Predicting reference soil groups using legacy data: A data pruning and Random Forest approach for tropical...

Area of Science:

• Soil Science
• Machine Learning
• Data Science

Background:

• Predicting soil taxonomic classes is difficult due to data irregularities from multiple surveyors.
• Source errors in soil datasets can significantly impact classification accuracy.

Purpose of the Study:

• To investigate if data pruning methods improve the prediction accuracy of minor soil groups.
• To compare data pruning with random oversampling for Random Forest classification.

Main Methods:

• Applied data pruning to create subsets of Plinthosols data (80% and 90% core range).
• Utilized Random Forest (RF) with recursive feature elimination.
• Compared pruned datasets against the entire dataset and random oversampling.

Main Results:

• The best prediction accuracy was achieved using RF with recursive feature elimination on the 90% core range, non-oversampled dataset.
• This model showed substantial agreement (kappa=0.57) and increased prediction accuracy for smaller soil groups by 7% to 35%.
• Wetness index was identified as the primary factor influencing soil groups in the Dano catchment.

Conclusions:

• Data pruning, specifically using a 90% core range dataset, effectively reduces source errors and enhances soil group prediction.
• Random Forest with recursive feature elimination is a robust method for improving classification accuracy in imbalanced soil datasets.
• Soil moisture distribution, indicated by the wetness index, is a key determinant of reference soil group distribution.