The graph structures generated by natural or simple heuristics often fail to represent the spatial correlations influenced by complex factors. Therefore, introducing graph structure learning (GSL) can enhance the graph neural network-based spatial interpolation models. However, the input features of the GSL module are systematically unbalanced in spatial interpolation tasks. For example, many natural variables follow Gaussian- or gamma-distribution, and sensor spatial distributions are generally uneven. Thus, the GSL module must systematically incorporate corresponding solutions to avoid negatively impacting its generalization ability and degrading model performance. Our proposed model utilizes two encoders to embed feature deviations of node readings and centroid distance from preset distributions, respectively. Notably, these encoders are jointly optimized with other model components, and their generalization ability is improved through an adaptively adjustable information bottleneck. Consequently, the GSL module can offer a more robust graph structure by explicitly perceiving feature deviations in the input. Experimental results demonstrate that our model outperforms existing state-of-the-art baselines across multiple real-world datasets with diverse characteristics.