📄 earth_science

Comparative Assessment of Feature Selection Methods for Machine Learning-Based Soil pH Prediction

This study demonstrates that the effectiveness of feature selection methods for predicting soil pH is highly dependent on the chosen machine learning algorithm, with the combination of XGBoost and simulated annealing (SA-FS1) emerging as the optimal strategy for balancing accuracy, stability, and transferability.

Original authors: Mir Nasser Navidi, Shahrokh Fatehi, Ayeh Lotfollahi, Ahmad Aliyari Boroujeni

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mir Nasser Navidi, Shahrokh Fatehi, Ayeh Lotfollahi, Ahmad Aliyari Boroujeni

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to bake the perfect loaf of bread, but instead of flour and water, your ingredients are 106 different environmental clues about the soil: how steep the hill is, what kind of rocks are underneath, how much sunlight hits the ground, and how green the plants are. Your goal? To predict the soil's "personality," known as its pH (a measure of how acidic or alkaline the soil is).

The problem is, if you throw all 106 ingredients into the mixing bowl at once, the recipe gets messy. Some ingredients are just duplicates of each other, and others might be red herrings that confuse the baker. This is where Feature Selection comes in. It's the art of picking the best handful of ingredients to get the perfect loaf.

But here's the twist: not every baker (machine learning algorithm) likes the same recipe.

The Great Ingredient Hunt

The researchers in this study acted like a team of detectives testing six different ways to pick the best ingredients. They tried:

VIF: A strict librarian who throws out any ingredient that sounds too similar to another (removing duplicates).
RFE: A sculptor who starts with the whole block of stone and chips away the least important parts one by one.
Simulated Annealing (SA): A clever explorer who wanders around the ingredient shelf, sometimes taking a step backward to find a better path forward, avoiding dead ends.
Genetic Algorithm (GA): A digital evolution lab that mixes and matches ingredient groups, keeping the "fittest" combinations and letting the weak ones die out.

They then tested these ingredient lists against four different "bakers" (Machine Learning models): SVM, Random Forest (RF), Cubist, and XGBoost.

The Big Discovery: One Size Does Not Fit All

The main finding is a bit like discovering that a Ferrari engine doesn't run well on diesel, and a diesel truck engine doesn't run well on high-octane racing fuel. The best way to pick ingredients depends entirely on who is doing the baking.

The "Over-Thinkers" (SVM and Cubist): These models were very sensitive. When the researchers used strict methods like the "Librarian" (VIF) or the "Sculptor" (RFE) to pick ingredients, these models got confused. They memorized the training data too perfectly (a problem called overfitting) and failed miserably when tested on new data. It's like a student who memorizes the practice test answers but fails the real exam because the questions were slightly different.
The "Adaptable" Bakers (Random Forest and XGBoost): These models were more robust. They could handle a wider variety of ingredient lists. However, they still shined brightest when paired with the explorers (Simulated Annealing and Genetic Algorithms).

The Winning Combo

After running the simulations, the researchers found a clear champion. The XGBoost model, when fed the specific list of ingredients found by the Simulated Annealing (SA-FS1) method, produced the most reliable results.

The Score: This combo achieved a validation R² of 0.62 and a concordance of 0.74.
What this means: In the world of soil prediction, this is a solid performance. It suggests the model can generalize well to new areas, unlike the other combinations which often stumbled when faced with fresh data.

What the Paper Rules Out

The authors explicitly warn against the idea that there is a "magic bullet" or a single best way to pick ingredients for everyone.

No Universal Winner: They argue against the idea that one feature selection method (like VIF or RFE) is always the best. In fact, for complex models like XGBoost, the strict "Librarian" (VIF) actually hurt performance by removing useful, non-redundant information.
No Free Lunch: Simply throwing more variables at the problem doesn't help. The study showed that sometimes picking fewer variables (like the 5 selected by SA-FS1) works better than picking a huge list (like the 24 selected by SA-FS2), provided those few are the right ones.

How Sure Are We?

The paper is very confident in its measurements but cautious about generalization.

Measured: The results (like the R² of 0.62 for XGBoost/SA-FS1) are based on actual data collected from 120 soil samples in a specific region in Iran (about 57,850 hectares). The models were rigorously tested using a method called "repeated cross-validation," which is like testing the bread recipe on 10 different batches of dough to ensure it works every time.
Suggested: The authors suggest that this approach (using stochastic methods like Simulated Annealing with XGBoost) is the way to go for digital soil mapping. However, they admit that because their study area was relatively small and lacked climate variables (due to low variability in that specific spot), the findings might need more testing in larger, more diverse regions before we declare it a global solution.

The Takeaway

If you want to predict soil pH accurately, don't just grab a random list of environmental clues. You have to match your ingredient-picking strategy to your baking style. For the powerful XGBoost baker, the best strategy is to let a smart, wandering explorer (Simulated Annealing) find the perfect, small set of ingredients. If you try to force a strict librarian to pick the ingredients for this baker, you might end up with a flat, unappetizing loaf.

Technical Summary: Comparative Assessment of Feature Selection Methods for Machine Learning-Based Soil pH Prediction

Problem Statement
Soil pH is a critical indicator of soil quality, particularly in arid and semi-arid environments, yet its spatial variability is complex and influenced by numerous natural and anthropogenic factors. In Digital Soil Mapping (DSM), the selection of appropriate environmental covariates is a pivotal step for improving prediction accuracy. However, the vast number of available environmental variables often leads to redundancy, multicollinearity, and increased computational costs. While Feature Selection (FS) methods are widely used to identify relevant predictors, their effectiveness is not uniform; it varies significantly depending on the specific machine learning (ML) algorithm employed. There is a lack of comprehensive understanding regarding which FS strategies yield the most robust and transferable models when paired with different ML architectures for soil pH prediction.

Methodology
The study was conducted in a 57,850-hectare area in Kermanshah Province, Iran, characterized by calcareous geology and diverse land uses.

Data Collection: 120 soil samples were collected from the 0–30 cm depth using the conditional Latin Hypercube Sampling (cLHS) approach.
Covariates: A total of 106 environmental covariates were generated based on the SCORPAN framework, including topographic attributes, hydrological indices, vegetation indices (derived from Landsat-8 OLI), and geological/chemical variables.
Feature Selection (FS) Strategies: Six distinct FS methods were evaluated:
1. VIF (Variance Inflation Factor): A filter-based method to remove multicollinearity.
2. RFE (Recursive Feature Elimination): A wrapper-based backward elimination method.
3. SA-FS1 & SA-FS2: Two variants of Simulated Annealing (stochastic wrapper) producing different subset sizes (5 and 24 variables, respectively).
4. GA-FS1 & GA-FS2: Two variants of Genetic Algorithms (stochastic wrapper) producing different subset sizes (5 and 10 variables, respectively).
Machine Learning Models: Four algorithms were tested in combination with the FS methods:
- Support Vector Machine (SVM)
- Random Forest (RF)
- Cubist (rule-based regression)
- XGBoost (Extreme Gradient Boosting)
Evaluation: Models were trained on 80% of the data and validated on 20%. Performance was assessed using repeated 10-fold cross-validation, with metrics including the coefficient of determination ( $R^2$ ), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Concordance.

Key Results
The study revealed that feature selection performance is strongly algorithm-dependent, with no single FS method proving universally optimal across all models.

Variable Selection Variability: The number of selected predictors varied significantly, ranging from 4 variables (RFE) to 24 variables (SA-FS2). Despite these differences, certain topographic and hydrological variables (e.g., Slope Length, Local Upslope Curvature, Morphometric Protection Index) were consistently selected across multiple methods, indicating their structural importance in pH modeling.
Algorithm Sensitivity:
- SVM and Cubist: These models exhibited high sensitivity to variable selection. They showed strong calibration performance but suffered from significant overfitting and poor generalization during validation, particularly when paired with deterministic methods like VIF and RFE.
- Random Forest (RF): Demonstrated robustness to changes in predictor sets. While it achieved high calibration accuracy, its validation performance was moderate, showing a better balance between accuracy and generalizability, especially with VIF and SA-FS2.
- XGBoost: Showed the most consistent behavior and the best generalization capabilities.
Optimal Combination: The combination of XGBoost with the SA-FS1 method (selecting 5 variables) yielded the highest validation performance ( $R^2 = 0.62$ , Concordance = 0.74). This combination offered the optimal balance between prediction accuracy and model transferability.
Stochastic vs. Deterministic: Stochastic optimization methods (Simulated Annealing and Genetic Algorithms) generally outperformed deterministic methods (VIF, RFE) when coupled with complex, non-linear models like RF and XGBoost. Deterministic methods tended to oversimplify the feature space, potentially removing complementary information necessary for boosting algorithms.

Significance and Claims
The authors claim that this study provides a critical framework for understanding the interaction between feature selection strategies and machine learning algorithms in digital soil mapping. The primary significance lies in the finding that feature selection is not an algorithm-independent process. The effectiveness of a specific FS method is contingent upon the learning behavior and complexity of the predictive algorithm.

The paper concludes that for soil pH prediction in high-dimensional environmental datasets:

Stochastic search-based methods (SA and GA) are generally more compatible with complex ensemble models (XGBoost, RF) than deterministic filter or wrapper methods.
A "one-size-fits-all" feature selection strategy does not exist; tailored FS strategies are required to enhance the reliability, interpretability, and transferability of soil property predictions.
The XGBoost model, when paired with a carefully tuned stochastic feature selection method (specifically SA-FS1), represents a robust approach for achieving accurate and generalizable soil pH maps.

The authors acknowledge limitations, including the relatively small spatial extent of the study area and the exclusion of climatic variables due to low spatial variability, which may affect the scalability of these findings to larger or more heterogeneous regions. They suggest future work should explore hybrid FS strategies and larger, climatically diverse study areas.