Abstracting Geo-specific Terrains to Scale Up Reinforcement Learning