Urban green spaces are an important part of the urban ecosystem and hold significant ecological value. To effectively protect these green spaces, urban managers urgently need to identify them and monitor their changes. Common urban vegetation positioning methods use deep learning segmentation models to process street view data in urban areas, but this is usually inefficient and inaccurate. The main reason is that they are not applicable to the variable climate of urban scenarios, especially performing poorly in adverse weather conditions such as heavy fog that are common in cities. Additionally, these algorithms also have performance limitations such as inaccurate boundary area positioning. To address these challenges, we propose the UGSAM method that utilizes the high-performance multimodal large language model, the Segment Anything Model (i.e., SAM). In the UGSAM, a dual-branch defogging network WRPM is incorporated, which consists of the dense fog network FFA-Net, the light fog network LS-UNet, and the feature fusion network FIM, achieving precise identification of vegetation areas in adverse urban weather conditions. Moreover, we have designed a micro-correction network SCP-Net suitable for specific urban scenarios to further improve the accuracy of urban vegetation positioning. The UGSAM was compared with three classic deep learning algorithms and the SAM. Experimental results show that under adverse weather conditions, the UGSAM performs best in OA (0.8615), mIoU (0.8490), recall (0.9345), and precision (0.9027), surpassing the baseline model FCN (OA improvement 28.19%) and PointNet++ (OA improvement 30.02%). Compared with the SAM, the UGSAM improves the segmentation accuracy by 16.29% under adverse weather conditions and by 1.03% under good weather conditions. This method is expected to play a key role in the analysis of urban green spaces under adverse weather conditions and provide innovative insights for urban development.