Estimating Soil Available Phosphorus Content through Coupled Wavelet–Data-Driven Models

Abstract

Soil phosphorus (P) is a vital but limited element which is usually leached from the soil via the drainage process. Soil phosphorus as a soluble substance can be delivered through agricultural fields by runoff or soil loss. It is one of the most essential nutrients that affect the sustainability of crops as well as the energy transfer for living organisms. Therefore, an accurate simulation of soil phosphorus, which is considered as a point source pollutant in elevated contents, must be performed. Considering a crucial issue for a sustainable soil and water management, an effective soil phosphorus assessment in the current research was conducted with the aim of examining the capability of five different wavelet-based data-driven models: gene expression programming (GEP), neural networks (NN), random forest (RF), multivariate adaptive regression spline (MARS), and support vector machine (SVM) in modeling soil phosphorus (P). In order to achieve this goal, several parameters, including soil pH, organic carbon (OC), clay content, and soil P data, were collected from different regions of the Neyshabur plain, Khorasan-e-Razavi Province (Northeast Iran). First, a discrete wavelet transform (DWT) was applied to the pH, OC, and clay as the inputs and their subcomponents were utilized in the applied data-driven techniques. Statistical Gamma test was also used for identifying which effective soil parameter is able to influence soil P. The applied methods were assessed through 10-fold cross-validation scenarios. Our results demonstrated that the wavelet–GEP (WGEP) model outperformed the other models with respect to various validations, such as correlation coefficient (R), scatter index (SI), and Nash–Sutcliffe coefficient (NS) criteria. The GEP model improved the accuracy of the MARS, RF, SVM, and NN models with respect to SI-NS (By comparing the SI values of the GEP model with other models namely MARS, RF, SVM, and NN, the outputs of GEP showed more accuracy by 35%, 30%, 40%, 50%, respectively. Similarly, the results of the GEP outperformed the other models by 3.1%, 2.3%, 4.3%, and 7.6%, comparing their NS values.) by 35%-3.1%, 30%-2.3%, 40%-4.3%, and 50%-7.6%, respectively.