Machine-learning-based quantitative estimation of soil organic carbon content by VIS/NIR spectroscopy

Soil organic carbon (SOC) is an important soil property that has profound impact on soil quality and plant growth. With 140 soil samples collected from Ebinur Lake Wetland National Nature Reserve, Xinjiang Uyghur Autonomous Region of China, this research evaluated the feasibility of visible/near infrared (VIS/NIR) spectroscopy data (350–2,500 nm) and simulated EO-1 Hyperion data to estimate SOC in arid wetland regions. Three machine learning algorithms including Ant Colony Optimization-interval Partial Least Squares (ACO-iPLS), Recursive Feature Elimination-Support Vector Machine (RF-SVM), and Random Forest (RF) were employed to select spectral features and further estimate SOC. Results indicated that the feature wavelengths pertaining to SOC were mainly within the ranges of 745–910 nm and 1,911–2,254 nm. The combination of RF-SVM and first derivative pre-processing produced the highest estimation accuracy with the optimal values of R(t) (correlation coefficient of testing set), RMSE(t) and RPD of 0.91, 0.27% and 2.41, respectively. The simulated EO-1 Hyperion data combined with Support Vector Machine (SVM) based recursive feature elimination algorithm produced the most accurate estimate of SOC content. For the testing set, R(t) was 0.79, RMSE(t) was 0.19%, and RPD was 1.61. This practice provides an efficient, low-cost approach with potentially high accuracy to estimate SOC contents and hence supports better management and protection strategies for desert wetland ecosystems.