AlphaEarth and evapotranspiration prediction in Jordan
A retrospective study in Jordan found that while conventional models significantly outperformed baseline predictors for monthly evapotranspiration, augmenting them with AlphaEarth foundation embeddings generally led to a slight deterioration in local prediction accuracy, suggesting the value of such embeddings is conditional on the specific prediction setting and comparator.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Water is the lifeblood of agriculture, and in the arid landscapes of the Middle East, managing it is a matter of survival. For decades, scientists have relied on satellites to watch how much water crops drink and release back into the air, a process called evapotranspiration. These satellite snapshots, such as those from the FAO's WaPOR system, provide a consistent view of water use across vast fields, helping governments and farmers understand where resources are going. However, predicting exactly how much water a field will use next month is difficult. The environment is complex, and the data is often noisy. Recently, a new type of artificial intelligence tool has emerged, capable of learning deep patterns from years of Earth observation data. These tools create compact digital summaries, or "embeddings," that describe the surface of the land in a way that traditional maps cannot. The big question for water managers in places like Jordan is whether these advanced digital summaries can actually improve predictions beyond what we already know from past weather, soil, and crop history.
In a recent study focused on the Jordan Valley and the Mafraq region, researcher Samer Abaddi set out to answer this question with a rigorous, real-world test. The goal was to see if adding these new AI summaries to a standard prediction model would help forecast next month's water use more accurately than the model could on its own. The study looked at 352 specific patches of farmland over a period of several years, using data from 2019 through 2023 to build the models and reserving 2024 and 2025 to test them. The researchers compared two approaches: one that used only conventional information like temperature, rainfall, and the history of water use in that specific field, and another that added the new AI summaries from the previous year. They also tested how well these models worked when moved from one region to another, simulating a scenario where a model trained in one area is used to predict water use in a different landscape.
The results were clear and somewhat surprising. When the researchers tested the models on the same fields where they had been trained, the addition of the new AI summaries did not help. In fact, the predictions became slightly less accurate. The standard model, which relied on familiar data like past rainfall and local history, made errors of about 6.43 millimeters per month. When the AI summaries were added, that error grew to 6.59 millimeters. This means the new information did not provide any extra value; the model was already using the most useful details available. The researchers found that the AI summaries likely contained information that overlapped with what the standard model already knew, such as seasonal patterns and past water use, making the extra data redundant rather than helpful.
However, the story changed when the researchers tested the models in new locations. When they took a model trained in one region and applied it to the other, the AI summaries did provide a boost. In the transfer from Mafraq to the Jordan Valley, the AI-enhanced model reduced the prediction error by about 6.5 percent, performing better than even the simplest historical guesses. Yet, this improvement was not enough to make the model perfect; it still lagged behind what a simple model trained specifically on local data could achieve. In the reverse direction, from the Jordan Valley to Mafraq, the AI help was present but insufficient to overcome the challenges of the new environment. The study concluded that while these advanced AI tools are powerful, they are not a universal upgrade. Their value depends entirely on the specific situation and what information is already available.
This research offers a crucial lesson for the future of water management in Jordan and similar regions. It suggests that simply adding more complex artificial intelligence to a system does not guarantee better results. If a model already has access to a rich history of local weather and water use, adding a new layer of data might not improve the forecast. The most effective approach remains one that respects the specific history of the land. For water managers, this means that before adopting new AI technologies, they must carefully test whether these tools actually add something new to the mix. The study confirms that while these digital summaries are impressive, they are not a magic solution that works everywhere. Instead, they are a tool that works best in specific gaps, such as when trying to predict water use in a place where no local history exists, rather than in places where the past is already well understood.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.