MROD-YOLO: Multimodal Joint Representation for Small Object Detection in Remote Sensing Imagery via Multiscale Iterative Aggregation

Accurate and efficient small object detection using multimodal remote sensing images on resource-constrained aerial platforms is a challenging task. Most existing solutions rely on complex networks to extract object features, which often incurs significant computational overhead. In this article, we propose MROD-YOLO, a novel multimodal object detection framework designed to enhance the accuracy and efficiency of remote sensing imagery analysis. First, we design a multimodal joint representation network (MJRNet), which employs a global context attention block (GCB) to maximize feature retention during fusion. MJRNet integrates complementary features from multimodal images, addressing the limitations of single-modal methods in complex scenarios. Furthermore, we introduced the receptive field expansion mechanism (RFEM) to the spatial feature representation of backbone networks. Finally, a multiscale iterative aggregation (MSIA) module is developed to refine feature interactions and improve the detection of small objects in challenging environments. By replacing the path aggregation network (PANet) with a feature pyramid network (FPN), the model ensures effective feature preservation for texture details. Experimental results show that MROD-YOLO achieves an accuracy of 77.9% mAP50, which is 4.2% higher than the SOTA large models CFT. MROD-YOLO provides a promising solution for real-time, high-precision object detection in multimodal remote sensing applications.