A novel framework of reverse mining of multi-level co-location patterns was proposed aiming at the problems that traditional methods of multi-level co-location pattern mining did not consider the grid characteristics of the real data distribution, and the multi-level mining framework from global to regional led to the algorithm inefficiency. The regional co-location patterns were first mined, and the global co-location patterns were deduced based on the mined regional patterns. Some pruning strategies were proposed to enhance the mining efficiency. The grid characteristics of the data distribution in real datasets were considered, and the grid neighbor relationship between instances was defined. The concept of grid spatial cliques with a novel method for calculating grid spatial cliques was defined. An adaptive grid density peak clustering strategy for partitioning regions was proposed in the regional division stage, and clusters were assigned based on the similarity of two-size grid spatial cliques. Extensive experiments were conducted on both synthetic and real-world datasets. The experimental results validated the effectiveness, efficiency and scalability of the proposed method. A pruning rate of up to 78% was achieved on real datasets.