Dimensionality optimized machine learning retrieval of canopy chlorophyll, nitrogen, and phosphorus from google satellite embeddings

• Google embeddings retrieved CCC, CNC, and CPC at 10 m over maize parcels. • CIRBS excelled for CCC/CNC; VOPSS excelled for CPC (R² = 0.83 to 0.93). • Accuracy saturated at 17–33 bands or 5–14 PCs, enabling compact models. • Six recurrent channels (A57, A51, A00, A15, A48, A61) drove predictions. • Ridge and ElasticNet outperformed ensemble and kernel methods. Long-term studies show that canopy chlorophyll, nitrogen, and phosphorus differ in spectral signatures and retrieval stability. This work assessed the retrieval of these biochemical parameters from a global 64-dimensional Google satellite embedding at 10 m resolution over maize parcels at the Kellogg Biological Station in Michigan, USA. Strategies were employed to compress the embedding into compact, information-rich subsets. Eleven machine learning regressors were applied across four workflows, each implementing distinct approaches to control dimensionality and enhance interpretability under a unified cross-validation framework. Results showed that Consensus Importance-Ranked Band Selection (CIRBS) achieved the highest accuracy for canopy chlorophyll (CCC) and nitrogen (CNC), while Variance-Ordered Principal Component Subset Selection (VOPSS) performed best for canopy phosphorus (CPC). Baseline workflows using full-band or fixed PCA representations exhibited lower performance. Accuracy saturated rapidly in both workflows: 17–33 bands for CIRBS and 5–14 principal components for VOPSS captured nearly all predictive information. A compact set of embedding channels (A57, A51, A00, A15, A48, A61) ranked highest across models and targets, indicating that key pigment, structural, and phenological contrasts were preserved in the learned 64-feature space. Regularized linear models (Ridge and ElasticNet) consistently outperformed ensemble and kernel methods. The embedding pre-compressed key vegetation contrasts into an analysis-ready vector, enabling compact models (R² = 0.83 – 0.93, NRMSE = 0.08 – 0.11) that supported efficient tiled inference and treatment-level differentiation. This framework provides accurate, interpretable, and scalable solutions for field-scale biochemical mapping and precision agriculture decision-making.