From Hours to Daily Extremes: Evaluating ERA5 and ERA5-Land Biases Across Morocco’s Diverse Climates

Weather forecasting has entered a new era where powerful computer programs, trained on vast archives of past weather data, are beginning to rival traditional physics-based models. These artificial intelligence systems do not simulate the atmosphere from first principles; instead, they learn patterns from historical records to predict what comes next. The most trusted source for these historical records is a global dataset called ERA5, a detailed reconstruction of the atmosphere that fills in gaps where no weather stations exist. A newer, higher-resolution version called ERA5-Land focuses specifically on the Earth's surface, promising even greater accuracy for the ground we walk on. Because these AI models rely entirely on this data to learn, the quality of the reconstruction is paramount. If the historical record contains hidden errors or smooths over local extremes, the AI will inherit those flaws, potentially failing to predict the very dangerous events society needs to know about.

In Morocco, a country where the landscape shifts dramatically from coastal plains to high mountain ranges and arid deserts, researchers set out to test the reliability of these two datasets. They wanted to know if the digital maps of the past truly matched the reality recorded by weather stations on the ground. The team gathered hourly observations from thirty-six different stations across the country, covering a wide range of climates and topographies from 2016 to 2024. They compared the temperature, humidity, wind speed, and rainfall recorded by the stations against the values provided by both the standard ERA5 and the land-focused ERA5-Land. The goal was to see if the digital reconstructions could capture the subtle, local details of the weather, especially during the most extreme heat, cold, and wind events that define the region's climate.

The investigation revealed that for understanding general climate patterns, both datasets are remarkably good. They successfully reproduce the daily rhythm of the weather, capturing how temperatures rise and fall throughout the day and how seasons change over the year. For average conditions, the two products are largely interchangeable, though the standard ERA5 tends to be slightly more accurate for temperature and humidity, while the land-specific ERA5-Land offers a small advantage in measuring wind speed over complex, mountainous terrain. However, the story changes when the researchers looked at the extremes. While the models handle moderate weather well, their ability to predict the most severe events depends heavily on how rare those events are.

When the researchers tested the models against thresholds for extreme heat, cold, and wind, they found a clear pattern. Both datasets performed well at detecting moderate extremes, such as a very hot day or a strong wind gust. But as the events became rarer and more intense, the models' skill dropped significantly. For the hottest days or the strongest winds, the models often missed the mark entirely. The researchers noted that this failure is not necessarily a unique flaw in the software but a fundamental limitation of trying to match a broad, grid-based digital map to a single, specific point on the ground. A weather station sits in a unique micro-climate, perhaps in a valley or on a specific slope, while the digital model represents a large area that averages out those local details. This mismatch means that even a perfect model cannot predict the exact peak value at a specific station if that peak is driven by local geography the model cannot see.

The study also uncovered a specific technical hurdle for the land-focused dataset. Because ERA5-Land is designed only for land, it leaves out the ocean. This creates a problem for stations located on the coast or near islands, where the model has no valid grid point to match with the observation. In some cases, the data for these coastal stations simply does not exist in the land-only version, a gap that could easily be mistaken for missing weather records if not checked carefully. Furthermore, the researchers found that the higher resolution of the land dataset did not always help. In some desert and mountain areas, the finer grid actually introduced more noise into the temperature and humidity readings, particularly during the night, without improving the overall accuracy.

Ultimately, the research suggests that while these global datasets are excellent tools for studying general climate trends and training AI models, they must be used with caution when predicting specific, life-threatening weather events at a local level. The models are not broken, but they are not designed to see the smallest details of the landscape. The authors recommend that anyone using this data for impact studies or AI verification should focus on relative extremes—how unusual a day is compared to the local average—rather than absolute thresholds like "a day hotter than 45 degrees." They also advise explicitly checking for coverage gaps near the coast and reporting errors in a way that accounts for the natural variability of the wind and temperature. By understanding these limitations, scientists and forecasters can better interpret the digital maps of the past, ensuring that the next generation of weather predictions remains grounded in the complex reality of the world.