<p>Data-driven energy benchmarking has become a key strategy for promoting energy conservation in the building sector. However, its scalability is constrained by the limited availability of high-quality building data. Data-driven approaches employing deep learning have demonstrated the ability to help address data scarcity using remote sensing data but lack interpretability. This study introduces a novel deep learning framework that combines satellite imagery at multiple zoom levels with tabular building and weather data to estimate monthly site energy use intensity (EUI). Unlike prior approaches, this method explores how contextual resolution impacts estimation accuracy and incorporates explainability techniques to address deep learning model interpretability to identify which attributes in a building’s context contribute most to energy usage. The approach is applied to buildings in Washington, D.C. and evaluated against widely used energy benchmarking models. The results show the proposed dual-scale EB-CNN, which fuses satellite image data at zoom levels 16 and 18 with tabular building and weather data, outperforms conventional approaches and achieves an R² of 93.55% and an MAE of 0.87 kBtu/sqft/month. Notably, even when relying solely on satellite and weather data, the model performs comparably to the baseline models, effectively mitigating data scarcity constraints. A novel saliency analysis process is also used to analyze the highest performing model, demonstrating that it leverages scale-specific visual cues across satellite zoom levels to improve accuracy. By addressing both data limitations and model interpretability, the research offers a scalable and explainable pathway to support building energy conservation and carbon reduction across diverse urban environments.</p>