top of page

Foundation Models 101: Why we need Marine Geospatial Foundation Models

  • Writer: Vincent Opitz
    Vincent Opitz
  • Jun 19
  • 6 min read

Updated: Jul 1


Artificial intelligence has traditionally been built around specialists - models trained to excel at one specific task. You might have a model that recognises faces, another that translates languages, and a third that recommends films. Each requires its own dataset, training process, and expertise. This approach has served us well, but it's fundamentally limited by its narrow focus.


Enter foundation models - a revolutionary approach that's changing how we think about AI entirely. Rather than training models for specific tasks, foundation models are designed to understand the world more broadly, learning general patterns from massive datasets without being told exactly what to predict. Once trained, they can be quickly adapted to tackle multiple different challenges with remarkable effectiveness.


Traditional AI requires separate models for each task, whilst foundation models provide unified intelligence that can power multiple applications. Source: Armand Ruiz



The Foundation Model Revolution


Think of foundation models like a brilliant research assistant who has read every book in the library. They haven't been told to become an expert in any particular subject, but through this comprehensive reading, they've developed a deep understanding of how knowledge connects across disciplines. When you later ask them to write about history, science, or literature, they can draw upon this vast understanding to provide insights that a specialist trained only in one field might miss.


Consider how this has transformed language technology. Rather than training separate models to translate text, write summaries, answer questions, and generate code, we now have foundation models like GPT that understand language so fundamentally they can tackle all these tasks and more. The same revolution is happening across other domains: computer vision models that can classify images, detect objects, and generate art; protein folding models that understand molecular structure well enough to predict drug interactions and design new medicines.


The common thread is profound: these models don't just memorise patterns - they develop genuine understanding of their domains that transfers across applications.

At Ocean OS, we've been asking a crucial question: could this same approach revolutionise how we understand and monitor our oceans?


Existing Geospatial Foundation Models


Most AI systems designed for Earth observation focus on land-based applications - tracking deforestation, monitoring urban development, or predicting crop yields. These terrestrial features tend to be relatively stable over time. A forest remains a forest for decades, and cities don't fundamentally restructure themselves overnight.


Marine environments operate by entirely different rules. Ocean eddies form and dissipate over weeks. Phytoplankton blooms can appear and vanish in days. Temperature gradients shift with currents that span thousands of kilometres. The ocean is a fluid, three-dimensional system where everything is connected across vast spatial and temporal scales.


Visual summary of the breakdown of phases to solve geospatial problems using a geospatial foundation model. Source: Google


Why the Ocean Needs Foundation Models


The ocean represents one of the most complex systems on our planet, yet remains one of the least understood. Every drop of seawater contains a story - temperature, salinity, nutrients, microscopic life - that connects to currents thousands of kilometres away, weather patterns forming overhead, and ecosystems thriving on the seafloor.


Marine science has traditionally relied on two main approaches. Physics-based numerical models attempt to simulate ocean behaviour using mathematical equations that describe fluid dynamics, thermodynamics, and biogeochemical processes. These models, whilst scientifically rigorous, require enormous computational resources and often struggle with the ocean's inherent complexity and chaotic behaviour. They're excellent at capturing broad patterns but can miss local variations and struggle with phenomena that don't fit neatly into their mathematical frameworks.


The alternative has been narrow machine learning models that focus on specific aspects of the ocean: one for predicting storms, another for tracking fish populations, a third for understanding currents. Each requires its own expensive data collection, specialised expertise, and focused training. When we need to understand how these systems interact - which is almost always - we're left trying to piece together insights from separate models that don't communicate with each other.


Foundation models offer a fundamentally different approach that combines the best of both worlds. Like physics-based models, they can capture complex system interactions, but unlike traditional numerical models, they learn these relationships directly from data rather than requiring us to specify every equation. They're faster to run, more adaptable to local conditions, and better at handling the messy, incomplete data that characterises real-world ocean observations. Most importantly, they can identify patterns and relationships that we might not have thought to include in our mathematical models.


From isolated insights and lengthy processing to comprehensive ocean understanding - foundation models can transform how we study marine systems. Source: Ocean OS



Transforming Critical Marine Predictions


The practical impact of foundation models in marine science is already becoming apparent across some of the most challenging prediction problems we face. Because these models understand the ocean as an integrated system, they can provide insights that traditional approaches - whether physics-based or narrow machine learning - simply cannot match.


Marine Heatwave Forecasting: Marine heatwaves can devastate coral reefs, disrupt fisheries, and alter entire ecosystems. Traditional physics-based models struggle to predict these events because they emerge from complex interactions between atmospheric conditions, ocean currents, and local geography. Foundation models excel at identifying the subtle precursor patterns across multiple variables that signal an approaching heatwave, often providing weeks or months of advance warning where traditional models might only manage days.


Storm Prediction and Intensity Forecasting: Whilst we've made significant progress in predicting storm paths, understanding storm intensity remains challenging. Foundation models can integrate sea surface temperatures, wind patterns, atmospheric pressure, and even biological indicators to provide more accurate predictions of storm development and intensification. They're particularly valuable for understanding how storms might behave in changing climate conditions where historical patterns may no longer apply.


Harmful Algal Bloom Prediction: Predicting toxic algal blooms requires understanding complex interactions between nutrient levels, temperature, salinity, light conditions, and existing marine life. Physics-based models struggle with the biological complexity, whilst traditional machine learning approaches often miss the interconnected environmental factors. Foundation models can identify the multi-variable signatures that precede bloom events, enabling earlier warnings that protect both marine ecosystems and human health.


Species Distribution Forecasting: Understanding where marine species will be found—and how their distributions might shift with changing conditions—requires integrating oceanographic, biological, and climatic data. Foundation models can predict not just where species are likely to be today, but how their habitats might change as ocean conditions evolve. This capability is transforming conservation efforts and fisheries management, allowing for proactive rather than reactive approaches to marine stewardship.


The Perfect Storm of Technological Advancement


Several technological developments have converged to make marine foundation models practical. The explosion in ocean data from satellites, autonomous sensors, and monitoring networks provides the comprehensive datasets these models need. Advances in computing power, particularly graphics processing units originally designed for gaming, now enable us to train models on these massive datasets.


Perhaps most importantly, breakthrough algorithms called transformers have made it efficient to learn from complex, interconnected data - perfect for understanding ocean systems where everything affects everything else. These algorithms excel at identifying patterns across time and space, crucial for understanding phenomena like how El Niño events influence global ocean circulation.


The timing is particularly crucial for marine science. As climate change accelerates and human impacts on ocean systems intensify, we need tools that can understand and predict ocean behaviour with unprecedented accuracy and speed. Foundation models provide exactly this capability.


Beyond Technology: Democratising Ocean Understanding


Foundation models represent more than just technological advancement - they're democratising access to sophisticated ocean analysis. Rather than requiring separate teams to develop models for each specific application, organisations can adapt a single foundation model to multiple uses, dramatically reducing development time and costs.

This efficiency is particularly important for conservation organisations, government agencies, and research institutions working with limited resources. A marine protected area manager can use the same underlying model for species monitoring, pollution detection, and climate impact assessment. A coastal community can access early warning systems for storms and better fisheries management tools, all powered by the same foundational understanding.


The technology is also accelerating scientific discovery by identifying patterns that humans might miss across vast datasets. Foundation models can process years of satellite data, sensor readings, and environmental measurements simultaneously, revealing insights about ocean system interactions that would take human researchers decades to uncover.


Ocean OS: Building the Future of Marine Intelligence


At Ocean OS, we're at the forefront of developing a foundation model specifically designed for marine environments. Our vision extends beyond creating another AI application - we're building the foundational intelligence that can adapt to the diverse challenges facing our oceans.


Whether you're tracking marine heatwaves, planning sustainable aquaculture, studying ecosystem changes, or managing coastal resources, our foundation model aims to provide the deep ocean understanding needed to make better decisions faster. We recognise that the ocean doesn't operate in silos - currents affect temperature, temperature affects marine life, marine life affects nutrient cycles. Our models reflect this interconnected reality.


Foundation models aren't just the future of marine science - they're transforming how we understand and protect our oceans today. As we face increasingly complex challenges from climate change to biodiversity loss, we need tools that can comprehend the ocean as the complex, interconnected system it truly is. That's exactly what foundation models deliver, and it's why we're so excited about their potential to revolutionise marine science and ocean stewardship.


bottom of page