<p>The increasing volume and heterogeneous nature of remote sensing data necessitate efficient and accurate image retrieval systems, particularly for land use and land cover (LULC) mapping. This paper presents an advanced framework that integrates an optimized label propagation network (OLPNet) with a two-stage hybrid hierarchical classification (TS-H₂C) approach for efficient remote sensing image retrieval (RSIR). The TS-H₂C-OLPNet framework employs two sparse kernel learning machines- the relevance vector machine (RVM) and support vector machine (SVM), to enhance label distribution in complex, high-dimensional feature spaces. The framework utilizes a reconstruction-based relational autoencoder (RAE) to extract robust deep features with reduced dimensionality. In the Stage-1 hierarchy, the RVM generates confidence scores to determine super-class labels, from there the OLPNet propagates respective sub-class labels to the Stage-2 SVM. This approach efficiently manages the search space, speeds up retrieval, and effectively handles highly overlapping LULC classes, improving the recognition of unseen categories. Extensive experiments on benchmark RSI datasets with challenging scene categories demonstrate that the TS-H₂C-OLPNet framework achieve SOTA Precision, Recall, and F1-score performance. Results ensure that integrating optimized label propagation and hybrid hierarchical classification can offer a robust solution for large-scale retrieval using complex LULC data.</p>