A Self-Consistent Approach to Hierarchical Remote Sensing Image Classification

While deep learning architectures have become the standard for remote sensing image classification, most existing frameworks adopt a flat classification paradigm. These approaches often overlook the inherent semantic dependencies within predefined label hierarchies, focusing exclusively on fine-grained labels and failing to exploit the topological structure of the taxonomy. This paper proposes the Semantics-Aware Hierarchical Consensus (SAHC) method, a novel framework designed to explicitly model hierarchical relationships by integrating granularity-specific classification heads into a unified architecture. Unlike rigid rule-based systems, the proposed approach utilizes trainable hierarchy matrices that parameterize the transitions between taxonomic levels, enabling the network to learn the structural mapping between coarse and fine categories in a self-consistent, data-driven manner. Furthermore, we introduce a hierarchical consensus mechanism to align probability distributions across various levels of the taxonomy. This mechanism acts as a learned weighted ensemble that regularizes the feature space by leveraging inherent class dependencies, effectively reducing semantic ambiguity in complex scenes. The SAHC framework is evaluated on a benchmark RS dataset characterized by a complex taxonomy using diverse convolutional and transformer-based backbones. Experimental results demonstrate that SAHC significantly improves classification robustness and feature discriminability, outperforming traditional flat-labeling approaches across multiple evaluation metrics.