Res-UNet-GRU: A Data-Driven Near-Field Sensing Approach in ISAC-Based Low-Altitude Networks

The rise of the low-altitude network (LAN) has driven the demand for situational awareness in dynamic environments, where integrated sensing and communication (ISAC) serves as a key enabler. To meet these requirements, ISAC systems increasingly employ high-frequency carriers and large-scale antenna arrays, thereby making near-field propagation effects non-negligible. However, existing studies have primarily focused on static near-field targets, which limits their applicability in dynamic LANs. In response to this limitation, we investigate a near-field ISAC system in which a ground-based ISAC transceiver simultaneously communicates with multiple uncrewed aerial vehicles (UAVs) and exploits their signal echoes for joint sensing. Specifically, we aim to estimate the angle of arrival (AoA), distance, and velocity of each UAV to support mobility-aware operations. Given the substantial number of parameters to be estimated, conventional methods such as maximum likelihood estimation (MLE) or matched filtering incur prohibitive computational costs. To address this issue, we propose Res-UNet-GRU, a neural network that integrates a residual UNet (Res-UNet) with a gated recurrent unit (GRU) to facilitate the aggregate estimation of location and motion parameters. The proposed network can estimate the UAV state without requiring exhaustive search or complex optimization, making it well-suited for high-mobility LANs and providing a foundational perception layer for future artificial general intelligence (AGI)-driven autonomous systems. Simulation results demonstrate that the proposed Res-UNet-GRU consistently achieves the lowest mean squared error (MSE) in estimating AoA, distance, and velocity parameters across all signal-to-noise ratio (SNR) levels compared to benchmark deep learning (DL)-based estimators.