The Florida Keys National Marine Sanctuary (FKNMS) is monitored by an extensive but fragmented constellation of remote sensing and in-situ observation platforms: Sentinel-2 and Landsat multispectral satellite imagery, NOAA Coral Reef Watch satellite-derived thermal products, autonomous underwater vehicle photography, passive acoustic hydrophone arrays, vessel-tracking AIS transponders, and multi-agency water-quality sensor networks—collectively covering 2,900 square nautical miles of reef, seagrass, and mangrove habitat supporting more than 6,000 species. These data streams operate in silos, producing isolated assessments that cannot support the integrated, time-sensitive management decisions the sanctuary requires. Meanwhile, the sanctuary’s static zone boundaries, designed from 1990s-era reef assessments, now protect degraded substrate in some areas while leaving climate-resilient coral assemblages unprotected in others. We propose an energy-aware, tiered AI architecture that fuses these multi-source remote sensing and in-situ data streams into a digital twin of the FKNMS ecosystem. The architecture assigns remote sensing analytics tasks across three computational tiers—classical machine learning for structured sensor and satellite-derived indices, deep learning for unstructured imagery and acoustic spectrograms, and foundation models for cross-modal reasoning and multi-agency report synthesis—reserving the most energy-intensive techniques for tasks where they are irreplaceable. A reinforcement learning agent operates on the fused remote sensing state representation to recommend real-time adjustments to sanctuary zone boundaries, optimizing for coral recovery, sustainable fish stocks, and biodiversity under socioeconomic constraints. We ground each architectural component in existing, publicly available remote sensing infrastructure and demonstrate the framework through three prescriptive scenarios: satellite-triggered bleaching early warning with adaptive closures, remote-sensing-informed lionfish invasion management, and climate-adaptive rezoning driven by multi-temporal imagery analysis. The architecture directly addresses the energy-conservation paradox—the risk that computationally expensive remote sensing analytics intended to protect the environment may themselves cause environmental harm—by formalizing energy cost as a first-class constraint in model selection. This work complements a companion paper applying the same tiered architecture to the Greater Yellowstone Ecosystem, demonstrating extensibility across terrestrial and marine remote sensing domains.