A New Approach to Predict Radio Map via Learning-Based Spatial Loss Field

Accurately predicting radio maps is essential for various applications. Learning-based methods have recently gained widespread attention for precision and speed in radio map prediction. However, many existing methods in this field require a substantial amount of measurement data for training, hindering practical applications due to the associated costs. To overcome the challenge of limited training data, this paper explores the use of the spatial loss field to extract radio propagation patterns, aiming to enhance prediction accuracy and reduce the required data volume. Specifically, we propose regression clustering to address interpolation within the same region and combine deep learning to predict radio maps across different regions. Verification results on the publicly available dataset demonstrate the superiority of our approach in scenarios with limited data.