Reinforcement learning guided active crop localization with CNN detectors and an interactive decision dashboard for precision agriculture

Accurate crop localization is a foundational requirement for automated field monitoring in precision agriculture, yet conventional one-shot convolutional neural network (CNN) detectors frequently yield misaligned or incomplete bounding boxes under variable illumination, occlusion, complex canopy structures, and heterogeneous plant morphology. To overcome these limitations, we propose a reinforcement learning (RL), guided active crop localization framework that reformulates localization as a sequential decision-making problem, enabling adaptive refinement of initial detector predictions through learned geometric transformations. The framework integrates three CNN backbones, YOLOv12, VGG16, and ConvNeXt-V2, with three deep RL agents (DQN, PPO, and SAC), where the agent state combines visual embeddings, normalized bounding-box coordinates, and detector confidence, and optimization is driven by a continuous ΔIoU reward function. Extensive experiments across nine backbone agent combinations demonstrate that RL-based refinement consistently outperforms static detectors. The YOLOv12 + DQN configuration achieves the highest localization accuracy, with a Mean IoU of 0.969 ± 0.071, a 100% localization success rate (IoU ≥ 0.5), and convergence typically within a single refinement step. In contrast, VGG16 and ConvNeXt-V2 based configurations yield Mean IoU values ranging from 0.60 to 0.75, while requiring substantially higher computational cost. Importantly, the YOLOv12 + DQN model maintains high operational efficiency, completing inference in ≈0.09 s per sample, making it suitable for real-time agricultural applications. Beyond algorithmic performance, we deploy the proposed model within an interactive decision support dashboard that visualizes bounding-box refinement trajectories, Grad-CAM attention dynamics, and class-level localization statistics, facilitating transparent model interpretation and practical adoption by agricultural practitioners. Overall, the proposed RL-guided active localization framework delivers substantial accuracy gains, low-latency performance, and operational interpretability, establishing a scalable and actionable solution for next-generation precision agriculture systems.