Soil organic carbon (SOC) content plays a key role in soil biological, chemical and physical properties and its processes. Modeling and prediction of SOC distribution across field and larger regions base on laboratory measurements are costly and time consuming. Prediction SOC using remote sensing data is one of the rapid and cost-effective indirect methods that lead to relative accurate predictions. Most studies have developed Multiple-Linear Regression (MLR) models for predicting SOC from remote sensing data. However multiple non-linear relationships may be increase the accuracy of predictions. The main objective of this study was to develop multiple-non linear models for SOC prediction from remote sensing data by artificial neural networks and compares their predictive performance to MLR models. Thus, MLR and Artificial Neural Network (ANN) models were developed using digital number (DN) and Top –Of- Atmosphere (TOA) reflectance data as inputs, by SPSS and MATLAB software respectively. At first, MLR models were developed and the best of them was determined based on regression coefficient, then inputs of this model were used for neural networks development. Finally both networks and MLR models were subsequently evaluated through validation data, and their predictive performance was compared with each other. Results show neural networks have better performance than MLR models. Moreover it is characterized that TOA reflectance data have higher efficiency than digital number values to predict SOC.
Soil organic carbon; remote sensing; multiple-linear regression; artificial neural networks