Comparative Evaluation of Empirical and Data‐Driven Techniques for Estimating Seepage Losses in Irrigation Canals

In irrigation canal systems, accurately quantifying seepage losses is crucial for enhancing water conveyance efficiency. This study analyses seepage estimation methods using 6 years of inflow–outflow data from the Thal Canal System in Pakistan, focusing on plain cement concrete (PCC)–lined, brick‐lined and unlined reaches. Eight empirical equations were evaluated against observed seepage losses, and their results were compared with those of the metaheuristic‐optimised CatBoost model and the symbolic regression (PySR) model. The seepage losses were 5% for PCC‐lined canals, 13% for brick‐lined canals and 22% for unlined canals. The CatBoost model demonstrated the highest predictive performance relative to the inflow–outflow‐derived reference estimates ( R 2 = 0.97, RMSE = 0.80 m 3 /s), whereas PySR also performed well ( R 2 = 0.94, RMSE = 1.22 m 3 /s) and produced a single closed‐form equation. Among the empirical methods, the Chinese equation was most effective for unlined canals ( R 2 = 0.63, RMSE = 1.58 m 3 /s), and the Punmia equation performed best among the empirical equations evaluated for the PCC‐lined reach ( R 2 = 0.61, RMSE = 1.49 m 3 /s). Overall, CatBoost provided the strongest predictive performance, whereas PySR offered a useful balance between predictive accuracy and mathematical transparency. The empirical equations showed limited uncalibrated transferability across the investigated lining conditions.