The performance of two-stage remote sensing object detection methods largely depends on the quality of oriented region proposals generated in the first stage. However, most existing methods’ region proposal networks face significant challenges: 1) Manually designed anchor boxes struggle to balance recall and the number of anchors due to variations in scale, aspect ratio, and orientation of targets in remote sensing images. This also necessitates separate anchor box designs for different datasets, causing inconvenience. 2) Horizontal bounding boxes fail to capture the orientation of objects, leading to misclassification of some negative samples as positive in label assignment, thereby impairing training. Additionally, static label assignment strategies can cause misalignment between classification and localization tasks. To address these issues, we propose an Adaptive Anchor-based Region Proposal Network (AA-RPN). Without the need for predefined anchor boxes, our approach dynamically generates Adaptive Anchor Boxes (AAB) based on the target’s scale, aspect ratio, and orientation during training. The proposed Dynamic Label Assignment (DLA) strategy assigns labels dynamically based on center priors and network outputs. Additionally, we introduce an Adaptive Feature Pyramid Network (AFPN) to provide specific scale contexts for different targets. Using Oriented R-CNN as a baseline, extensive experiments on public benchmarks demonstrate significant improvements in accuracy and computational efficiency.