Sampling-Assisted Neural Network Radio Map Generation with Shortcut Channel and Selective Sampling

Mengfan Wu, Marco Skocaj, Mate Boban

Keywords: Wireless communications, radio maps, machine learning, computer vision

TL;DR: We propose a sample-assisted CNN with shortcut paths for indoor radio map prediction. Optimized sampling strategies boost accuracy, especially at high sampling rates. Developed for MLSP 2025, our method enables fast, refined predictions.

Abstract: In this paper, we present a machine learning approach for fast indoor radio map generation assisted by sample measurements in the target environment as extra input. Our solution is developed for MLSP 2025 The Sampling-Assisted Pathloss Radio Map Prediction Data Competition. In addition to feature engineering for input augmentation, we design a shortcut path in the convolutional neural network that routes the sample input channel directly to the deeper layers, which facilitates efficient refinement of the output radio map. We further propose selective sampling strategies for measurement locations to enhance the accuracy of the generated radio maps. The proposed method demonstrates particularly strong performance under conditions of relatively high sampling rates.

Submission Number: 4