ANYF: A hybrid deep learning model for radiation sources localization through radio maps automatic interpretation

Abstract

To address the insufficient accuracy challenge of traditional radiation sources localization methods in complex urban environments, this paper proposes a hybrid deep learning model for radiation sources localization through radio maps (RMs) automatic interpretation. Firstly, a multimodal dataset is generated for training hybrid deep learning networks and locating radiation sources, with RMs as key data to be interpreted. Secondly, an atrous spatial pyramid pooling (ASPP) and convolutional block attention module (CBAM) enhanced UNet (ACU-Net) network is designed as the UNet network combined with the ASPP module and the CBAM to realize the segmentation of the radiation source area. Finally, a You Only Look Once version 8 (YOLOv8) with focal modulation (FM) (YOLO-FM) network is built as the YOLOv8 improved by replacing the spatial pyramid pooling feature (SPPF) module with the FM to detect small objects and perform high-accuracy radiation sources localization. These two networks, ACU-Net and YOLO-FM, thus constitute the hybrid deep learning model named ACU-Net-YOLO-FM (ANYF). The simulation experiment results show that the mean localization error of the proposed model in the target area is 2.37 meters, and the mean error when directly transferred to another urban area without any fine-tuning is 3.54 meters. In addition, we conducted real-world measurement experiments, which resulted in a mean error of 6.75 meters, demonstrating that the proposed model provides a feasible solution for supervising radiation sources and detecting interference.