Rooftop agriculture is a good practice to improve food security in cities by using under-utilized surfaces of buildings. Beyond food production, it promotes community based sharing, income generation, and urban greening through surplus sales, exchanges, or donation exchanges within communities. Ineffective matching between surplus and demand at proximity level generates waste and induces economic loss. This paper introduces a hybrid surplus-demand matching approach which combines classical graph-theoretic optimization with Graph Neural Network (GNN) based link prediction on a case of rooftop farming networks in the city of Dhaka. We modeled a realworld dataset, which includes spatial, temporal and categorical features, as a bipartite graph. Optimal initial matches were extracted based on MWBM, whereas GCN, GAT, and GraphSAGE models are used to learn the latent node representations and suggest further feasible matches. Experiments demonstrate that GraphSAGE significantly outperforms classical and other GNN baselines. The model presented shows that the integration of deterministic optimization and data-driven graph learning can enhance the surplus sharing, minimize the biomass wasted, and sustain urban agriculture.