DDQN-GeoAWN: multi-objective adaptive scheduling for geospatial analysis workflows in cloud environments using deep reinforcement learning

Large-scale geospatial workflow scheduling in cloud environments is challenging because heterogeneous data, workflow dependencies, and spatiotemporal correlations interact with dynamic task arrivals and fluctuating resource loads. To address this, we propose DDQN-GeoAWN, a multi-objective adaptive scheduling method that employs deep reinforcement learning to model the online decision process. Its core design follows a closed-loop logic: a multi-priority feedback queue (MPFQ) dynamically organizes incoming workflows and captures queue-state information; this queue-state information, together with resource status and workflow spatiotemporal features, is ingested by a geospatial adaptive weighting network (GeoAWN) to generate adaptive reward weights for completion time, load balancing, and energy consumption; these weights are incorporated into the reward function of the double deep Q-network (DDQN), which then executes real-time scheduling decisions and policy updates, closing the loop. Experiments on two representative scenarios show that DDQN-GeoAWN outperforms the best baseline, reducing completion time by 9.8%–11.6%, load imbalance by 19.2%–22.8%, and energy consumption by 8.6%–9.3%. The results demonstrate that state-adaptive scalarization can effectively coordinate competing scheduling objectives and improve overall resource scheduling performance for geospatial cloud workflows.