Document Type : Research Paper
Authors
1 Assistant professor, Department of Water Science and Engineering, Faculty of Agriculture, Shahid Bahonar University of Kerman, Kerman, Iran
2 PhD Student, Department of Water Science and Engineering, Ke.C., Islamic Azad University of Kerman, Kerman
Abstract
Keywords
Main Subjects
The increasing strain on water resources, resulting from climate change and rising demand for food production, has posed significant challenges to the transition toward sustainable agriculture. Consequently, the use of traditional irrigation methods such as surface irrigation—which are often associated with low efficiency and substantial water loss—is no longer justifiable. This has led to a growing interest among farmers in pressurized irrigation systems, such as drip irrigation, due to its precise control over applied water volume and improved water use efficiency. However, the performance of drip irrigation systems is substantially enhanced when accurate information regarding the dimensions of the soil moisture distribution pattern is available. Numerous analytical, numerical, and empirical models have been developed to estimate the infiltration depth and surface wetted width of the moisture distribution pattern. Nevertheless, these models
have certain limitations, including the requirement for a wide range of input variables, the definition of complex boundary conditions, or the need for recalibration. With the emergence and advancement of machine learning models, robust algorithms have been introduced that can provide accurate estimations of the moisture distribution pattern dimensions without the aforementioned constraints. Therefore, the present study evaluates the accuracy of five machine learning models for estimating the infiltration depth and surface wetted width of the soil moisture distribution pattern over time under saline water application in surface drip irrigation. As the first step, a laboratory setup was constructed to record the aforementioned characteristics.
To investigate the soil moisture distribution pattern, a physical model was constructed in the Irrigation Laboratory of Shahid Bahonar University of Kerman. The setup consisted of several components including a water supply source, a soil reservoir, and a surface drip irrigation system. The physical model comprised three separate compartments, with dimensions of 0.8 × 0.8 × 1 m. A transparent Plexiglas sheet was installed on the front side of each compartment, gridded into 5 × 5 cm squares to facilitate visual observation and measurement. Each compartment was filled layer by layer with soil samples passed through a 2 mm sieve, and each layer was compacted using 15 blows of a 5 kg weight. A surface drip emitter with 4 L/h discharge rate was placed on the soil surface near the transparent Plexiglas wall so that the infiltration depth and surface wetted width could be clearly observed and measured. Five salinity treatments with electrical conductivities (EC) of 1.5, 3, 6, 9, and 12 dS. m⁻¹ were applied to examine the moisture distribution pattern. The infiltration depth and surface wetted width were recorded at 10 time intervals of 30 minutes, with three replications. The collected data from two soil textures were used to simulate the dimensions of the moisture distribution pattern and to evaluate the accuracy of selected machine learning models.
RF: Random Forest is a supervised ensemble learning model that consists of a large number of decision trees. The model is developed using the bagging technique combined with random feature selection at each splitting node. This process generates multiple independent decision trees, and the final output is determined by majority voting in classification problems or by averaging the predictions in regression tasks.
XGBoost: eXtreme Gradient Boosting is an efficient and scalable implementation of the Gradient Boosting framework based on decision trees. The algorithm builds trees in a sequential (additive) manner, where each successive tree is trained to correct the residual errors of the preceding ones. With using advanced regularization, gradient-based optimization and parallel processing, XGBoost delivers high accuracy and strong generalization performance.
CatBoost: Categorical Boosting is a gradient boosting algorithm built on decision trees that is specifically engineered for the direct handling of categorical features without requiring preprocessing. The algorithm utilizes Ordered Boosting and Ordered Target Statistics to minimize target leakage, control overfitting, and deliver superior performance in both accuracy and training speed.
SVR: Support Vector Regression (SVR) is a supervised learning algorithm developed for regression problems. It aims to find a function that deviates from the actual observed values by no more than a specified margin , while maintaining the flattest possible function. For nonlinear relationships, SVR employs kernel functions (such as the Radial Basis Function - RBF kernel) to implicitly map the input data into a higher-dimensional feature space. This transformation allows the algorithm to perform linear regression in the new space, which corresponds to nonlinear regression in the original input space.
Elnet: Elastic Net is a regularized linear regression model that integrates the L1 (Lasso) and L2 (Ridge) penalties. The model minimizes an objective function consisting of the residual sum of squares augmented by a convex combination of L1 and L2 regularization terms. This hybrid regularization enables both automatic feature selection (sparsity) and the handling of multicollinearity through the grouping effect, making Elastic Net highly effective for high-dimensional data and datasets with strongly correlated predictors.
The implementation of the aforementioned machine learning models was carried out in Python 3.13. The input variables included sand percentage, silt percentage, irrigation water salinity, cumulative time, cumulative water volume, bulk density, and saturated hydraulic conductivity. These inputs were used to predict the surface wetted width (width) and the infiltration depth (depth) of the soil moisture distribution pattern.
The dataset was normalized using the min-max normalization method. To improve model generalization and prevent data leakage, shuffled stratified k-fold cross-validation was employed. The performance of the models was evaluated using the statistical indices coefficient of determination (R²), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE).
The analysis of the obtained results confirms the acceptable accuracy of the tree-based models in estimating the depth and surface width of the soil moisture distribution pattern in both soil textures. The most accurate estimation of depth was achieved by the CatBoost model in sandy clay soil (R² = 0.94, RMSE = 0.27 cm, MAE = 0.10 cm) and in loamy sand soil (R² = 0.92, RMSE = 1.62 cm, MAE = 0.77 cm). The weakest performance in estimating depth in both soil textures was observed with the Elnet model. Similarly, the most accurate estimation of surface width was obtained by the CatBoost model in sandy clay (R² = 0.98, RMSE = 1.58 cm, MAE = 0.91 cm) and in loamy sand (R² = 0.97, RMSE = 1.76 cm, MAE = 1.06 cm). The weakest results for width prediction in both textures were also recorded by the Elnet model.
In the present study, the performance of five machine learning models including CatBoost, XGBoost, RF, SVR and Elnet was evaluated for estimating the infiltration depth and surface width of the soil moisture distribution pattern in two different soil textures under surface drip irrigation. The results indicated that in both soil textures, tree-based models exhibited higher accuracy compared to the other models in estimating the aforementioned variables. Among them, CatBoost provided the most accurate estimates of both infiltration depth and surface width across both soil textures, often showing close agreement with the results obtained from the XGBoost model. This finding suggests that these models are more capable of capturing complex and hidden patterns among the variables.
In contrast, the presence of nonlinear relationships between input and target variables in simulating soil moisture distribution patterns, along with significant interactions among variables such as water volume, irrigation duration and soil properties led to weaker performance in models such as Elnet regression, particularly in soils with high sand content.
Furthermore, comparison of the results revealed that in soils with higher clay content, water movement in both vertical (depth) and horizontal (surface width) directions is more regulated and system variations occur in a smoother and more continuous manner. Consequently, the relationships among variables are more structured, resulting in more stable model performance for both target variables.
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Conceptualization, Golestani Kermani S.; methodology, Golestani Kermani S. and Saeidi Abbasabad M.; software, Golestani Kermani S.; validation, Golestani Kermani S.; writing—original draft preparation, Golestani Kermani S. and Saeidi Abbasabad M; writing—review and editing, Golestani Kermani S and Saeidi Abbasabad M. All authors have read and agreed to the published version of the manuscript.
Statement: During the preparation of this work the author(s) used chatgpt in order to translate extended abstract to English. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication.
Data available on request from the authors.
The authors avoided data fabrication, falsification, and plagiarism, and any form of misconduct.
The authors declare no conflict of interest.
)