SpringerLink

Remote sensing, artificial neural networks, and spatial interpolation methods for modelling soil chemical characteristics - Modeling Earth Systems and Environment

The increase in global population, rapid urbanization, and continuous soil degradation have disrupted the balance between food demand and supply. Precision agriculture (PA) is an effective solution for overcoming this challenge. The first step in the PA was to determine the physicochemical attributes of the soil. This research aimed to predict soil chemical properties using Geographic Information Systems (GIS), Remote Sensing (RS), and artificial neural networks (ANN). This research explored soil chemical properties using classical statistics, Geostatistics, ANN, and Spatial Interpolation (SI). Multiple Linear Regression (MLR), Ordinary Least Square Regression (OLS), Group Method of Data Handling (GMDH) Neural Networks, Inverse Distance Weighting (IDW), and Kriging Interpolation were used to model and predict soil chemical properties. These modeling techniques accurately predicted soil organic matter (SOM) and pH. This study compared and evaluated models based on Landsat-8 and Sentinel-2 to predict soil chemical properties. Among the soil chemical properties, SOM was significantly modeled using MLR, with an accuracy of 0.54 Sentinel and 0.39 Landsat. Landsat band-TIRS2 and Sentinel band-9 significantly defined the SOM. The exploratory regression technique predicts SOM by incorporating more variables, such as Landsat bands 10 and 11 and Sentinel bands 1, 2, 5, 9, and 10. OLS regression predicted SOM with an accuracy of 0.55 and 0.54 for Sentinel and Landsat, respectively. Using GMDH Neural Network, a more accurate model was developed incorporating Sentinel bands 10, 9, and 8 and Landsat bands 5, 10, and 11, with the best accuracy of 0.68%.