A Deep Ensemble Semantic Segmentation Framework for Efficient Spectrum Sensing in Cognitive Radio Networks

Received: 28 December 2025 | Revised: 11 February 2026 | Accepted: 14 February 2026 | Online: 4 April 2026

Corresponding author: Sk. Shariful Alam

Abstract

With 5G networks transitioning to congested frequency bands, distinguishing Long Term Evolution (LTE) and New Radio (NR) signals from noise becomes critical for effective interference management. However, simultaneously locating local signal boundaries and establishing a global context can be difficult when traditional semantic segmentation techniques are utilized. To address this issue, we propose a robust ensemble deep learning framework that integrates two distinct architectures. The first is DeepLabV3+ with Atrous Spatial Pyramid Pooling (ASPP) to acquire multi-scale contextual features. The second is a U-Net architecture integrated with the spatial and channel Squeeze-and-Excitation (scSE) attention enhancement mechanism. The integration of the scSE attention enhancement mechanism with the U-Net model helps adjust the feature maps when needed, reducing insignificant areas and revealing valuable spectral features. In order to ensure model generalizability, the training process adopted large-scale data augmentation and used median frequency balancing to correct the effect of class imbalance. Evaluations demonstrate that our dual-stream ensemble approach is much more effective than baseline models and previous studies. The proposed ensemble framework achieved an excellent mean Intersection over Union (mIoU) score of 0.9878. The average mIoUs of common architectures such as PixelMLP, SimpleFCN and SimpleSegNet were 0.2962, 0.4998, and 0.5931, respectively. These findings demonstrate that combining attention-based feature refinement with global context extraction is a highly dependable and precise technique for automating 5G spectrum detection.