面向轻量化的地基云图分割技术研究

Segmentation Technology of Ground-Based Cloud Image for Lightweight

云是气象领域比较重要的要素之一,准确快速获取云量,在天气预报、飞行安全、气候研究等方面均具有积极意义,因此为了快速准确地获得云量,就需要精准地分割云图。但是现有云图分割网络存在分割效果差、运算复杂度高、参数量大等问题,为了解决这些问题,提出一种轻量化的地基云图分割网络模型LGCSegNet。该模型利用Encoder-Decoder网络模型框架进行设计,利用通道拼接思想在通道深度上实现不同层次图像特征融合,避免损失特征边界,实现精确地分割地基云图,可获取准确的云量信息。在地基云图数据集HBMCD和HBMCD_GT上进行实验,经多组对比实验证明所提出的网络对云的表征能力更强,分割准确率比较高,可达到96.83%,分割的平均交并比可以达到86.00%,为实际应用提供了一定的理论基础。

Cloud is one of the most important elements in the field of meteorology.Accurate and rapid acquisition of cloud cover is of positive significance in weather forecast, flight safety, climate research, etc.Therefore, in order to obtain the cloud amount quickly and accurately, it is necessary to segment the cloud image accurately.However, the existing cloud image segmentation network has some problems, such as poor segmentation effect, high computational complexity and large number of parameters.In order to solve these problems, a lightweight ground cloud image segmentation network model LGCSegNet is proposed.This model is designed using the framework of Encoder-Decoder network model.By using the idea of channel splicing, the image features of different levels are fused on the channel depth to avoid the loss of feature boundary, which achieves accurate segmentation of ground cloud image and obtains accurate cloud cover information.The experiments are carried out on the ground-based cloud image data sets HBMCD and HBMCD_GT.Through several groups of comparative experiments, it is proved that the proposed network has stronger cloud characterization ability and higher segmentation accuracy that can reach 96.83%,and the average crossover ratio of segmentation can reach 86.00%,which provides a certain theoretical basis for practical application.