Stability-Oriented Deep Learning for Hyperspectral Soil Organic Matter Estimation.

Area of Science:

• Soil Science
• Remote Sensing
• Machine Learning

Background:

• Soil organic matter (SOM) is crucial for soil fertility and ecosystem health.
• Hyperspectral technology offers rapid, non-destructive SOM estimation.
• Challenges exist in SOM modeling due to spectral variability and small sample sizes, impacting model stability.

Purpose of the Study:

• To develop a robust deep learning framework for accurate SOM estimation under small-sample conditions.
• To enhance the stability and practical applicability of hyperspectral SOM modeling.
• To address spectral covariance issues and improve predictive performance.

Main Methods:

• Proposed a multi-strategy collaborative deep learning framework (SE-EDCNN-DA-LWGPSO).
• Integrated spectral preprocessing (SG-1DR), data augmentation, dilated convolutions, SE channel attention, and LWGPSO optimization.
• Utilized subtropical red soil samples and SPXY partitioning for rigorous validation through repeated experiments.

Main Results:

• The SG-1DR preprocessing scheme demonstrated superior stability.
• Progressive introduction of framework components (dilated convolution, data augmentation, attention) significantly reduced prediction error fluctuations and performance dispersion.
• The final model achieved high consistency and stability with R² = 0.938 ± 0.010, RMSE = 2.256 ± 0.176 g·kg⁻¹, and RPD = 4.050 ± 0.305.

Conclusions:

• The proposed deep learning framework significantly improves the consistency and numerical stability of hyperspectral SOM estimation under small-sample conditions.
• The integrated multi-strategy approach effectively mitigates challenges posed by spectral variability and limited data.
• This framework holds promise for practical, reliable soil fertility assessment using hyperspectral remote sensing.