The field of weather forecasting makes extensive use of radar big data to extract information about precipitation, storms, lightning, and other weather phenomena to aid in the prediction and monitoring of weather changes. To improve the quality of radar data, machine learning and fuzzy logic algorithms are often used to identify and classify non-meteorological clutter in weather data. However, these methods often require dozens of texture features as inputs and need to manually adjust the thresholds to cope with different clutter types, which leads to significant time costs. In this paper, we propose a multi-scale weighted connected UNet to address these challenges by combining the channel attention feature fusion module and the UNet structure model. The task of recognizing non-meteorological clutter is regarded as a semantic segmentation problem, which eliminates the need to manually set thresholds for clutter pixel-level classification. Additionally, the channel-focused feature fusion mechanism is able to analyze the deep latent features of the input parameters and suppress the useless features, so that only six polarization parameters are required as inputs. Furthermore, the model incorporates full-scale deep supervision to improve the edge segmentation accuracy of clutter and meteorological echoes. Experiments confirm that our proposed model outperforms the compared models in clutter identification with Critical Success Index (CSI) of 0.808.