Mapping soil organic carbon (SOC) with high precision is useful for controlling soil fertility and comprehending the global carbon cycle. Low-relief locations are characterized by minimal variability in traditional soil-forming elements, such as terrain and climatic conditions, which make it difficult to reflect the spatial variation of soil properties. In the meantime, vegetation cover makes it more difficult to obtain direct knowledge about agricultural soil. Crop growth and biomass are reflected by the normalized difference vegetation index (NDVI), a significant indicator. Rather than using conventional soil-forming variables.
In this study, a novel model for predicting SOC was developed using Landsat-8 Operational Land Imager (OLI) band data (Blue (B), Green (G), Red (R), and Near Infrared (NIR), NDVI data as the supporting variables, and Artificial Neural Networks (ANNs). A total of 120 surface soil samples were collected at a depth of 25 cm in the northeastern Nile Delta near Damietta City. Of these, 80% (96 samples) were randomly selected for model training, while the remaining 24 samples were used for testing and validation. Additionally, Gaussian Process Regression (GPR) models were trained to estimate SOC levels using the Matern 5/2 kernel within the Regression Learner framework.
The results demonstrate that both the ANN with a multilayer feedforward network and the GPR model offer effective frameworks for SOC prediction. The ANN achieved an R2 value of 0.84, while the GPR model with the Matern 5/2 kernel achieved a higher R2 value of 0.89. These findings, supported by visual and statistical evaluations through cross-validation, confirm the reliability and accuracy of the models.
The systematic application of GPR within the Regression Learner framework provides a robust tool for SOC prediction, contributing to sustainable soil management and agricultural practices.