The Center for Data Innovation recently spoke with co-founder Fernando Yu. suyanais a Massachusetts-based company that uses satellite imagery and machine learning models to provide climate risk insurance. Yu explained how Suyana's models identify climate patterns, assess the likelihood and severity of damaging events, and incorporate the results into insurance products appropriate for specific regions.
David Cartai: What challenges is Suyana solving in the climate-driven insurance industry?
Fernando Yu: Traditionally, when farmers and businesses seek climate insurance for risks such as droughts and floods, they work with insurance companies that rely on loss adjusters, field inspectors, and reporting agencies to assess disaster risk. These teams manually visit sites, verify past claims, and rely on sparse weather station data. Gathering this information is time-consuming, costly, and often has significant gaps, so insurance companies use broad regional averages to set insurance prices rather than the actual conditions of a particular property. As a result, claims are often processed more slowly, costs increase, and risks are misrepresented.
Suyana takes a different approach by combining high-resolution satellite imagery from providers such as Planet Labs with ground-based data such as soil moisture, precipitation, sea level measurements, and output from global climate models. Our machine learning model combines these inputs to produce a hyper-specific risk assessment with 1 km × 1 km resolution. This is about a 400x improvement over the 20 km x 20 km grids common today. This level of detail allows us to price climate risk at the scale of individual parcels rather than entire regions, resulting in more accurate coverage.
Keltai: How do models translate climate data into risk predictions and insurance triggers?
excellent: We use machine learning models to predict the likelihood and severity of specific climate hazards in specific locations, such as droughts, floods, and coastal storm surges. At each location on the map, the model estimates how often a damaging situation is likely to occur and how severe the situation is.
To do this, we first group locations that behave similarly from a climate perspective, based on factors such as rainfall patterns, soil properties, temperature fluctuations, and coastal exposure. This step, known as clustering, helps identify areas that tend to experience similar stress. We then build parametric models that track specific environmental metrics, such as soil moisture and wave height, within each group and estimate how often those metrics exceed damage-related thresholds. These forecasts ultimately determine insurance prices and payments for individual properties.
Keltai: How do you build location-specific climate risk models when you have a location with little historical climate data?
excellent: In climate insurance, we refer to these data gaps as missing or unreliable regional measurements such as rainfall, soil moisture, and storm exposure, especially in areas with less dense networks of weather stations. These gaps make it difficult to accurately estimate risk at the farm or community level.
We fill these data gaps through satellite coverage, transfer learning, and local validation. Global satellite archives provide us with a consistent, decades-long record of soil moisture, vegetation health, and precipitation, even in areas without weather stations.
It also uses transfer learning. This allows a model trained in one domain to apply learned patterns to similar domains. For example, identify comparable agro-climatic zones and apply validated models from places like Brazil to regions in Bolivia and Paraguay, adjusting to local conditions.
Finally, we are investing heavily in ground truth verification. We work with development organizations to compare satellite-derived metrics to real-world conditions. For example, during the Bolivian drought, we worked with farmers and agronomists to ensure that our indicators matched what was observed on the ground. This human-in-the-loop approach enhances model accuracy and ensures that the right insurance products are designed for users.
Keltai: How do you maintain model reliability in the face of climate change?
excellent: As climate change alters historical weather patterns, models must continually evolve. Rather than relying solely on historical data, we place greater emphasis on recent satellite and ground-based observations to capture emerging trends.
We also integrate forward-looking results from global climate models and incorporate the latest climate science research into our updates. This will help assess increasing risks, such as prolonged drought due to rising storm surge levels and changing rainfall patterns. We combine these updates with a conservative pricing strategy to account for uncertainty and prioritize long-term resilience.
Keltai: Can you tell us how Suyana is starting to roll out its products into the real world market?
excellent: In Bolivia, we are working with four of the country's five largest agricultural banks and the country's largest grain wholesaler to provide drought insurance covering 400,000 hectares. The 2023-2024 growing season saw Bolivia experience its worst drought in 30 years, devastating uninsured farmers and causing massive loan defaults. Our built-in model directly bundles insurance and farm credit, so any loan a farmer takes out for seed, fertilizer, or equipment is automatically covered. This eliminates the need to find another provider or go through complex paperwork, which are often barriers for smallholders to access protection.
Peru is piloting storm surge insurance for fishermen who currently rely on slow and inconsistent government subsidies when extreme weather closes ports. When ports close, fishermen lose their only source of income, and subsidies often take weeks or months to arrive. Our parametric products track wave height and automatically trigger payments when conditions force a closure, providing speed and reliability that subsidy systems cannot provide.
