Remote sensing-based prediction of organic carbon in agricultural and natural soils influenced by salt and sand mining using machine learning.

ABSTRACT

It is important to keep soil organic carbon (SOC) in balance to ensure soil health and quality. In this manner, mining activities have crucial impacts on SOC stocks, especially in semi-arid and arid regions such as Iran. For this purpose, SOC was measured at 180 randomly selected points in both natural and agricultural soils in the central part of Iran. Machine learning methods, such as GEP (Genetic Expression Programming), SVR (Support Vector Regression), and ANNs (Artificial Neural Networks), were developed and employed to estimate SOC for all sampled points, including both natural and agricultural soils. Following that, topography and remotely sensed data were employed as input variables to improve SOC prediction influenced by mining. The remotely sensed data and topography factors were extracted from Landsat 9 images and Digital Elevation Models (DEMs), respectively. Input variables were considered in three scenarios, including the use of topography factors (scenario I), the use of remote sensing data (scenario II), and the use of both topography factors and remote sensing data (scenario III). The results of this study showed that the most effective model for predicting SOC across all sampled data was SVR (ME = -0.1539%, R2 = 0.642 and RMSE = 0.620%) when employing scenario III. Furthermore, the results indicated that the optimal method for both natural and agricultural soils was the SVR method when employing scenario III. Further analysis through mapping SOC contents showed that mining activities influenced the distribution of SOC in the studied region. Overall, the predicted maps of SOC contents indicated that lower SOC contents were predominantly distributed in the vicinity of salt and sand mines, particularly in salt-rich areas, for both natural and agricultural soils.