An Explainable Deep Learning Framework for Evapotranspiration Prediction Using Hydro-Meteorological Data

Water is the thread that stitches the atmosphere to the land, moving constantly between the sky and the soil. When rain falls, some of it soaks into the ground, some flows into rivers, and some disappears back into the air as vapor through a process called evapotranspiration. This is a combination of water evaporating from the ground and plants releasing moisture through their leaves. Understanding exactly how much water makes this journey is critical. It tells farmers when to irrigate, helps engineers design reservoirs, and allows disaster managers to predict when rivers might swell into floods. For decades, scientists have relied on complex formulas to estimate these water movements, but these methods often require mountains of weather data and careful tuning that makes them difficult to use in places where records are incomplete.

A new approach, detailed in recent research, suggests that computers can learn these water patterns directly from history, without needing rigid formulas. This study introduces a framework that uses deep learning, a type of artificial intelligence capable of finding hidden patterns in vast amounts of information. The researchers combined several different computer architectures to act as a team: some to spot local patterns in the landscape, others to remember how weather changes over time, and a special system to explain exactly why the computer made a specific prediction. By weaving together weather data, geographic location, and this new kind of artificial intelligence, the study offers a clearer, more transparent way to forecast how water moves across the Earth, a vital step for managing resources in a changing climate.

The researchers set out to build a system that could predict evapotranspiration and potential evapotranspiration, which is the maximum amount of water that could evaporate if water were unlimited. To do this, they gathered secondary data on rainfall, temperature, humidity, wind speed, solar radiation, and the specific latitude and longitude of a location. Instead of relying on a single computer model, they trained four different types of deep learning networks to analyze this information. One type, known as a neural network, served as a baseline to find general relationships. Another, a convolutional network, was designed to look at the data as if it were an image, helping to identify how features like temperature or rainfall vary across different parts of a landscape. Two other types, known as recurrent networks, were specifically built to understand time. They could look at a sequence of weather events and remember how yesterday's rain or humidity influenced today's conditions.

The study found that these different computer minds had distinct strengths. The networks designed to handle time, which can remember past weather events, proved better at capturing the flow of hydrological patterns than the standard networks that only look at the present moment. This makes sense, as water movement is a story that unfolds over days and weeks, not just a single snapshot. The networks that analyzed spatial features also improved the predictions by recognizing that a hot, dry patch of land behaves differently than a humid valley, even if the air temperature is the same. By combining these approaches, the framework created a more complete picture of how water moves through the environment.

However, a major hurdle with artificial intelligence is that it often acts as a "black box," giving an answer without explaining how it reached that conclusion. In fields like water management, where decisions affect millions of lives, knowing the "why" is just as important as the "what." To solve this, the researchers integrated a tool called SHAP, which acts like a magnifying glass for the computer's decision-making. This tool breaks down the final prediction to show exactly how much each factor contributed. The analysis revealed that rainfall, temperature, and solar radiation were the primary drivers of evapotranspiration, while humidity and wind speed also played significant roles. Crucially, the system showed that the geographic location itself mattered, confirming that where a place sits on the map influences its water cycle.

The results were visualized on maps that showed how evapotranspiration varies across a region. Areas with higher temperatures and intense sunlight showed higher rates of water loss, while more humid locations retained more moisture. This spatial view allowed the researchers to see the water cycle not just as a number, but as a landscape that shifts and changes. The study suggests that by integrating these different computer models with geographic maps and explainable tools, scientists can move beyond simple guessing. The framework does not just predict the future; it explains the logic behind the prediction, making it a trustworthy tool for planners.

This research does not claim to have solved every problem in water forecasting. The study relied on existing data and offered a qualitative comparison of how these different computer models performed, rather than providing a final, definitive set of numbers for every possible scenario. The authors note that future work will need to test these methods with larger datasets and more complex variables, such as soil moisture and vegetation health. Yet, the work establishes a clear path forward. It demonstrates that we can build intelligent systems that are not only accurate but also understandable. For the managers of our water resources, this means having a tool that can forecast floods and droughts while clearly showing the weather and geography that drive those forecasts, turning complex data into actionable knowledge for a more resilient future.