The research tackles a problem that has quietly plagued evapotranspiration modeling for decades. The gold standard for computing ETo, the FAO-56 Penman-Monteith equation, requires a full suite of meteorological measurements including air temperature, solar radiation, wind speed, and humidity. Many weather stations around the world, particularly in developing regions, simply do not record all of these variables reliably. Machine learning models have emerged as powerful alternatives because they can learn the complex, non-linear relationships between whatever meteorological inputs are available and the evapotranspiration rates observed at a site. Yet feeding a model every possible variable is not always wise: redundant or irrelevant inputs inflate model complexity, slow computation, and can degrade accuracy through noise. The question of which variables to keep, and which to discard, is known as feature selection, and it is precisely where the new study makes its mark.

The researchers built a hybrid framework that couples seven distinct feature selection methods with an Extreme Gradient Boosting model, better known as XGBoost. XGBoost was chosen deliberately for its proven track record with tabular meteorological datasets, its built-in regularization that helps prevent overfitting, its native ability to capture non-linear interactions between variables, and its computational efficiency on moderately sized station records. The seven selection techniques span the two major families of feature selection. Filter-based methods, which evaluate variables independently of any predictive model, included the Pearson correlation coefficient, the minimum redundancy-maximum relevance criterion known as mRMR, the ReliefF algorithm, the Fisher score, and the Stability method. Embedded methods, which perform selection as part of model training, included LASSO regression and SHAP, the SHapley Additive exPlanations approach borrowed from the field of interpretable machine learning.

Each of these techniques interrogates the data in a fundamentally different way. Pearson correlation measures simple linear association between each meteorological variable and ETo. The mRMR criterion, rooted in information theory, seeks variables that are highly relevant to the target while being minimally redundant with one another, avoiding the trap of selecting several measurements that all encode the same information. ReliefF evaluates how well a variable distinguishes between neighboring instances that are similar or dissimilar in the target value, making it sensitive to feature interactions. The Fisher score assesses class separability in the feature space, while the Stability method quantifies how consistently a selection algorithm picks the same variables across resampled datasets, a property that matters enormously when models must be trusted on new data. LASSO, by contrast, shrinks regression coefficients toward zero and effectively eliminates weak predictors through regularization, and SHAP derives feature importance from game-theoretic principles by computing each variable’s marginal contribution to model predictions.

To test the framework under genuinely different atmospheric regimes, the team used data from two contrasting climatic conditions, an arid region and a humid region, with records obtained from the Islamic Republic of Iran Meteorological Organization. The study was structured in two modeling stages. In the first stage, the three most important features identified by each selection technique were fed into XGBoost as inputs. In the second, more aggressive stage, only the two most important features were used. This staged design functions as a feature-based domain adaptation procedure, systematically evaluating the relevance and redundancy of key meteorological variables while minimizing the input space, which in turn reduces model complexity and computational cost. The practical payoff is a set of modeling tools suited to data-scarce conditions, where reliable measurements are limited and every sensor counts.

The results are striking. In the arid region, the best-performing combinations achieved a coefficient of determination of at least 0.980, a root mean square error no greater than 0.344 millimeters per day, and a mean absolute error of at most 0.270 millimeters per day. The humid region performed even better, with R-squared values of at least 0.982, RMSE as low as 0.188 millimeters per day, and MAE down to 0.142 millimeters per day. For a variable that typically ranges across several millimeters per day seasonally, sub-0.2 millimeter daily errors represent an extraordinary level of precision, comparable to the uncertainty of the reference equation itself.

Perhaps the most actionable finding is that the optimal feature selection method depends on climate. In the arid region, ReliefF and LASSO delivered the lowest error indices when only the two most important features were used. In the humid region, the Pearson, SHAP, Fisher, and Stability methods produced the best results with the same two-feature approach. This divergence makes physical sense. In arid environments, where atmospheric demand is driven largely by energy availability and vapor pressure gradients, instance-based and regularization-based selectors appear to isolate the dominant drivers most effectively. In humid conditions, where radiation and temperature interplay differently and air moisture is less limiting, correlation-based and interpretability-driven methods capture the signal more cleanly. The lesson for practitioners is that feature selection is not a one-size-fits-all exercise; the choice of selector should be matched to the climatic context of the station being modeled.

The authors did not stop at raw accuracy figures. They subjected their models to bootstrap analysis, effect size assessment, and decision boundary analysis to evaluate robustness and statistical significance rather than relying on single-point error metrics alone. Bootstrap resampling tests whether performance holds across repeated random draws from the data, guarding against lucky splits. Effect size analysis quantifies the magnitude of differences between competing models, ensuring that reported improvements are meaningful rather than trivially small. Decision boundary analysis probes how models behave across the full range of input conditions, revealing whether accuracy is uniform or concentrated in particular regimes. Together, these diagnostics lend the headline numbers a level of statistical credibility that many applied machine learning studies lack.

As a final benchmark, the XGBoost models were compared against Random Forest and a feed-forward Artificial Neural Network, two of the most widely used alternatives in hydrological machine learning. XGBoost came out on top, confirming that its gradient-boosted tree architecture, with its regularization and efficient handling of non-linear interactions, is particularly well suited to the structure of daily meteorological records. This finding adds to a growing body of evidence that gradient boosting is a formidable default choice for tabular environmental data, often outperforming deep networks on datasets of modest size.

The implications reach well beyond Iranian weather stations. As climate change intensifies pressure on freshwater supplies, water managers in arid and semi-arid regions worldwide face the twin challenges of rising crop water demand and sparse monitoring infrastructure. A modeling recipe that achieves R-squared values above 0.98 using only two meteorological inputs means that stations lacking full instrumentation can still produce trustworthy evapotranspiration estimates for irrigation scheduling and drought assessment. The framework’s emphasis on stability and interpretability, through methods like SHAP and the Stability criterion, also aligns with the broader push for transparent artificial intelligence in environmental science, where black-box predictions alone are no longer sufficient for policy decisions. By demonstrating that the marriage of thoughtful feature selection and gradient boosting can squeeze remarkable accuracy from minimal data, the study offers a practical template for the next generation of evapotranspiration models, one that could help stretch every drop of information as far as every drop of water.

Subject of Research: Machine learning-based feature selection for estimating daily reference evapotranspiration from meteorological data

Article Title: Assessing machine learning-based feature selection methods for estimating daily reference evapotranspiration using meteorological data

Article References: Sadeghzadeh, M., Karimi, S., Shiri, J., & Sivakumar, B. (2026). Assessing machine learning-based feature selection methods for estimating daily reference evapotranspiration using meteorological data. Theoretical and Applied Climatology, 157(10), Article 629. https://doi.org/10.1007/s00704-026-06568-2

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06568-2

Keywords: reference evapotranspiration, feature selection, XGBoost, machine learning, meteorological data, ReliefF, LASSO, SHAP, arid climate, humid climate, irrigation scheduling, Theoretical and Applied Climatology