Development of MLR and variedly optimized ANN models for forecasting the detachability and liquefaction potential index of erodible soils

Abstract

• DLPI which quantifies soil erodibility potentials, was computed for southern Anambra. • MLR and four ANNs successfully predicted the DLPI of the study area. • Low modeling errors were obtained in the developed MLR and ANN models. • Soft computing algorithms can enhance soil erosion risk assessment and management. Artificial intelligence modeling techniques, such as artificial neural network (ANN) and multiple linear regression (MLR), have been found useful in the prediction of environmental hazard indicators, even in areas with data scarcity. The detachability and liquefaction potential index ( DLPI ) is a numerical model that quantifies soil erodibility potential. This paper aims to design ANN and MLR models that can predict the DLPI of erodible soils in southeastern Nigeria. Feedforward back-propagated ANN models (ANN1, ANN2, ANN3 and ANN4) optimized using different algorithms and output layer activation functions were proposed. To implement the four models, the geotechnical data of the eroding soils, which were used for the DLPI calculation, were utilized. Sixteen geotechnical variables analyzed on the soils were used as input variables. While ANN1 and ANN3 were optimized using scaled conjugate gradient algorithm, ANN2 and ANN4 were optimized using gradient descent algorithm. However, the output layers of the ANN1 and ANN2 were activated using identity function while those for the ANN3 and ANN4 were activated using hyperbolic tangent function. The results of the four models showed high performance with determination coefficient (R 2 ) values of 0.971 (ANN1), 0.998 (ANN2), 0.987 (ANN3), and 0.992 (ANN4). Further, low modeling errors were detected from the residual plots, sum of square errors (0.000-0.024), and relative errors (0.003-0.080); thus, confirming the reliability of the models. The MLR gave a perfect model for the DLPI prediction, with its multiple correlation coefficient = 1.000, R 2 = 1.000, R 2 adjusted = 1.000, and standard error of estimate = 0.000. Therefore, it suffices to say that the ANN and MLR algorithms are suitable for DLPI prediction.