Hyperspectral inversion of soil organic matter based on improved ensemble learning method.

Area of Science:

• Soil Science
• Remote Sensing
• Spectroscopy

Background:

• Soil organic matter (SOM) is critical for soil health and carbon sequestration.
• Hyperspectroscopy offers an efficient and cost-effective method for SOM detection.
• Ensemble Models (EM) show promise in soil spectroscopy but require optimized base learner selection and weighting.

Purpose of the Study:

• To develop innovative base learner weight assignment methods for EM in SOM detection.
• To identify optimal base learner counts for different EM techniques (WA, Blending, Stacking).
• To enhance the performance of EM for quantitative SOM assessment using Vis-NIR spectroscopy.

Main Methods:

• Utilized Vis-NIR spectroscopy on 704 soil samples from the Tarim River Basin.
• Investigated various weight coefficient assignment methods (including R², RMSE, MAE) and base learner counts.
• Evaluated EM performance within Weighted Averaging (WA), Blending, and Stacking frameworks.

Main Results:

• A multi-index weight assignment (R², RMSE, MAE) significantly improved EM performance over traditional methods.
• Optimal base learner count varied by ensemble technique; Stacking and Blending peaked at 12, WA at 15.
• Stacking demonstrated superior precision, achieving R² of 0.889, RMSE of 0.957 g kg⁻¹, and MAE of 0.803 g kg⁻¹.

Conclusions:

• Optimizing base learner count and employing a multi-index comprehensive evaluation for weight assignment are crucial for EM performance.
• The Stacking method with 12 base learners and multi-index weighting is the optimal strategy for SOM hyperspectral inversion.
• This refined approach enhances the accuracy and reliability of soil organic matter detection via hyperspectroscopy.