Improved multivariate modeling for soil organic matter content estimation using hyperspectral indexes and...

Area of Science:

• Soil Science
• Remote Sensing
• Data Analysis

Background:

• Soil organic matter (SOM) is crucial for soil fertility and agricultural productivity.
• Hyperspectral data offers rich information but requires processing to reduce redundancy and enhance prediction accuracy.
• Spectral indices and characteristic bands are common methods for feature extraction in hyperspectral analysis.

Purpose of the Study:

• To compare the effectiveness of spectral indices (SI) and characteristic bands selected by CARS in improving soil organic matter (SOM) prediction models.
• To evaluate different spectral transformation techniques and machine learning algorithms for SOM prediction.
• To identify the optimal model for accurate SOM estimation using hyperspectral data.

Main Methods:

• Collected 178 topsoil samples and measured visible and near-infrared (VNIR) reflectance spectra.
• Applied spectral transformations (LR, CR, FDR) and calculated optimal spectral indices.
• Selected characteristic bands using the CARS algorithm.
• Developed SOM prediction models using Random Forest, SVR, DNN, and PLSR with both SI and CARS features.

Main Results:

• SI-based models accurately predicted SOM (R2: 0.80-0.87, RPD: 2.14-2.52).
• CARS-based models generally showed higher accuracy, with the CARS-CR-SVR model achieving the best performance (R2: 0.92, RMSE: 1.91 g/kg, RPD: 3.23).
• Model accuracy varied significantly with spectral transformations, with CARS-based models outperforming SI-based models overall.

Conclusions:

• Characteristic band selection using CARS is more effective than spectral indices for enhancing SOM prediction accuracy from hyperspectral data.
• The CARS-CR-SVR model demonstrates superior performance for SOM estimation.
• Spectral transformation and modeling methods significantly influence prediction accuracy, highlighting the importance of careful data processing and model selection.