A Location-Aware Hybrid Deep Learning Framework for Dynamic Near-Far Field Channel Estimation in Low-Altitude UAV Communications

TL;DR

This work tackles the challenge of estimating dynamic A2G channels in low-altitude UAV communications where near-field and far-field propagation co-exist due to mobility and large antenna arrays. It introduces a location-aware hybrid deep learning framework that fuses CNN-based spatial feature extraction, multi-head self-attention for global dependencies, BiLSTM for temporal evolution, and real-time transmitter/receiver position priors to learn joint spatio-temporal channel representations across near- and far-field regimes. The proposed RACNN-BiLSTM architecture outperforms several baselines, achieving at least a 30.25% NMSE reduction at 15 dB and showing robustness to pilot length, antenna count, and mobility. By integrating geometric priors with advanced neural components, the approach provides a practical and scalable solution for accurate CSI in dynamic UAV relay systems, with potential impact on 6G XL-MIMO UAV networks.

Abstract

In low altitude UAV communications, accurate channel estimation remains challenging due to the dynamic nature of air to ground links, exacerbated by high node mobility and the use of large scale antenna arrays, which introduce hybrid near and far field propagation conditions. While conventional estimation methods rely on far field assumptions, they fail to capture the intricate channel variations in near-field scenarios and overlook valuable geometric priors such as real-time transceiver positions. To overcome these limitations, this paper introduces a unified channel estimation framework based on a location aware hybrid deep learning architecture. The proposed model synergistically combines convolutional neural networks (CNNs) for spatial feature extraction, bidirectional long short term memory (BiLSTM) networks for modeling temporal evolution, and a multihead self attention mechanism to enhance focus on discriminative channel components. Furthermore, real-time transmitter and receiver locations are embedded as geometric priors, improving sensitivity to distance under near field spherical wavefronts and boosting model generalization. Extensive simulations validate the effectiveness of the proposed approach, showing that it outperforms existing benchmarks by a significant margin, achieving at least a 30.25% reduction in normalized mean square error (NMSE) on average.

Figures (9)

• Figure 1: Network model
• Figure 2: The framework of the proposed channel estimation algorithm
• Figure 3: RACNN-BiLSTM training vs. testing Loss
• Figure 4: NMSE vs. SNR for far-field and near-field.
• Figure 5: near-field: NMSE vs. pilot lengths