Due to the great application value of agricultural greenhouses in modern agriculture, detecting agricultural greenhouses in remote sensing images has attracted much attention. However, agricultural greenhouses in remote sensing images have remarkable features of large-scale variations, arbitrary orientations, and dense distributions, which pose significant challenges for fast and accurate detection. To address this problem, this study proposes YOLO-GH, an enhanced model based on YOLOv11n for detecting oriented agricultural greenhouses. First, a multi-scale bidirectional attention module is proposed to handle the characteristics of greenhouses in remote sensing images, such as arbitrary orientations and scale variations. This module is incorporated into the backbone network to enhance vertical and horizontal feature extraction while improving multi-scale bidirectional feature representation. Subsequently, a lightweight multi-scale attention module is designed to better capture multi-scale greenhouse features in the feature fusion network. The study also introduces adaptive rotated convolution to replace standard convolution in certain backbone layers, enabling adaptive extraction of rotation-invariant greenhouse features from different images. Additionally, this study constructs a new greenhouse dataset (UJNGH) and conducts comprehensive experiments. Results demonstrate that YOLO-GH achieves 97.2% average accuracy, surpassing existing advanced detection methods.