<p>Accurate real-time segmentation of urban street-view green landscapes is an essential prerequisite for ecological monitoring and urban planning. However, street-view vegetation targets exhibit complex fractal boundary structures and are affected by seasonal appearance variations and multi-scale morphological differences, making it challenging for existing instance segmentation methods to simultaneously achieve segmentation accuracy and real-time performance. To this end, this paper proposes Mask-MMG, a lightweight real-time instance segmentation method specifically designed for urban street-view green landscapes. First, a cross-stage feature fusion mechanism incorporating re-parameterized networks is designed to enhance vegetation texture extraction through efficient local connection patterns while reducing computational overhead. Second, a multi-scale feature aggregation strategy based on a bidirectional focus propagation mechanism is proposed to effectively handle scale variations from leaves to canopies while maintaining low computational cost. Third, a lightweight convolutional architecture based on group operations and channel shuffling is introduced to address the challenges of irregular vegetation boundaries and rich fine-grained features. Experimental results demonstrate that the proposed algorithm achieves 84.7% Mask mAP@0.5 and 84.2% MIoU at 110.7 FPS on the CamVid dataset, reducing model parameters by 52.8%, computational cost by 62.6%, and memory footprint by 61.2% while maintaining segmentation accuracy. On the Cityscapes dataset, it attains 84.2% Mask mAP@0.5 and 82.3% MIoU. Furthermore, the 56.4% Mask mAP@0.5 achieved on the COCO dataset demonstrates state-of-the-art performance among lightweight models of comparable computational scale. Crucially, the algorithm sustains 81.4 FPS when deployed on resource-constrained edge computing hardware, validating its practical real-time capability for supporting large-scale urban vegetation monitoring systems under limited computational resources.</p>