The protection of the meteorological detection environment is a pivotal responsibility for meteorological stations, closely related to the accuracy of meteorological data and the quality of meteorological forecasts. Targeting the subtle boundary changes of naturally growing obstacles like trees and crops in time-series images, this study proposes an implanted boundary perception module that focuses on extracting obstacle boundary feature information in a supervised and guided manner. The boundary features are then integrated with backbone network features through the boundary-guided module to enhance the model’s attention to the boundaries. The experiment selects the lightweight model HRNet-W18 as the baseline for the method and validates it on a real meteorological detection environment dataset. The boundary guidance method effectively improves the mIoU by 2.29% for tree and crop obstacle categories, achieving an overall segmentation accuracy of 95.82%. This performance surpasses 14 other mainstream semantic segmentation methods and achieves a better balance between model complexity and accuracy.