and wet air in front of the mon-omer movement to form a squall front, also known as the gust front. Based on the acquisition of radar data from CINRAD radar, we can be further automatically identified and detected on the field of radar images. Wind crossing is often accompanied by severe catastrophic weather, which is particularly important to improve the short warning ability of such catastrophic weather, and is also of great significance to meteorological disaster prevention and mitigation. Rafale front automatic recognition task still has the following disadvantages: The recognition rate of traditional mathematical and physical methods needs to be improved. Due to the regional difference of gust front as the target is small and the features are not obvious, there is still a considerable difficulty in accurately identifying it. Currently, there is no truly commercialized approach in China Automatic recognition algorithm for gust fronts. This paper adopts the deep learning method of recent fire, based on the Mask RCNN model and the resnet101 backbone network, after detecting the effective gust front, combining the emerging problems and the gust front features, combining the attention mechanism of the channel and space dimensions, extracts the target refinement features, and per-forms the deep small target detection. Using the CINRAD weather radar data of all cities in Henan province in 2013-2016 four years, the number of datasets was augmented using the Data Augment for Label Me method in Data Aug for Object Segmentation, enhanced by the diversity of the data. Given that the gusts front is usually accompanied by strong thunderstorms, the radar intensity diagram shows a narrow-band echo. The design uses the Mask RCNN model to complete the automatic gust front identification on the radar map. The specific morphology of the gust front is divided after the specific position is detected. The recognition accuracy can be further improved.
出处
《气候变化研究快报》
2023年第3期576-588,共13页
Climate Change Research Letters
