Development support vector machine, artificial neural network and artificial neural network - genetic algorithm hybrid models for estimating erodible fraction of soil to wind erosion

• Published in: International journal of river basin management Vol. 22; no. 3; pp. 379 - 388
• Main Authors: Nouri, Alireza, Esfandiari, Mehrdad, Eftekhari, Kamran, Torkashvand, Ali Mohammadi, Ahmadi, Abbas
• Format: Journal Article
• Language: English
• Published: Abingdon Taylor & Francis 02.07.2024; Taylor & Francis Ltd
• Subjects: Accuracy; administrative management; Algorithms; ANN; Artificial neural networks; Calcium; Calcium carbonate; Calcium carbonates; Carbonates; Clay; EF-hand; erodibility; Estimation; Genetic algorithms; Neural networks; Organic matter; Performance prediction; prediction; sand; Saturated soils; silt; Soil; Soil characteristics; Soil erosion; Soil properties; soil salinity; Soil surfaces; Soil testing; Support vector machines; watersheds; wind; Wind erosion
• ISSN: 1571-5124; 1814-2060; 1814-2060
• DOI: 10.1080/15715124.2022.2153856

The index of Erodible Fraction (EF) component against wind is well known as a criteria of soil wind erodibility. This study was conducted to estimate the index of EF by utilizing support vector machine (SVM), artificial neural network (ANN) and genetic algorithm (GA-ANN) methods. For this reason, 95 soil samples were collected 10 cm above the soil surface of Allahabad plain in Qazvin province. Then the percentage of aggregates with a diameter of less than 0.84 mm, pH, EC, soil saturation capacity, SAR, equivalent calcium carbonate, organic matter, clay, sand and silt were measured and used in SVM, ANN and GA-ANNs model construction and testing. Results revealed that the EF had a significant correlation at the level of 1% with five soil characteristics, including pH, EC, SAR, clay and organic matter. On the other hand, the ANNs model developed for EF prediction had higher accuracy (R 2  =  0.49) than the GA-ANNs model (R 2  = 0.37) and SVMs model (0.02). All models were approximately overestimated and the GMER values for SVMs, ANNs and GA-ANNs models were 1.11, 1.15 and 1.08, respectively. However, according to the Akaike information criterion (AIC), all of them had similar power model estimation. Although SVMs offer good accuracy and perform fast prediction in this study, it did not perform well. This issue may be related to the complexity of the relationship between EF and soil properties.