Hybrid wireless sensor networks (HWSNs) combine sensors of varying costs to balance budget and deployment density. However, their data products often exhibit high heterogeneity and noise, presenting new challenges for spatial interpolation models. Traditional spatial interpolation models take dense input. When working on HWSN datasets, a large part of the dense input must be obtained through imputation, leading to feature distribution changes and error accumulation. To address these challenges, we propose the Context Encoder Spatial Interpolation (CESI) Model, designed to work directly with sparse, narrow-format input. CESI integrates a GraphSAGE-based backbone with a Transformer-based context embedding module, leveraging probabilistic encoding for better generalization to unseen coordinates and a self-supervised signal to balance inductive biases between the two modules. Experimental results demonstrate that CESI consistently outperforms baseline models across several publicly available real-world datasets.