Multi-scale digital soil mapping with deep learning.

Area of Science:

• Geosciences
• Soil Science
• Machine Learning

Background:

• Traditional multi-scale feature construction in digital soil mapping (DSM) using neighborhood filtering is susceptible to outliers.
• Decomposing elevation data can introduce artifacts, limiting its utility for accurate DSM.

Purpose of the Study:

• To introduce and evaluate 'mixed scaling,' a novel method for multi-scale terrain feature construction.
• To compare the effectiveness of mixed scaling with traditional methods for digital soil mapping using Deep Learning.

Main Methods:

• Developed 'mixed scaling' by extending the Gaussian pyramid with intermediate scales to preserve landscape features.
• Applied mixed scaling and traditional methods to construct terrain attributes for three different datasets.
• Modeled soil data using Deep Learning and Random Forests algorithms.

Main Results:

• Mixed scaling with an extended Gaussian pyramid yielded superior covariates for soil mapping.
• Deep Learning models achieved 4-7% higher prediction accuracy on average compared to Random Forests.
• The new method effectively preserves landscape features identifiable at various scales.

Conclusions:

• Mixed scaling offers a robust approach to multi-scale terrain feature construction, overcoming limitations of existing methods.
• Deep Learning, coupled with mixed scaling, provides highly accurate digital soil mapping predictions.
• The extended Gaussian pyramid ensures crucial soil-formation-related scales are included in the models.