The accuracy of fingerprint-based indoor localization strongly relates to the precision of the wireless radio map. However, radio maps are vulnerable to deployment changes and require constant maintenance, which is labor-intensive and time-consuming. To address this issue, we propose an automated radio map update method that employs a mobile robot to collect fingerprints along its path. Considering the relationship between signal strength and propagation distance, we design a reconstruction network combining multi-scale features in radio map updating, which estimates signal strengths based on local and global distributions. In addition, we propose a dynamic strategy with an attention mechanism to select a set of crucial representative points (RPs) for a sparse survey in the update stage. The chosen points have stronger relations to the surroundings than other points, improving our method’s reconstruction accuracy. The experimental results show that the proposed method reconstructs the radio maps of our two test sites which outperforms the state-of-art approaches in reconstruction accuracy by 32.03%. And the localization accuracy of the updated radio maps is significantly improved by about 51.39%.