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

Area of Science:

• Environmental Science
• Remote Sensing
• Soil Science

Background:

• Soil organic carbon (SOC) is crucial for soil quality and plant growth.
• Arid wetland ecosystems require effective monitoring and management strategies.

Purpose of the Study:

• To assess the feasibility of using visible/near infrared (VIS/NIR) spectroscopy and simulated EO-1 Hyperion data for estimating SOC in arid wetland regions.
• To identify optimal machine learning algorithms and spectral features for accurate SOC prediction.

Main Methods:

• Collected 140 soil samples from Ebinur Lake Wetland National Nature Reserve.
• Employed VIS/NIR spectroscopy (350-2,500 nm) and simulated EO-1 Hyperion data.
• Utilized Ant Colony Optimization-interval Partial Least Squares (ACO-iPLS), Recursive Feature Elimination-Support Vector Machine (RF-SVM), and Random Forest (RF) algorithms for feature selection and SOC estimation.

Main Results:

• Key spectral features for SOC estimation were identified in the 745-910 nm and 1,911-2,254 nm ranges.
• The RF-SVM algorithm combined with first derivative pre-processing achieved the highest accuracy (R=0.91, RMSE=0.27%, RPD=2.41).
• Simulated EO-1 Hyperion data with RF-SVM yielded good SOC estimates (R=0.79, RMSE=0.19%, RPD=1.61).

Conclusions:

• VIS/NIR spectroscopy and simulated satellite data offer an efficient and accurate method for estimating SOC in arid wetlands.
• Machine learning approaches enhance the precision of SOC content estimation.
• This research supports improved management and protection of desert wetland ecosystems.