A novel methodological framework for predicting and mapping agriculture-related soil attributes using Euclidean...

Area of Science:

• Soil science
• Geostatistics
• Machine learning

Background:

• Traditional geostatistical methods like Ordinary Kriging (OK) and Inverse Distance Weighting (IDW) have limitations in predicting soil attributes at fine spatial scales.
• Comparative performance and reliability of various statistical and machine learning (ML) techniques for soil attribute prediction remain unclear.

Purpose of the Study:

• To compare the performance and reliability of Ordinary Kriging (OK), Inverse Distance Weighting (IDW), and various ML algorithms in predicting and spatializing soil attributes.
• To evaluate prediction uncertainty and computational processing time for different modeling approaches.
• To assess the impact of grid size on prediction accuracy and computational cost.

Main Methods:

• Utilized Euclidean distance-based predictors from X-Y coordinates and regular grids (5, 7, 10 divisions) in Minas Gerais State, Brazil.
• Generated soil attribute maps (CEC, phosphorus, sand, clay) using OK, IDW, Random Forest (RF), Cubist, Support Vector Machine (SVM), and Earth algorithms.
• Assessed model performance using R2, Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coefficient of variation.

Main Results:

• ML methods, particularly RF and SVM (R2 = 0.62–0.70), demonstrated superior predictive accuracy compared to IDW and OK (R2 = 0.52–0.58).
• RF achieved the highest accuracy for most attributes, while SVM excelled for sand prediction.
• Processing time was shortest for IDW, followed by OK; among ML models, Earth was fastest, followed by RF, SVM, and Cubist.

Conclusions:

• ML algorithms outperform traditional geostatistical interpolators for soil attribute mapping due to their ability to handle numerous covariates and flexible structures.
• While ML methods require greater computational time, they offer improved prediction accuracy and spatialization, especially with larger grids.
• The findings underscore the robustness and practical potential of ML approaches for precise soil attribute mapping.