Integrating eco-biogeographical knowledge into deep learning: An approach for fine-grained mountain vegetation classification

Fine-grained mountain vegetation classification remains a critical challenge in remote sensing due to spectral similarities among vegetation canopies and the complexity of mountainous ecosystems. Although deep learning (DL) has demonstrated superior performance in generic land cover classification, its application to vegetation formation classification suffers from three inherent limitations: (1) poor generalizability across heterogeneous terrains, (2) constrained accuracy when handling ecologically similar taxa, and (3) lack of interpretability due to the black-box nature. To address these challenges, this study develops a DL vegetation classification framework enhanced by elevation-dependent eco-biogeographical knowledge computation. The framework comprises three key steps: (1) Extracting computable knowledge rules from the Mountain Vegetation Knowledge Graph; (2) Designing an eco-biogeography-informed loss function to integrate these rules to train the DL model, thereby enhancing its interpretability and generalizability; (3) Performing post-classification refinement through eco-biogeographical knowledge-based similarity calculations, which further improves model accuracy. Experimental results with FCN8s-ResNet50 show significant improvements, achieving 77%–87% precision, 70%–88% recall, 74%–92% F1-score, and 82% overall accuracy for the training model. The framework achieves an additional 21.67% accuracy boost (from 69.84% of Exp.1 to 91.51% of Exp.4) for vegetation formation classification. This approach systematically enhances the accuracy, interpretability, and generalizability of vegetation classification models. Additionally, it provides an efficient framework for automated vegetation mapping and monitoring. • Novel eco-biogeography-informed loss function enhances DL model interpretability and generalizability. • Eco-biogeographical knowledge-based similarity computation further improves model accuracy and interpretability. • The eco-biogeographical DL model attains 91.51% accuracy with a 21.67% gain for formation classification. • Integrating eco-biogeographical knowledge into DL models offers a new paradigm for fine-grained vegetation mapping.