Timely and accurate monitoring of opioid overdose risks is critical for public health, particularly in underserved rural regions. Traditional surveillance systems often lack the spatial and temporal resolution needed to support proactive interventions. While socioeconomic indicators, such as the social vulnerability index and housing value, show moderate correlations with opioid-related outcomes, existing methods rarely incorporate high-resolution environmental data. Previous research relies largely on static census data and coarse geographic indicators, limiting its ability to detect localized risk patterns. Moreover, the connection between built-environment features and opioid overdose remains underexplored—especially in rural areas like Alabama’s Black Belt. To address this gap, we propose a multiscale spatio-temporal framework that integrates satellite imagery and machine learning to monitor opioid-related emergency room (ER) visit rates. We collected 201,967 housing images from Black Belt counties and classified them using computer vision models, including ResNet and external attention transformers. To overcome limitations in labeled data, we developed four unsupervised pipelines combining k-means clustering with autoencoders, masked autoencoders, VGG16, and household-image ratios. Our results show that unsupervised embeddings outperform supervised classification in capturing signals associated with ER visits. Descriptive features, such as roof type, road layout, and environmental openness, significantly inform predictions. Although Black Belt counties report lower absolute ER visit rates, they show faster year-over-year growth. Our study demonstrates the potential of combining satellite imagery with multimodal artificial intelligence to improve rural health surveillance and supports the development of scalable, interpretable monitoring tools for early intervention and policy planning.