Poverty detection relies on census records, household surveys, per capita income reports, and on-site community assessments. These traditional methods are costly in terms of data and resources as they require on-the-ground surveys. In contrast, satellite imagery, when combined with advanced land use and land cover classification, offers a scalable and cost-effective alternative for poverty detection. Although many deep learning methods have been explored for poverty detection, their practical application is limited due to the lack of interpretation. Moreover, the lack of explainability in existing models requires human intervention. This study uses satellite image to identify key features such as airports, roads, buildings, and bridges to detect poverty. To improve the interpretation of the model’s predictions, we apply explainable AI methods, namely LIME, SHAP, and Grad-CAM. In terms of model performance, EfficientNetB2 outperformed all other evaluated models, achieving the highest validation accuracy of 97.62%, along with a precision of 97.89% and recall of 97.29%. These results, when combined with explainable AI techniques such as Grad-CAM, LIME, and SHAP significantly enhance the interpretability of predictions. This capability is especially valuable for supporting government agencies and civic organizations in making informed decisions for poverty detection and resource allocation.