Deep Learning-Based Estimation of Seepage Loss in Lined Irrigation Canals

Abstract Accurate estimation of seepage loss from lined irrigation canals is essential for improving irrigation efficiency and water sustainability. This study evaluates and compares five deep learning (DL) models, namely a deep neural network (DNN), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory network (LSTM), and bidirectional long short-term memory network (BiLSTM), for predicting the dimensionless seepage-loss ratio ( q / ky ) from four dimensionless geometric and hydraulic variables. The main contribution of the work is a controlled benchmark of these DL architectures on a lined-canal numerical dataset using identical training/testing partitions, Grid Search Cross-Validation (GSCV), and a multi-criteria validation framework based on error indices, regression error characteristic (REC) curves, Taylor diagrams, uncertainty analysis, and generalization criteria. The dataset comprised 600 numerical cases and was divided into training (70%) and testing (30%) subsets. Results showed that all models achieved high predictive accuracy; however, the CNN provided the most balanced overall performance. In testing, the CNN yielded the best overall testing performance, with (R² = 0.996), the lowest RMSE (0.225), and the lowest testing uncertainty ( U 95 =0.6245), while preserving close agreement with the observed mean and standard deviation. Although the DNN and RNN produced competitive correlation statistics, their testing errors and/or uncertainties were slightly higher. The LSTM and BiLSTM showed larger dispersion and less stable generalization. These findings demonstrate that the CNN was the most reliable model for this dataset when accuracy, uncertainty, and generalization were considered jointly rather than on the basis of R² alone. The proposed framework offers a practical tool for rapid seepage assessment in lined-canal design and rehabilitation.