A lightweight hybrid ViT-GNN framework for data-centric land cover mapping in the amazon biome using graph structural priors

Continuous monitoring of the Amazon biome demands land cover classification models that are both highly sensitive and computationally feasible. To resolve the inherent trade-off between architectural complexity and predictive performance in spatial deep learning, this study introduces the Vision Transformer-Graph Neural Network with Feature Adaptation (ViT-GNN RFFA). In contrast to conventional end-to-end pixel models, this hybrid architecture operates exclusively on an 11-dimensional vector of extracted color-based vegetation indices (e.g., GRVI) and textural statistics. The dual-branch design isolates global sequence context via the ViT module while leveraging the GNN branch as a structural prior to learn non-linear covariance between specific features. Evaluated against a suite of benchmarks including MiniViT, Baseline CNN, Random Forest, XGBoost, and LightGBM, the proposed algorithm attained the highest Overall Accuracy of 0.930 utilizing merely 16,323 trainable parameters-a nearly 75% reduction in footprint versus pixel-based models. Crucially, a McNemar's statistical test confirmed that the accuracy gain over the strongest classical baseline (XGBoost, OA: 0.927) is statistically significant (p < 0.05). By pairing rigorous spatial cross-validation with an interpretable feature space, this work establishes that intelligent data pre-processing combined with graph-based relational learning offers a robust framework for high-precision environmental mapping under severe resource limitations.