Deep learning-based wireless channel prediction with propagation feature exploitation
Deep learning based wireless channel prediction can benefit from full exploitation of propagation features, yet how to construct effective inputs for accurate channel prediction remains under-explored. This paper presents a deep learning-based approach that optimizes environmental input construction for accurate channel path loss prediction. A campus measurement campaign is conducted to obtain datasets, where satellite image and semantic layers are aligned at a unified environmental granularity. Using a fixed network backbone and identical training protocol, results show that channel prediction accuracy is sensitive to spatial coverage. Overly local patches are insufficient, while enlarging coverage to include propagation-relevant surroundings yields error reduction with diminishing returns. It is further found that orientation normalization, implemented by aligning the Tx-Rx direction to a fixed reference axis, improves performance by reducing geometric variability. Building on this spatial extent, Shapley-based attribution and ablation analysis indicate that buildings and roads have dominating impacts, vegetation offers complementary information, and location-related descriptors mainly contribute through interactions. The results validate that the input feature construction improves channel prediction accuracy and provides guidance for future intelligent channel prediction.