Physics-Constrained Symbolic Regression from Imagery

We propose *SymbolicVision*, a physics-constrained symbolic regression framework that derives interpretable mathematical expressions directly from multi-spectral remote sensing imagery. Unlike black-box deep models, *SymbolicVision* combines a Vision-based image encoder with a Transformer-based symbolic decoder to enable cross-modal learning between visual features and symbolic formulas. A hybrid loss design ensures both numerical accuracy and physical plausibility. Evaluated on symbolic benchmarks (SRBench) and real satellite datasets (Open-Canopy), *SymbolicVision* achieves high predictive accuracy (\$R^2>0.99$\), and robust performance on geospatial tasks. This work highlights the potential of interpretable, physics-aware models for scientific remote sensing.