In this paper, air quality inference aims to estimate pollution levels at unmonitored locations on the basis of existing observations. In the context of multisource heterogeneous data, this task faces two critical challenges: effectively integrating diverse data sources with monitoring data and accurately modelling complex spatial relationships among monitoring stations and between stations and target locations. To address these challenges, we propose a novel air quality inference model, FSMoE-LSC, which integrates multiple key mechanisms. Specifically, a feature attention network is designed to fuse heterogeneous data sources with monitoring information; a self-attention-based mixture-of-experts (SMoE) module is introduced to capture intricate spatial dependencies; and a learnable skip connection vector is incorporated to enhance global information and target point information extraction. Experiments on real-world air quality data from the central urban area of Chongqing demonstrate that FSMoE-LSC significantly outperforms existing state-of-the-art baselines in terms of the RMSE and MAE, verifying its superior inference accuracy and generalizability. The source code and datasets used in this paper have been made publicly available at https://github.com/zhongcyi/ChongQing_AQI .