We introduce a challenging real-world planning problem where actions must be\ntaken at each location in a spatial area at each point in time. We use forestry\nplanning as the motivating application. In Large Scale Spatial-Temporal (LSST)\nplanning problems, the state and action spaces are defined as the\ncross-products of many local state and action spaces spread over a large\nspatial area such as a city or forest. These problems possess state\nuncertainty, have complex utility functions involving spatial constraints and\nwe generally must rely on simulations rather than an explicit transition model.\nWe define LSST problems as reinforcement learning problems and present a\nsolution using policy gradients. We compare two different policy formulations:\nan explicit policy that identifies each location in space and the action to\ntake there; and an abstract policy that defines the proportion of actions to\ntake across all locations in space. We show that the abstract policy is more\nrobust and achieves higher rewards with far fewer parameters than the\nelementary policy. This abstract policy is also a better fit to the properties\nthat practitioners in LSST problem domains require for such methods to be\nwidely useful.\n