Mesoscale Model Forecast Skill for Hourly Rainfall During Postfire Flood and Debris-Flow Events in California

When a wildfire sweeps through a steep, rugged landscape, it does more than consume trees and homes; it fundamentally alters the ground itself. The intense heat can bake the soil, making it repel water rather than absorb it, while the removal of vegetation leaves the earth exposed and loose. In the months and years following such a fire, these burned watersheds become dangerously sensitive to rain. A storm that would normally be harmless can suddenly trigger flash floods and debris flows—rushing torrents of water, mud, rocks, and trees that race down hillsides with devastating force. These events often happen quickly, driven by short bursts of heavy rain that last less than an hour. Because the danger is so immediate, emergency managers rely on weather forecasts to give people time to evacuate or close roads before the water arrives. However, predicting exactly where and when these intense, localized bursts of rain will occur is one of the most difficult challenges in modern meteorology.

A team of researchers set out to test how well current weather prediction models handle this specific, high-stakes problem. They focused on California, a state where wildfires and subsequent floods are a recurring cycle of disaster. The scientists examined ten separate instances where post-fire floods or debris flows actually occurred between 2021 and 2025. For each event, they looked at the weather forecasts that had been issued in the days leading up to the disaster. They compared the predictions from two advanced computer models, which are designed to simulate the atmosphere at a fine scale, against the actual rainfall recorded by rain gauges near the burned areas. Their goal was not just to see if the models predicted rain, but to determine if they could accurately forecast the intensity and timing of the specific rain bursts that trigger these dangerous flows, and to understand how the accuracy of these predictions changes as the forecast gets further into the future.

The researchers found that these computer models are quite good at telling emergency managers that rain is coming, but they often struggle to predict exactly how hard it will pour. When looking at forecasts issued within a day of the event, the models successfully signaled that heavy rain was likely to fall in the right general area. In technical terms, the models had a very high success rate in detecting that rainfall would exceed a critical threshold of 10 millimeters per hour. This means that if a storm was approaching a burned area, the models were reliable in warning that the conditions were right for a flood or debris flow to start. This is a crucial piece of information for decision-makers, as it provides the lead time needed to take protective action.

However, the study revealed a consistent blind spot in the models' ability to quantify the severity of the rain. While the models correctly identified that rain would fall, they frequently underestimated how intense that rain would be. Across all the events studied, the predicted rainfall amounts were typically lower than what the rain gauges actually measured. This underestimation became more pronounced for the most intense storms, where rainfall rates exceeded 20 millimeters per hour. The models also showed that their accuracy degrades as the forecast extends further into the future. While the predictions were reasonably close to reality within 24 hours of the event, the errors grew larger for forecasts made several days in advance. In some cases, the models missed the peak intensity entirely, predicting a moderate shower when a torrential downpour actually occurred.

The researchers also discovered that the way these models are evaluated matters significantly. If a forecaster looks at a single point on a map, the model might appear to have failed because it predicted the rain in a neighboring grid cell rather than the exact spot where the gauge was located. But when the researchers looked at a broader area, covering a neighborhood of about 9 kilometers around the rain gauge, the models performed much better. This suggests that while the models might not pinpoint the exact location of a raindrop, they are very effective at identifying the general zone where dangerous rainfall is likely to occur. This distinction is vital for emergency response, as it allows officials to focus on protecting entire burned watersheds rather than trying to predict the precise path of a single stream of water.

Despite these successes, the study highlights a persistent limitation: the models tend to be too conservative when it comes to the most extreme rainfall events. Since the volume and speed of a debris flow depend heavily on how hard the rain falls, underestimating the intensity can lead to underestimating the danger. If a model predicts a lower rainfall rate than what actually happens, the resulting flood models might suggest a smaller, less dangerous flow than what is truly coming. The researchers noted that this issue is particularly acute for the most intense storms, which are often the ones that cause the most damage. The models also struggled with specific types of weather systems, such as those driven by isolated pockets of low pressure, which are notoriously difficult to forecast compared to larger, more organized storm systems.

Ultimately, this work provides a clear, data-driven picture of what these weather models can and cannot do for post-fire hazard management. The models are a powerful tool for early warning, capable of identifying the threat of heavy rain days in advance with a high degree of confidence. They give emergency managers the critical heads-up needed to evacuate communities and secure infrastructure. Yet, the study serves as a reminder that these forecasts are not perfect. They often miss the full magnitude of the heaviest rains, and their accuracy fades as the forecast window widens. For those on the front lines of disaster response, this means that while the models are essential for planning, they must be used with an understanding of their tendency to downplay the worst-case scenarios. The best approach involves using these forecasts to prepare for the possibility of extreme conditions, rather than relying on them to predict the exact intensity of the storm.