A Position- and Scale-Aware Object-Agent Classifier via Deep Reinforcement Learning for Remote Sensing Imagery

Deep object classification is an effective method for land-cover classification in remote sensing images, requiring fewer samples and achieving clearer classification boundaries. However, determining the size and position of the image patches is crucial for accurate deep object classification, and this is often reliant on expert knowledge and trial and error, which limits its general applicability. In this article, a position- and scale-aware object-agent classifier for remote sensing imagery—named ObjectAgents—is proposed. Unlike prior pixel-level or global DRL approaches, this work pioneers an object-centric decision-making framework. The task of selecting representative image patches is modeled as an environment for multi-agent system interactions suitable for deep reinforcement learning (DRL). ObjectAgents consists of two agents: a PositionFilterAgent (PFA) and a ScaleControlAgent (SCA). These two agents work together to complete the classification process by observing the state of the environment and taking their own actions. The PFA is responsible for selecting optimal candidate points in the image for classification, while the SCA identifies the ideal image scale for accurate object identification. By using a learning mechanism where the agents receive rewards based on how much their actions improve the classification performance, the framework optimizes and learns the best decision-making strategies through DRL, enhancing the decision-making process and leading to better generalization in complex remote sensing scenarios. Experiments on IowaNet (1 m), the ISPRS Vaihingen benchmark (9 cm), and a constructed Global Tidal Wetlands dataset (10 m) demonstrate that ObjectAgents consistently outperforms manually designed baselines and remains robust in complex heterogeneous regions.