LKA-GFNet: Language Knowledge-Augmented Graph Fusion for Tri-Source Heterogeneous Remote Sensing Data Classification

Remote sensing (RS) image classification is pivotal for fine-grained land cover mapping and environmental monitoring. Recently, multi-source data fusion has emerged as a mainstream approach for enhancing classification accuracy. However, current fusion-based classification methods remain largely limited to dual-source integration, leaving the richer complementary information inherent in tri-source data—such as hyperspectral (HS), multispectral (MS), and radar data—underexplored. Furthermore, existing methods rely almost exclusively on image features and lack systematic incorporation of the textual semantic priors derived from geoscience classification standards. This oversight hinders the model’s capacity to capture the deeper associations between image features and land cover concepts, thereby considerably limiting classification performance. To address these gaps, this paper proposes a language knowledge-augmented graph fusion network (LKA-GFNet) for the fusion classification of HS, MS, and radar data. By incorporating textual semantic priors of geoscience labels into tri-source feature learning and alignment, LKA-GFNet enables semantics-driven high-accuracy classification. Specifically, a strong-weak dual-stage alignment framework is introduced to enable progressive image semantic perception, leveraging unified textual priors as shared references to collaboratively optimize both geoscientific knowledge and tri-source heterogeneous image features. Concurrently, we design a compound threshold-constrained graph-driven decision fusion module for interactive fusion of tri-source heterogeneous features. It constructs intra- and cross-modal graphs via cumulative probability and adaptive distance thresholds, mining latent complex dependencies and uncertainties within the trisource data. Extensive experiments on three tri-source datasets demonstrate that LKA-GFNet significantly outperforms state-of- the-art methods, achieving robust land cover mapping with enhanced interpretability. The source code will be uploaded to https://github.com/cwlnnu/LKA-GFNet.