Optimized convolutional neural network - kernel ridge regression-error correction method: an advanced model for predicting soil saturated hydraulic conductivity Authors: Parisa Kahkhamoghadam, Mohammad Mahdi Chari, Mohammad ehteram Published: 2026 ABSTRACT Predicting soil saturated hydraulic conductivity (Ks) is essential for understanding water movement in soils and designing efficient irrigation and drainage systems. Thus, our paper proposes the mutated grasshopper optimization (MGRO) algorithm- convolutional neural network (CNN)- kernel ridge regression (KRR)-error correction (ERC) method to predict Ks. The MGRO-CNN-KRR-ERC (MCKE) model is executed in several steps. First, the MGRO algorithm optimizes the hyperparameters of the CNN and KRR techniques. Then, the CNN technique is trained on the input data to produce a set of outputs called the summary vector, which represents the compact and meaningful descriptor of the original dataset for subsequent KRR processing. Next, the summary vector is fed into KRR to produce predicted Ks data. Finally, the ERC method reduces the difference between these data and measured Ks values to improve the overall prediction accuracy. The model uses four soil properties as input data. Our study uses the mutual information index to select the most informative features, and the results indicate that all four properties significantly affect the precision of predictions. The current research also utilizes the kernel estimation method to determine the uncertainty of the predicted Ks values. The study results demonstrate that the MCKE model significantly outperforms KRR, CNN, and hybrid approaches in terms of predictive accuracy, error reduction, and robustness. Specifically, the MCKE model achieves a reduction of 51–93% in mean absolute percentage error (MAPE to all other predictive models. In general, the MCKE model provides a robust and reliable framework for predicting Ks under diverse environmental conditions. Introduction Soil saturated hydraulic conductivity (KS) is a fundamental soil physical property that quantifies the ability of soil to transmit water under fully saturated conditions (e.g., when soil pores are completely filled with water)1. This property plays a key role in regulating water retention and aeration in the root zone2. It also serves as a critical factor in determining irrigation efficiency, which directly influences the uniformity of water distribution in the field3. In recent years, KS has received increasing attention, and numerous studies have focused on predicting it. Predicting KS is essential for evaluating soil permeability and designing efficient irrigation and drainage systems2. It is also needed for accurately simulating solute transport and the movement of nutrients and contaminants in soil profiles under saturated conditions3–5. Thus, various tools such as machine learning (MAL) models have been developed to predict KS. MAL models can effectively model nonlinear relationships between soil properties, climate factors, and flows of water under both natural and managed conditions6. They can also uncover hidden patterns and nonlinear and handle complexity of the data7,8. MAL models can incorporate different types of data, such as soil physical properties (e.g., texture, porosity, and bulk density), which are needed for the accurate prediction of KS9. They also reduce the dependency on extensive field measurements by learning patterns directly from available datasets6. In recent years, scholars have proposed various MAL models to predict Ks. For example, Jain et al.10 predicted Ks using the random forest (RF) and ANN models. It was found that the ANN model achieved an R² of 0.83 and produced more accurate outputs compared to the RF model. However, these two models are sensitive to hyperparameter settings and may suffer from overfitting when applied to heterogeneous datasets. Albalasmeh et al.11 proposed the generalized regression neural network (GRNN) to predict Ks. It was found that the GRNN model achieved different prediction accuracies with various input combinations. However, the GRNN model was not able to effectively handle high-dimensional data and complex nonlinear interactions among soil properties. Moosavi et al.12 proposed the particle swarm optimization (PSO) algorithm-artificial neural network (ANN) to predict Ks. The study results demonstrated that the PSO-ANN model reduced the root mean square error (RMSE) of the genetic algorithm-ANN by 15%. However, these models may experience premature convergence and local optima issues during the optimization process. Table 1 lists previous studies that used different techniques to predict Ks. It also reports the performance of various predictive models and summarizes their results. Table 1Summary of previous studies on Ks prediction using different modeling techniques.AuthorModel InputsModelEvaluation metricsJian et al.10Bulk density (BD) and percentages of silt, sand, and clayANN and random forestR2 of ANN: 83R2 of random forest:0.54Albalasmeh et al.11Organic matter, bulk density, pHGRNNR = 0.93Granata et al.9BD, soil organic carbon content, standard deviation of soil particle diameter, geometric mean diameter, clay percentage, and silt percentageMulti-layer perceptron (MLP)Support vector machine (SVR)R2 of MLP = 0.63R2 of SVR = 0.76Moosavi et al.12Geometric mean diameter (GMD) of aggregates, pH, and calcium carbonate equivalent (CCE)PSO-ANNGA-ANNRMSE of PSO-ANN:0.34RMSE of GA-ANN:0.40Mozaffari et al.13BD, pH, potassium (K+), calcium (Ca2+), and magnesium (Mg2+), and sodium adsorption ratio (SAR)Radial basis function (RBFNN) and multiple linear regressionR2 of RBFNN = 0.58 to 0.78R2 of MLR = 0.52 to 0.63 Table 1Summary of previous studies on Ks prediction using different modeling techniques.AuthorModel InputsModelEvaluation metricsJian et al.10Bulk density (BD) and percentages of silt, sand, and clayANN and random forestR2 of ANN: 83R2 of random forest:0.54Albalasmeh et al.11Organic matter, bulk density, pHGRNNR = 0.93Granata et al.9BD, soil organic carbon content, standard deviation of soil particle diameter, geometric mean diameter, clay percentage, and silt percentageMulti-layer perceptron (MLP)Support vector machine (SVR)R2 of MLP = 0.63R2 of SVR = 0.76Moosavi et al.12Geometric mean diameter (GMD) of aggregates, pH, and calcium carbonate equivalent (CCE)PSO-ANNGA-ANNRMSE of PSO-ANN:0.34RMSE of GA-ANN:0.40Mozaffari et al.13BD, pH, potassium (K+), calcium (Ca2+), and magnesium (Mg2+), and sodium adsorption ratio (SAR)Radial basis function (RBFNN) and multiple linear regressionR2 of RBFNN = 0.58 to 0.78R2 of MLR = 0.52 to 0.63 Summary of previous studies on Ks prediction using different modeling techniques. R2 of ANN: 83 R2 of random forest:0.54 Multi-layer perceptron (MLP) Support vector machine (SVR) R2 of MLP = 0.63 R2 of SVR = 0.76 PSO-ANN GA-ANN RMSE of PSO-ANN:0.34 RMSE of GA-ANN:0.40 R2 of RBFNN = 0.58 to 0.78 R2 of MLR = 0.52 to 0.63 However, previous models such as numerical models, empirical techniques, ANN, and regression approaches have different drawbacks. For example, classical regression approaches are generally unable to capture strong nonlinear relationships between soil properties and Ks. The empirical and numerical models may not adequately capture temporal and spatial variability in hydraulic conductivity. The ANN technique may struggle to extract long-range dependencies from highly heterogeneous soil systems12,13. The existing Ks prediction models have disadvantages such as weak nonlinear fitting ability, large systematic errors, and difficulty in practical deployment. These models also suffer from limited generalization capability when applied to diverse environmental conditions and heterogeneous soil datasets. Therefore, the main motivation of the current paper is to propose a Ks prediction model that addresses these drawbacks and can accurately predict soil saturated hydraulic conductivity. In recent years, data scientists have proposed other types of MAL models, such as the kernel ridge regression (KRR) technique. The KRR technique combines the advantages of ridge regression and kernel methods to model complex nonlinear relationships in time series and hydrological processes14,15. In recent years, scholars have utilized the KRR technique to forecast different variables. For example, Ali et al.16 combined a decomposition algorithm with KRR to forecast rainfall. It was found that the hybrid technique reduced the mean absolute error (MAE) of the KRR by 59%. Ahmed et al.17 developed CEEMDAN- grey wolf optimization (GWO)- KRR technique to predict wheat yield. The study results showed that the proposed CEEMDAN- GWO-KRR approach provided reliable estimates and reduced prediction errors of the KRR by 20%. However, the performance of this approach was influenced by different factors such as the size and diversity of the training dataset. Deo et al.18 used the KRR technique to monitor solar energy. Their findings indicated that KRR produced higher accuracy compared to simple regression method. However, this technique was not able to effectively handle highly noisy input data under rapidly changing atmospheric conditions. Kordani et al.19 used different types of the KRR technique to forecast water quality index. It was found that the KRR-based models achieved high prediction accuracy and strong generalization performance. Specifically, the regularized locally weighted (GRKR) method reduced the MAPE of the LSSVM model by 1.3%. Ahmadianfar et al.20 combined the singular value decomposition (SVD) algorithm with the KRR technique to forecast water levels. The study results showed that the combined approach reduced the MAPE of the KRR method by 68%. However, hydrologists need more accurate predictions of KS to develop sustainable land and water management strategies under changing climate conditions. Moreover, these predictions are vital for ensuring sustainable crop production and minimizing soil erosion13. Thus, it is essential to enhance the predictive performance of the KRR technique by combining it with the advanced tools such as deep learning techniques. Deep learning techniques are advanced computational methods for simulating complex systems21. These techniques can capture complex nonlinear relationships in large and high-dimensional datasets22,23. They also offer significant improvements in modeling dynamic processes in soil profiles24. Deep leaning techniques includes different variants such as the convolutional neural network (CNN). The CNN technique improves the ability of the machine learning models to extract nonlinear features (e.g. structural dependencies and multiscale patterns) from input data24. This enhancement is necessary as it helps these models to obtain insights into spatial and temporal variability in soil properties under different climate and land-use conditions, ultimately leading to more accurate predictions of Ks24. In recent years, the CNN technique has been widely combined with different machine learning models to create advanced predictive frameworks for forecasting various environmental variables and ecosystem indicators. For example, Ehteram et al.21 coupled the CNN technique with a decision tree (DT) model to forecast water quality index. They reported that the CNN-DT model provided more reliable predictions compared DT. Ehteram et al.24 integrated the CNN technique with the support vector model to forecast rainfall in a large basin. Their findings indicated that the CNN-SVM model achieved higher prediction accuracy compared to traditional SVM and standalone CNN models. Thus, our study combines the CNN technique with KRR to improve its predictive performance. The resulting hybrid system, called CNN-KRR, is then utilized to predict Ks. However, to further enhance the performance of KRR, the CNN-KRR framework can be combined with tools such as error correction (ERC) method. The ERC method minimizes the difference between predicted and observed outputs25. This minimization is essential as it helps the prediction model reduce systematic errors and improve overall prediction accuracy. Recently, developers have widely coupled ERC with hybrid prediction approaches to improve their precision. For example, Ye et al.25 combined the error correction method with the gradient boosting tree (GBT) to forecast wind speed. It was found that the ERC-GBT model achieved more reliable outputs compared to standalone GBT. Xu et al.26 developed the ERC-artificial neural network (ANN) model to forecast runoff. It was found that the ERC-ANN model significantly improved runoff prediction accuracy. Thus, our study combines the ERC method with the CNN-KRR framework to develop the hybrid CNN-KRR-ERC (CKE) model for predicting KS. The model effectively reduces systematic prediction errors and provides reliable predictions of KS. It can also serve as an analytical tool for understanding spatial and temporal variability of water flow in heterogeneous soils. However, the CKE model includes CNN and KRR parameters, which should be effectively adjusted to achieve optimal prediction accuracy and generalization performance. Recently, modelers have developed various tools, such as optimization techniques, to effectively tune the parameters of hybrid models. These techniques systematically adjust parameters of prediction models, allowing them to achieve optimal predictive performance27. They also improve the adaptability of the hybrid model to different soil types and environmental conditions28. Optimization algorithms includes different types such as grasshopper optimization (GRO) technique, which is inspired by the natural behavior of grasshoppers in nature. The GRO technique adjusts model parameters by systematically exploring the parameter space28. It also offers several advantages such as the flexibility to handle both continuous and discrete variables27. However, to more effectively set the parameters of the prediction models, the effectiveness of the GRO technique should be enhanced by combining it with advanced tools such as the mutation operator. The mutation operator, which is an optimization strategy, alters the location of a selected grasshopper, which serves as a search agent within the search space. This alteration is essential as it enables GRO to diversify the search paths, leading to the discovery of high-quality solutions. The mutation operator imrpoves the search capability of the algorithm by introducing random variations into the population. It also improves the balance between exploration and exploitation, allowing the algorithm to search both locally and globally within the solution space. Thus, our study combines the mutation operator with the GRO technique to develop a new optimization algorithm called the mutated GRO (MGRO) technique. The MGRO technique offers enhanced global search capabilities, enabling it to explore a wider solution space. It can also handle multi-objective optimization problems, allowing simultaneous optimization of competing objectives such as accuracy, computational cost, and generalization performance. Thus, our research combines the MGRO technique with CNN-KRR-ERC to properly set its parameters. The resulting hybrid approach, called MGRO-CNN-KRR-ERC (MCKE), is then used to predict KS. Thus, the main novelty of the paper is the development of the MCKE model. The model serves as an effective framework for evaluating the influence of soil texture, porosity, and environmental factors on water movement. It also offers insights into soil hydraulic behavior under varying climate conditions. These insights are essential for supporting sustainable agricultural management and optimizing soil and water conservation practices. The MCKE model produce reliable outputs under variable environmental conditions by updating its predictions based on incoming weather and climate information, such as rainfall, temperature, and evapotranspiration. It can also incorporate irrigation and water management schedules, enabling the accurate simulation of dynamic soil–water interactions. The MCKE model consists of multiple components, which enable it to capture nonlinear relationships, correct prediction errors, and optimize model parameters simultaneously. Without using CNN, the model fails to effectively represent the complex nonlinear relationships in soil hydraulic processes. The lack of error correction method leads to the persistence of systematic biases in the predicted Ks values. While recent studies, including Zhang et al.29 and Zhang et al.25, have developed various hybrid models for predicting different target variables, the proposed MCKE model introduces a comprehensive framework that simultaneously improves feature extraction, nonlinear mapping, and error minimization in soil saturated hydraulic conductivity prediction. Compared to existing hybrid prediction models, the MCKE model significantly improves the representation of multiscale soil characteristics, leading to more reliable estimations of saturated hydraulic conductivity. Unlike existing hybrid prediction models, which may suffer from high bias and limited generalization, the MCKE model reduces systematic errors through an integrated error correction mechanism. The MCKE model also provides improved generalization performance across different soil types and environmental conditions due to its hybrid learning structure. In the next section of the study, the theories and mathematical formulation of the components of the MCKE model are presented to provide a comprehensive understanding of its structure. In the third section, a case study is presented to evaluate the performance of the MCKE model in predicting KS. The fourth section examines the performance of the MCKE model using various performance metrics and graphical tools. This section also provides a comprehensive discussion, which highlights the strengths and limitations of the model in terms of prediction accuracy, robustness, and generalization capability. Finally, the last section summarizes the main findings. Materials and methods Grasshopper optimization (GRO) technique The grasshopper optimization (GRO) technique is a nature-inspired metaheuristic algorithm that mimics the collective behavior of grasshoppers in nature30. This behavior is mainly governed by social interaction, gravity force, and wind advection, which together influence the movement and distribution of grasshoppers during foraging and migration31. By modeling these three components, the GRO technique can explore complex and nonlinear search spaces to identify optimal solutions. In the search space, the position of each grasshopper represents a candidate solution and is defined based on Eq. 131. This equation includes three components, which are described as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_i}={S_i}+{G_i}+{A_i}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_i}$$\end{document}: The location of a grasshopper, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S_i}$$\end{document}:The social interaction among grasshoppers,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${G_i}$$\end{document}: The gravity force on the i-th grasshopper, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A_i}$$\end{document}: The wind advection. The social interaction is a component that regulates the movement of each grasshopper based on the positions of its neighbors30. Equation 2 mathematically defines this component:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S_i}=\sum\limits_{{j=1}}^{N} {s\left( {{d_{ij}}} \right)} {\hat {d}_{ij}}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat {d}_{ij}}=\frac{{{x_i} - {x_j}}}{{{d_{ij}}}}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_j}$$\end{document}: The positions of the ith and jth grasshoppers, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d_{ij}}$$\end{document}: The distance between the i-th and the j-th grasshopper, s: It is a function that defines social forces. These forces refer to the attraction and repulsion interactions among grasshoppers, which allows them to maintain an appropriate distance from one another. The function s is mathematically defined as follows30.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s\left( r \right)=f{e^{\frac{{ - r}}{l}}} - {e^{ - r}}$$\end{document} Where r: a real value, f: The intensity of attraction and l: The attractive length scale. The gravity force is a component that moves grasshoppers toward the center of the search space. Equation 5 mathematically defines this component30.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${G_i}= - g \times {\hat {e}_g}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat {e}_g}$$\end{document}: A unity vector in the vertical direction of the surface and g: The gravitational constant. The wind direction is a component that guides the grasshopper population toward the best solution found so far. Equation 6 mathematically defines this component6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A=u{\hat {e}_w}$$\end{document} Where u:A constant drift and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat {e}_w}$$\end{document}: A unity vector in the direction of wind. Finally, Eq. 1 is rewritten using Eqs. 2–6 to produce a mechanism for updating the location of each grasshopper. Equation 7 shows the mathematical formulation of this mechanism.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{d}=c\left( {\sum\limits_{\begin{subarray}{l} j=1 \\ j \ne i \end{subarray} }^{N} {s\left( {\left| {{x_j} - {x_i}} \right|} \right)} \frac{{{x_j} - {x_i}}}{{{d_{ij}}}} - g{{\hat {e}}_g}+u{{\hat {e}}_w}} \right)+{\hat {T}_d}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{d}$$\end{document}: The updated position of the ith grasshopper, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_j}$$\end{document}: The position of the jth grasshopper, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_i}$$\end{document}: The position of the ith grasshopper, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat {T}_d}$$\end{document}: The position of the best grasshopper. Notably, the best grasshopper refers to the search agent that attains the optimal value of a predefined fitness function. Mutated GRO (MGRO) The MGRO algorithm is an advanced version of the GRO technique, which is implemented as follows: First, MGRO produces an initial population of grasshoppers.The locations of grasshoppers, which represent candidate solutions, are randomly initialized.An objective function (e.g. MAPE) is utilized to assess the quality of each candidate solution.The location of each grasshopper is updated using Eq. 7.The mutation operator is applied according to Eq. 8, which alters the location of a selected grasshopper based on its current position, the positions of two selected grasshoppers, and a random number. 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{{mut}}={X_i}+\beta \left( {{X_{rand1}} - {X_{rand2}}} \right)$$\end{document} First, MGRO produces an initial population of grasshoppers. First, MGRO produces an initial population of grasshoppers. The locations of grasshoppers, which represent candidate solutions, are randomly initialized. The locations of grasshoppers, which represent candidate solutions, are randomly initialized. An objective function (e.g. MAPE) is utilized to assess the quality of each candidate solution. An objective function (e.g. MAPE) is utilized to assess the quality of each candidate solution. The location of each grasshopper is updated using Eq. 7. The location of each grasshopper is updated using Eq. 7. The mutation operator is applied according to Eq. 8, which alters the location of a selected grasshopper based on its current position, the positions of two selected grasshoppers, and a random number. The mutation operator is applied according to Eq. 8, which alters the location of a selected grasshopper based on its current position, the positions of two selected grasshoppers, and a random number. Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_{rand1}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_{rand2}}$$\end{document}: The positions of two selected grasshoppers from the population, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{{mut}}$$\end{document}: The location of the ith grasshopper after applying the mutation operator, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_i}$$\end{document}: The location of the ith grasshopper, and β: A random number. The termination criterion (e.g., reaching a maximum number of iterations or achieving a predefined minimum fitness threshold) is controlled. If the condition is unmet, the algorithm goes back to the step 3; otherwise, it proceeds to the next stage.The grasshopper with the top fitness value is identified. Notably, the location of this grasshopper represents an optimal solution, which is reported as the final result. The termination criterion (e.g., reaching a maximum number of iterations or achieving a predefined minimum fitness threshold) is controlled. If the condition is unmet, the algorithm goes back to the step 3; otherwise, it proceeds to the next stage. The termination criterion (e.g., reaching a maximum number of iterations or achieving a predefined minimum fitness threshold) is controlled. If the condition is unmet, the algorithm goes back to the step 3; otherwise, it proceeds to the next stage. The grasshopper with the top fitness value is identified. Notably, the location of this grasshopper represents an optimal solution, which is reported as the final result. The grasshopper with the top fitness value is identified. Notably, the location of this grasshopper represents an optimal solution, which is reported as the final result. CNN technique The CNN technique is a deep learning model that has gained significant attention in recent years. This technique offers different advantages such as the ability to automatically extract features from complex datasets21. It also effectively handles heterogeneous data and complex structured data such as video frames, audio spectrograms, and biomedical signals24. The CNN technique includes multiple layers: Input layer: This layer first receives the raw input data, such as soil physical properties (texture, porosity, bulk density), climate variables (precipitation, temperature), and land-use factors. It then prepares and formats the data for subsequent processing.Convolutional layers: This layer consists of trainable filters that extract a set of nonlinear features (e.g., spatial and structural patterns) from the input data. Notably, this set is known as the feature map21.Pooling layer: This layer, which acts as the second component of CNN, reduces the spatial dimensions of the feature map. This reduction is essential because it helps decrease computational complexity32.Fully connected layer: The feature map with reduced dimensions is input to the fully connected layer, which converts it into a one-dimensional array called the summary vector. This vector serves as the final output of CNN. It also includes spatial, structural, and textural information, which is needed for the accurate prediction of Ks. Input layer: This layer first receives the raw input data, such as soil physical properties (texture, porosity, bulk density), climate variables (precipitation, temperature), and land-use factors. It then prepares and formats the data for subsequent processing. Input layer: This layer first receives the raw input data, such as soil physical properties (texture, porosity, bulk density), climate variables (precipitation, temperature), and land-use factors. It then prepares and formats the data for subsequent processing. Convolutional layers: This layer consists of trainable filters that extract a set of nonlinear features (e.g., spatial and structural patterns) from the input data. Notably, this set is known as the feature map21. Convolutional layers: This layer consists of trainable filters that extract a set of nonlinear features (e.g., spatial and structural patterns) from the input data. Notably, this set is known as the feature map21. Pooling layer: This layer, which acts as the second component of CNN, reduces the spatial dimensions of the feature map. This reduction is essential because it helps decrease computational complexity32. Pooling layer: This layer, which acts as the second component of CNN, reduces the spatial dimensions of the feature map. This reduction is essential because it helps decrease computational complexity32. Fully connected layer: The feature map with reduced dimensions is input to the fully connected layer, which converts it into a one-dimensional array called the summary vector. This vector serves as the final output of CNN. It also includes spatial, structural, and textural information, which is needed for the accurate prediction of Ks. Fully connected layer: The feature map with reduced dimensions is input to the fully connected layer, which converts it into a one-dimensional array called the summary vector. This vector serves as the final output of CNN. It also includes spatial, structural, and textural information, which is needed for the accurate prediction of Ks. The CNN model includes different parameters such as the number of convolutional filters, filter size, batch size, and epoch number, which should be carefully chosen. Thus, our study uses the MGRO technique to properly set these parameters. KRR technique The KRR technique is the predictive engine of the MCKE model. This technique is an extension of linear ridge regression that can effectively address nonlinear problems19. It can also effectively process heterogeneous data with nonuniform distributions, enabling robust modeling Ks under diverse environmental conditions. To provide an output for a training data set, the KRR technique uses Eq. 9 including a kernel function and a weight coefficient. This coefficient quantifies the influence of each training sample on the final prediction and is mathematically defined by Eq. 1020.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y=\sum\limits_{{i=1}}^{n} {{\alpha _i}} K\left( {{x_i},x} \right)$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha ={\left( {K+\lambda I} \right)^{ - 1}}y$$\end{document} Where α:A weight of training point Y: The predicted output, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K\left( {{x_i},x} \right)$$\end{document}: The kernel function, λ: The regularization parameter, I: The identity matrix, and y: The vector of observed outputs. The regularization parameter balances the complexity of the technique with the accuracy of predictions on the training data. Moreover, it is important to note that the kernel function enables the KRR technique to model nonlinear relationships between input variables and the target output. Equation 11 mathematically defines this function20.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K\left( {{x_i},{x_j}} \right)=\exp \left( { - \frac{{\left\| {{x_i} - {x_j}} \right\|}}{{2{\sigma ^2}}}} \right)$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_j}$$\end{document}: Input variables, and σ: The kernel parameter. It is important to note that the kernel parameter should be carefully tuned to achieve the highest prediction accuracy. Thus, our paper utilizes MGRO to systematically optimize the kernel parameter. Error correction (ERC) method The ERC method is the last component of the MCKE model, which is executed as follows: First, a machine learning model such as KRR predicts error predictions, which are produced from the initial prediction33.Then, the predicted errors are added to the outputs of the initial prediction, allowing MCKE to reduce systematic deviations and improve the overall accuracy of KS predictions. First, a machine learning model such as KRR predicts error predictions, which are produced from the initial prediction33. First, a machine learning model such as KRR predicts error predictions, which are produced from the initial prediction33. Then, the predicted errors are added to the outputs of the initial prediction, allowing MCKE to reduce systematic deviations and improve the overall accuracy of KS predictions. Then, the predicted errors are added to the outputs of the initial prediction, allowing MCKE to reduce systematic deviations and improve the overall accuracy of KS predictions. Equations 12 and 13 mathematically define this two-step process.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E_p}={Y_{obs}} - {Y_{pre}}$$\end{document}13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y_{final}}={Y_{pre}}+E_{p}^{{pre}}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E_p}$$\end{document}: The error sequence between observed (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y_{obs}}$$\end{document}) outputs and predicted (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y_{pre}}$$\end{document}) outputs, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y_{final}}$$\end{document}: The final model outputs, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{p}^{{pre}}$$\end{document}: The error predicted by the MCKE model. Development of the MCKE model To accurately predict Ks, our study develops the MCKE model, which is implemented as follows: MGRO generates an initial population of grasshoppers.The positions of grasshoppers, which represent different combinations of KRR and CNN parameters, are randomly initialized.The most important variables are determined using a technique, which is described in the next section.The CNN model is trained on the sequences of most important variables to produce the summary vector.The summary vector is fed into the KRR technique to produce predicted Ks data.The ERC method is executed using Eqs. 12 and 13 to enhance the overall prediction accuracy.MAE is chosen as the fitness function, which is calculated based on the difference between final outputs and observed KS data.The stop condition (e.g., the minimum improvement in prediction error over consecutive iterations) is controlled. If it is not satisfied, the process proceeds to the next step; otherwise, it goes to the step 10.The CNN and KRR parameters are updated using MGTO operator, which include Eqs. 1–8. The process then returns to the step 4.The combination of CNN and KRR parameters that achieves the lowest MAE is selected as the optimal configuration for the MCKE model.The MCKE model is retrained using the complete set of input data to generate the full sequence of Ks values. In Fig. 1, the flowchart of the MCKE model is presented. Moreover, Appendix A includes Fig. 1 A, which shows the role of each component. This appendix also includes Fig. 2A, which shows the advantages of the MCKE model over other hybrid prediction models. MGRO generates an initial population of grasshoppers. MGRO generates an initial population of grasshoppers. The positions of grasshoppers, which represent different combinations of KRR and CNN parameters, are randomly initialized. The positions of grasshoppers, which represent different combinations of KRR and CNN parameters, are randomly initialized. The most important variables are determined using a technique, which is described in the next section. The most important variables are determined using a technique, which is described in the next section. The CNN model is trained on the sequences of most important variables to produce the summary vector. The CNN model is trained on the sequences of most important variables to produce the summary vector. The summary vector is fed into the KRR technique to produce predicted Ks data. The summary vector is fed into the KRR technique to produce predicted Ks data. The ERC method is executed using Eqs. 12 and 13 to enhance the overall prediction accuracy. The ERC method is executed using Eqs. 12 and 13 to enhance the overall prediction accuracy. MAE is chosen as the fitness function, which is calculated based on the difference between final outputs and observed KS data. MAE is chosen as the fitness function, which is calculated based on the difference between final outputs and observed KS data. The stop condition (e.g., the minimum improvement in prediction error over consecutive iterations) is controlled. If it is not satisfied, the process proceeds to the next step; otherwise, it goes to the step 10. The stop condition (e.g., the minimum improvement in prediction error over consecutive iterations) is controlled. If it is not satisfied, the process proceeds to the next step; otherwise, it goes to the step 10. The CNN and KRR parameters are updated using MGTO operator, which include Eqs. 1–8. The process then returns to the step 4. The CNN and KRR parameters are updated using MGTO operator, which include Eqs. 1–8. The process then returns to the step 4. The combination of CNN and KRR parameters that achieves the lowest MAE is selected as the optimal configuration for the MCKE model. The combination of CNN and KRR parameters that achieves the lowest MAE is selected as the optimal configuration for the MCKE model. The MCKE model is retrained using the complete set of input data to generate the full sequence of Ks values. In Fig. 1, the flowchart of the MCKE model is presented. Moreover, Appendix A includes Fig. 1 A, which shows the role of each component. This appendix also includes Fig. 2A, which shows the advantages of the MCKE model over other hybrid prediction models. The MCKE model is retrained using the complete set of input data to generate the full sequence of Ks values. In Fig. 1, the flowchart of the MCKE model is presented. Moreover, Appendix A includes Fig. 1 A, which shows the role of each component. This appendix also includes Fig. 2A, which shows the advantages of the MCKE model over other hybrid prediction models. The training strategy includes the MGRO algorithm, which is executed using an iterative population-based optimization process. At each iteration, a population of candidate solutions (grasshopper positions) is generated, where each solution represents a specific combination of CNN and KRR hyperparameters. The quality of each candidate solution is evaluated using the MAE computed on the validation dataset. The CNN component is optimized using the MGRO algorithm, which searches for the best combination of structural and learning parameters. The tuned hyperparameters include: Number of convolutional layers (e.g., 1–3 layers tested).Number of filters per layer (e.g., 16–128 range).Filter/kernel size (e.g., 2 × 2, 3 × 3, 5 × 5).Activation function (ReLU selected as baseline candidate).Batch size (e.g., 16, 32, 64).Number of epochs (e.g., 50–300 depending on convergence).Dropout rate (0.0–0.5 to prevent overfitting, if applied). Number of convolutional layers (e.g., 1–3 layers tested). Number of convolutional layers (e.g., 1–3 layers tested). Number of filters per layer (e.g., 16–128 range). Number of filters per layer (e.g., 16–128 range). Filter/kernel size (e.g., 2 × 2, 3 × 3, 5 × 5). Filter/kernel size (e.g., 2 × 2, 3 × 3, 5 × 5). Activation function (ReLU selected as baseline candidate). Activation function (ReLU selected as baseline candidate). Batch size (e.g., 16, 32, 64). Batch size (e.g., 16, 32, 64). Number of epochs (e.g., 50–300 depending on convergence). Number of epochs (e.g., 50–300 depending on convergence). Dropout rate (0.0–0.5 to prevent overfitting, if applied). Dropout rate (0.0–0.5 to prevent overfitting, if applied). For the KRR module, MGRO optimizes key model parameters that control regularization and kernel mapping: Regularization coefficient (λ): It typically ranges from 10− 3 to 103.Kernel type: Radial Basis Function (RBF) kernel is selected as default candidate.Kernel parameter: It controls similarity width and nonlinear mapping scale. Regularization coefficient (λ): It typically ranges from 10− 3 to 103. Regularization coefficient (λ): It typically ranges from 10− 3 to 103. Kernel type: Radial Basis Function (RBF) kernel is selected as default candidate. Kernel type: Radial Basis Function (RBF) kernel is selected as default candidate. Kernel parameter: It controls similarity width and nonlinear mapping scale. Kernel parameter: It controls similarity width and nonlinear mapping scale. The proposed MGRO–CNN–KRR–ERC (MCKE) model is implemented on a high-performance computing environment. The configurations of the model include: Intel Core i7.Base clock speed ≥ 2.6 GHz.NVIDIA CUDA-enabled GPU.128 GB DDR4RAM to support large-scale data processing,64-bit operating system (e.g., Windows 11) for better memory management.Programming environment: Python 3.8.Deep learning libraries: TensorFlow/PyTorch.Scientific libraries: NumPy, SciPy, scikit-learn. Intel Core i7. Intel Core i7. Base clock speed ≥ 2.6 GHz. Base clock speed ≥ 2.6 GHz. NVIDIA CUDA-enabled GPU. NVIDIA CUDA-enabled GPU. 128 GB DDR4RAM to support large-scale data processing, 128 GB DDR4RAM to support large-scale data processing, 64-bit operating system (e.g., Windows 11) for better memory management. 64-bit operating system (e.g., Windows 11) for better memory management. Programming environment: Python 3.8. Programming environment: Python 3.8. Deep learning libraries: TensorFlow/PyTorch. Deep learning libraries: TensorFlow/PyTorch. Scientific libraries: NumPy, SciPy, scikit-learn. Scientific libraries: NumPy, SciPy, scikit-learn. Appendix B shows the computer code of the MCKE model while Appendix C explains the line by line of this code (Fig. 1). Fig. 1Flowchart of the MCKE model. Flowchart of the MCKE model. Determination of the most input variables Our study predicts KS using soil bulk density (SBD) and the percentages of clay (CL), sand (SA), and silt (SI) in the soil profile. These four physical properties are selected because of their significant influence on soil texture, which directly affects the soil’s ability to transmit water under saturated conditions. By including these properties, the MCKE model can accurately capture the variability in soil hydraulic behavior, which governs soil–water interactions and determines how water moves in the soil profile. However, these four properties produce 24−1 input combinations, which can increase the computational complexity of the prediction model. These input combinations may also include extreme or uncommon soil property values, which can negatively affect model stability and prediction accuracy. Thus, to address these challenges, it is essential to choose the most important input variables. These variables are defined as soil physical properties that have a significant effect on KS. They also contain critical information about the structure, texture, and porosity of soil, which directly influence its interactions with water. Scholars have developed such as the mutual information (MI) theory to choose the most crucial predictors34. The MI theory is a key tool in reducing model complexity by removing redundant or less informative variables35. The MI theory can identify the most input variables by producing the mutual information index (MII) value which measures the amount of shared information between each predictor and KS34. Equation 14 shows the mathematically formulation for computing this index34.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MII\left( {{X_i},{K_s}} \right)=\sum\limits_{{{x_i} \in {X_i}}} {\sum\limits_{{{k_s} \in {K_s}}} {p\left( {{x_i},{k_s}} \right)} } \log \left( {\frac{{p\left( {{x_i},{k_s}} \right)}}{{p\left( {{x_i}} \right)p\left( {{k_s}} \right)}}} \right)$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_i}$$\end{document}: A specific value of the variable, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_i}$$\end{document}: Th ith predictor, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MII\left( {{X_{i,}}{K_s}} \right)$$\end{document}: The mutual information index (MII) value, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_s}$$\end{document}: The soil saturated hydraulic conductivity, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p\left( {{x_i},{k_s}} \right)$$\end{document}: The joint probability distribution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_s}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p\left( {{x_i}} \right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p\left( {{k_s}} \right)$$\end{document}: The marginal probability distributions of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_s}$$\end{document}, respectively. Notably, soil physical properties with the highest MII values (e.g., MII ≥ 0.8) are selected as the most important input variables. Quantification of the effect of each input variable on model outputs Quantifying the effect of each input variable on the model outputs is essential for understanding the internal behavior of the predictive model36. It is also vital for developing optimal data collection strategies that focus on collecting data for the most influential soil properties, which have the greatest impact on KS predictions. In recent years, various techniques such as Shapley additive explanations (SHAP) have been widely used to quantify the effect of each predictor on the predicted outputs. The SHAP technique is an effective approach for enhancing model interpretability36,37. This technique offers different advantages such as the ability to support scenario simulations, allowing researchers to examine how changes in soil and environmental variables affect KS under various conditions. It can also provide deep insights into the nonlinear dependencies between hydraulic characteristics of soil and environmental variables38. These insights are essential for developing sustainable soil and water management practices, which optimize irrigation efficiency, minimize soil erosion, and maintain soil health. The SHAP technique quantifies the effect of each input variable on the outputs of the prediction models by providing a numerical indicator called the Shapley value. This value represents the dependency between each physical soil property and the predicted KS. Equation 15 shows how SHAP technique provides a Shapley value for each input variable36.15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\phi _i}=\sum\limits_{{S \subseteq N}} {\frac{{\left| S \right|!\left( {\left| N \right| - \left| S \right| - 1} \right)!}}{{\left| N \right|!}}} \left[ {f\left( {S \cup i} \right) - f\left( S \right)} \right]$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\phi _i}$$\end{document} : The Shapley value of input variable, N: The set of all input variables, S: A Subset of without variable i, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f\left( S \right)$$\end{document}: Model outputs based on input variables in the subset S. Notably, a high Shapley value indicates a strong influence of the corresponding soil property on the outputs of the models. Quantifying the uncertainty of the predictions Quantifying the uncertainty of the model outputs is necessary for validating the effectiveness of the prediction models. By quantifying the uncertainty, researchers can provide insights into the sensitivity of the model to input variability and assumptions. Therefore, various methods, such as the kernel density estimation (KDE) technique, have been proposed to determine the uncertainty of predicted Ks values. The KDE technique is an estimation method for approximating the probability density function of a random variable39. This technique can be applied to large and heterogeneous datasets, which are commonly used for predicting Ks. It also provides probabilistic insights into model predictions, which are essential for understanding the likelihood of different outcomes40. Hence, our article utilizes the KDE method to estimate the uncertainty of MCKE outputs. This method is executed through a three step-process, which is explained as follows: First, the MCKE model generates predicted Ks values.Next, prediction errors are obtained by calculating the difference between the estimated and measured Ks values.Finally, the KDE technique computes the probability density function (PDF) of the prediction errors. By estimating the PDF, the method generates confidence intervals that indicate the degree of uncertainty for each predicted Ks value. First, the MCKE model generates predicted Ks values. First, the MCKE model generates predicted Ks values. Next, prediction errors are obtained by calculating the difference between the estimated and measured Ks values. Next, prediction errors are obtained by calculating the difference between the estimated and measured Ks values. Finally, the KDE technique computes the probability density function (PDF) of the prediction errors. By estimating the PDF, the method generates confidence intervals that indicate the degree of uncertainty for each predicted Ks value. Finally, the KDE technique computes the probability density function (PDF) of the prediction errors. By estimating the PDF, the method generates confidence intervals that indicate the degree of uncertainty for each predicted Ks value. The mathematical model of the KDE method includes three equations, which are described as follows:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat {f}=\frac{1}{{Nh}}\sum\limits_{{i=1}}^{N} {K\left( {\frac{{er - e{r_i}}}{h}} \right)}$$\end{document}17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K\left( v \right)=\frac{1}{{\sqrt {2\pi } }}\exp \left( {\frac{{ - {v^2}}}{2}} \right)$$\end{document}18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v=\frac{{er - e{r_i}}}{h}$$\end{document} Where N: The total number of error samples, er: A given error value,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat {f}$$\end{document}: The estimated PDF, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e{r_i}$$\end{document}: The ith prediction error, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_i}$$\end{document}: the bandwidth. Comparative models To further validate the performance of the MCKE model, it is compared with several empirical model and machine learning techniques that have been widely utilized for estimating Ks. These models and techniques are explained in the next section. Radial basis function neural network (RBFNN) The RBFNN technique is a machine learning model that consists of three layers: an input layer, a middle layer, and a linear output layer. These three layers enables the model to approximate complex nonlinear relationships between inputs and outputs41. The RBFNN technique offers several advantages such as the capacity to extract hidden patterns in input sequences. The technique also demonstrates fast convergence during training, making it distinct from other neural network models that require longer training times. The RBFNN technique is executed as follows: First, the input layer receives the input matrix and passes it forward to the hidden layer42.Second, the hidden layer, including radial basis activation functions, extract informative features from the input matrix. These features include soil characteristics that influence the movement and retention of water under saturated conditions.Finally, the extracted features are input into the output layer to produce Ks values. First, the input layer receives the input matrix and passes it forward to the hidden layer42. First, the input layer receives the input matrix and passes it forward to the hidden layer42. Second, the hidden layer, including radial basis activation functions, extract informative features from the input matrix. These features include soil characteristics that influence the movement and retention of water under saturated conditions. Second, the hidden layer, including radial basis activation functions, extract informative features from the input matrix. These features include soil characteristics that influence the movement and retention of water under saturated conditions. Finally, the extracted features are input into the output layer to produce Ks values. Finally, the extracted features are input into the output layer to produce Ks values. The parameters of the RBFNN technique include the number of hidden neurons, learning rate, batch size, and epoch number, which should be carefully tuned to achieve the best prediction accuracy. Thus, our study develops the MGRO-RBFNN model, which uses the MGRO technique to optimally tune the parameters of the RBFNN model. Multiple linear regression (MLR) The MLR model is a parametric approach for modeling the relationship between a set of explanatory variables and a response variable43. The model provides different advantages such as the ability to test hypotheses and evaluate the statistical significance of individual predictors44. It can also provide information about trends in the data over time, enabling the detection of temporal shifts or seasonality. To provide the final output, the MLR technique uses Eq. 19, which represents the linear relationship between the dependent variable and all predictors.19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O={\delta _0}+{\delta _1}{U_1}+{\delta _2}{U_2}+...+{\delta _n}{U_n}+\varepsilon$$\end{document} Where \varepsilon: The error term, O: The output variable, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\delta _0}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\delta _1}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\delta _2}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\delta _n}$$\end{document}: Regression coefficients. Notably, the regression coefficients should be carefully estimated to minimize prediction errors. Thus, our paper proposes the MGRO-MLR model, which uses the MGRO technique to systematically optimize the regression coefficients. The model is then applied to predict Ks. Empirical techniques Various empirical equations have been developed to predict Ks based on physical soil properties. Some of these equations are described below: Julia et al.45. 20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_s}=30.84Exp\left( { - 0.144CL} \right)$$\end{document} Julia et al.45. Julia et al.45. Ottoni et al.46. 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${K_s}=51.1 \times {\left( {0.582 - 0.00216SI - 0.00232CL - 0.203BD} \right)^{1.853}}$$\end{document} Ottoni et al.46. Ottoni et al.46. Where CL: The percentage of clay, SI: The percentage of silt, and BD: The bulk density. Evaluation metrics To comprehensively assess the predictive performance of the proposed MCKE model, several statistical metrics are employed. These metrics quantify the accuracy, robustness, and uncertainty of each prediction model. They are explained as follows: Index of agreement (IA): IA is a statistical metric that quantifies the degree of agreement between predicted and observed values. A high IA value shows that the differences between predicted and observed values are relatively small. 22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$IA=1 - \frac{{\sum\limits_{{i=1}}^{n} {{{\left( {{K_{siob}} - {K_{sipr}}} \right)}^2}} }}{{\sum\limits_{{i=1}}^{n} {\left( {\left| {{K_{siob}} - {{\bar {K}}_{sob}}} \right|+\left| {{K_{sipr}} - {{\bar {K}}_{sob}}} \right|} \right)} }}$$\end{document} Index of agreement (IA): IA is a statistical metric that quantifies the degree of agreement between predicted and observed values. A high IA value shows that the differences between predicted and observed values are relatively small. Index of agreement (IA): IA is a statistical metric that quantifies the degree of agreement between predicted and observed values. A high IA value shows that the differences between predicted and observed values are relatively small. Mean absolute percentage error (MAPE): MAPE is a statistical metric that measures the average percentage difference between model outputs and actual observations47. A low MAPE value indicates that the model produces precise estimates. 23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAPE=\frac{1}{n}\sum\limits_{{j=1}}^{n} {\left| {\frac{{{K_{siob}} - {K_{sipr}}}}{{{K_{siob}}}}} \right|}$$\end{document} Mean absolute percentage error (MAPE): MAPE is a statistical metric that measures the average percentage difference between model outputs and actual observations47. A low MAPE value indicates that the model produces precise estimates. Mean absolute percentage error (MAPE): MAPE is a statistical metric that measures the average percentage difference between model outputs and actual observations47. A low MAPE value indicates that the model produces precise estimates. Legates and McCabe Index (LMI): LMI is a performance indicator that offers a normalized measure of agreement between forecasted and measured values. A higher LMI indicates stronger model performance. 24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LMI=1 - \frac{{\sum\limits_{{i=1}}^{n} {\left| {{K_{Siob}} - {K_{Sipr}}} \right|} }}{{\sum\limits_{{i=1}}^{n} {\left| {{K_{siob}} - {{\bar {K}}_{sob}}} \right|} }}$$\end{document} Legates and McCabe Index (LMI): LMI is a performance indicator that offers a normalized measure of agreement between forecasted and measured values. A higher LMI indicates stronger model performance. Legates and McCabe Index (LMI): LMI is a performance indicator that offers a normalized measure of agreement between forecasted and measured values. A higher LMI indicates stronger model performance. Coverage probability of the Prediction interval (CPPI): This index quantifies the fraction of actual observations that lie within the prediction intervals. A lower CPPI value shows higher uncertainty in the predicted outputs. 25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CPPI=\frac{1}{N}\sum\limits_{{i=1}}^{N} {{\psi _i}}$$\end{document} 26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\psi _i}=\left[ \begin{gathered} 1 \leftarrow if\left( {L\left( {{K_s}_{i}} \right)<{K_s}_{i}