Sensor-specialized JEPA foundation models complement planetary-scale embeddings for agentic hydrologic reasoning

Geospatial foundation models compress multispectral observations into dense embeddings that are increasingly applied in natural-language environmental reasoning systems. A single planetary-scale model, e.g. Google AlphaEarth, handles broad characterization well but may compromise on specialized hydrologic signals. Such generalist models are also often inaccessible, expensive, and require large-scale compute. We propose a fleet of small sensor-specialized Joint Embedding Predictive Architecture (JEPA) foundation models that can be consulted by a routing agent for specialized questions. We pretrained five (22M parameter) Mini-JEPAs , that share an identical Vision Transformer backbone, an identical JEPA training recipe, and a 64-dimensional output space, using images from Sentinel-2 optical, Sentinel-1,SAR, MODIS thermal, multi-temporal Sentinel-2 phenology, and a topography and soil stack. Each Mini-JEPA reconstructs the environmental variable matched to its sensor, with cross-validated R 2 reaching 0.97 for elevation under the topography-soil model, 0.97 for mean temperature under the thermal model, and 0.81 for precipitation under the phenology model. Across all five models, the variable each one predicts best is the variable its sensor physically observes. The five manifolds differ in geometric structure, with global participation ratios from 8.9 (SAR) to 20.2 (thermal) and local intrinsic dimensionalities from 2.3 to 9.0; the spread is driven mainly by the thermal model, while the two Sentinel-2 optical models are nearly identical. Combining the topography-soil and phenology models with AlphaEarth Foundation Model adds small but consistent predictive value for soil moisture, aridity, and precipitation ( Δ R 2 up to 0.031). On the variables their sensors directly observe, the specialists also hold their skill across unseen regions where AlphaEarth alone degrades. A router LLM (Claude Sonnet/Opus) reads the per-modality reference cards and picks the sensors that fit each question, choosing at least one expected sensor on every question. In paired LLM-as-Judge evaluation over 90 hydrologic questions, adding the routed fleet to AlphaEarth helps most on questions that defeat optical imagery, such as cloud cover and water under vegetation (Cohen’s d = 0.46 , p = 0.041 ). Our findings suggest that small specialized Mini-JEPAs can be built with modest compute, complement planetary-scale models, and match them on the variables their sensors observe.