Soil moisture simulation using individual versus ensemble soft computing models

Mohammad Zounemat‐Kermani

doi:10.1007/s13762-022-04202-y

2022-05-12

PRE

Mohammad Zounemat‐Kermani

Soudabeh Golestani Kermani *

Meysam Alizamir

Marzieh Fadaee

DOI：10.1007/s13762-022-04202-y

原文链接

原文求助

摘要

En 中文

Soil moisture plays an important role in water distribution among various components of hydrological cycle and energy exchanges between the atmosphere and the earth's surface. Its accurate estimation is necessary for optimal water management in agriculture, environment, and other related fields. This study describes the development and applications of individual machine learning models including artificial neural networks, radial basis function, multi-layer perceptron, multivariate adaptive regression splines, and extreme learning machine as well as the ensemble Bayesian model averaging methodology for computing soil moisture modeling. Eight climatological inputs are used for constructing the models and two distinct scenarios of (i) predicting soil moisture (t) and (ii) forecasting soil moisture (t + 1) are designed based on different feature selection methods, such as the best subset selection and the historical correlation functions. The statistical evaluation shows that in predicting strategy, the Bayesian model averaging had the best (root mean square error = 0.0127 (m(3)/m(3)), mean absolute errors = 0.0092 (m(3)/m(3))) and multivariate adaptive regression splines model had the weakest (root mean square error = 0.0227 (m(3)/m(3)), mean absolute errors = 0.0196 (m(3)/m(3))) performance in the test stage. Also, in the forecasting strategy, the Bayesian model averaging had the best (root mean square error = 0.0023 (m(3)/m(3)), mean absolute errors = 0.00111 (m(3)/m(3))) and radial basis function had the weakest (root mean square error = 0.0022 (m(3)/m(3)), mean absolute errors = 0.00062 (m(3)/m(3)) performance during the testing stage. Overall, the modeling efforts confirm that the Bayesian model averaging optimizes both the predicted and forecasted results.

Keyword:

Artificial neural networks

Forecasting

Machine learning

Predicting