Objective In the context of urban regeneration, green roofs represent a critical strategy for improving urban ecological environments. Existing studies typically treat rooftop greening as isolated elements and over-rely on large parks as cores for constructing ecological networks, thereby overlooking the synergistic holistic effects of integrating potential rooftop greening with existing green spaces. Precisely identifying the greening potential of building rooftops and quantifying their critical role in optimizing ecological network connectivity within high-density urban areas will provide a scientific foundation for urban ecological development.

Methods This research presents a systematic approach for the extraction of potential green roofs and the assessment of ecological network reconstruction, with the central urban area of Tianjin as an example. By synthesizing four key suitability characteristics including flat roof features, roof color, additional structures, and building height, the research develops a deep learning model that integrates attention mechanisms and multi-scale feature fusion strategies to efficiently identify potential green roof areas. Furthermore, based on the Least-Cost Path (LCP) model, the research proposes a methodological framework for incorporating potential green roofs into urban ecological network reconstruction. The contribution of potential green roofs to urban ecological network optimization is systematically assessed from two dimensions (basic elements and overall structure) through three aspects: source patch extensibility, corridor connectivity, and corridor connectivity index (CCI) of source patches.

Results The research findings reveal several key aspects. 1) In the central urban area of Tianjin, 21,244 roofs are identified with greening potential, covering a total area of 1,345.22 hm², with an average area of 633 m² and the largest potential green roof spanning 24,927 m². These potential sites exhibit a “polycentric and dispersed” spatial distribution pattern. 2) The integration of potential green roofs with existing green spaces has significantly enhanced the urban ecological network, with the number of source patches increasing from 100 to 131, and the total area increasing by 302.35 hm². Notably, potential green roofs directly contribute 126.28 hm². The number of ecological corridors has expanded from 4,950 to 8,515. The average cumulative resistance of corridors increases slightly — likely due to the introduction of new corridors traversing high-resistance areas — while the decreased average corridor length and a 72% increase in network density indicate the formation of a more compact ecological network, suggesting enhanced opportunities for species dispersal between source patches through more diverse and shorter pathways. 3) Potential green roofs expand source patches through two mechanisms: First, they have enlarged 67 existing source patches, contributing 61.34 hm² (7% of the expanded source patches). Second, they have facilitated the formation of 31 new source patches, adding 64.94 hm² (30% of the newly formed source patches). The analysis demonstrates that potential green roofs make significant contributions to newly-formed source patches. These green roofs not only synergize with existing green spaces to exceed critical thresholds for creating new source patches, but also account for nearly one-third of the total area of these newly-formed patches. The improvement in corridor connectivity is achieved through increased existing − incremental (E-I) and incremental − incremental (I-I) connections. E-I corridors, representing connections between new and existing source patches, have increased by 3,100, while I-I corridors, reflecting connections between incremental source patches, have added 465 new links. These new connections demonstrate superior performance metrics compared to the existing urban ecological network, indicating expanded spatial coverage and enhanced network hierarchy with more diverse dispersal pathways. From a functional perspective, the average CCI of source patches increases by 53%. The CCI values of source patches are redistributed, resulting in two distinct phenomena: Some source patches with originally low dispersal capability gain increased connectivity opportunities and become key bridging areas, while others with high dispersal capability experience reduced network control due to the influence of surrounding and newly added nodes. Among these, expansionary source patches demonstrate higher mean CCI values compared to other types, and work synergistically with newly-added source patches to enhance the overall redundancy and resilience of the ecological network.

Conclusion In conclusion, although individual potential green roofs may have limited area, their integration with existing green spaces can effectively expand urban ecological sources, optimize corridor connectivity, and enhance network functionality, thereby providing a significant technical approach for optimizing urban ecological networks. This research presents two major innovations: First, it transforms the traditional multi-step identification process by developing an end-to-end multi-task deep learning network that directly identifies potential green roofs based on their key characteristics, breaking through the existing complex process of “complete extraction followed by stepwise elimination”; second, the research presents methods to assess how scattered, small-scale green roofs can strengthen urban ecological networks. Through quantitative studies, the research demonstrates how potential rooftop gardens contribute to enhancing city ecosystems.