Learning-Based Multi-Criteria Decision Making Model for Sawmill Location Problems

This study proposes a Learning-Based Multi-Criteria Decision-Making framework that integrates machine learning and GIS to identify optimal sawmill locations, demonstrating through a Mississippi case study that the Random Forest algorithm effectively identifies 10–11% of the landscape as highly suitable, with the Supply-Demand Ratio being the most influential selection criterion.

The Gist

Imagine you are a master chef trying to open the world's best new restaurant. You know the food has to be great, but the location is the secret ingredient that makes or breaks the business. You need to be close to the farmers who grow your ingredients, near the highway so customers can find you, close to a town for your staff to live, and far away from places that flood or are too steep to build on.

Now, imagine doing this not for one restaurant, but for a sawmill (a factory that turns trees into wood) across an entire state like Mississippi. There are thousands of potential spots, and picking the wrong one could cost millions of dollars.

This paper is about a new, super-smart way to find that perfect spot. Instead of asking a group of experts to guess which factors are most important (which can be biased or subjective), the authors built a digital detective using Artificial Intelligence (AI).

Here is how their "Learning-Based Multi-Criteria Decision-Making" (LB-MCDM) framework works, broken down into simple steps:

1. The Old Way vs. The New Way

• The Old Way (The "Expert Guess"): Traditionally, planners would say, "I think being close to a road is 40% important, and being close to a forest is 30% important." They decide these numbers based on their own opinions. This is like a chef guessing the recipe without tasting the food first. It's risky and can be unfair.
• The New Way (The "AI Detective"): The authors let the data speak. They fed a computer 11,000 different potential locations, along with real-world data about roads, forests, rain, and jobs. They asked the computer: "Based on where successful sawmills already are, what actually matters most?" The computer figured out the recipe for them.

2. The Ingredients (The Data)

The AI looked at a "soup" of ten different ingredients to decide if a spot was good:

• Roads & Railroads: How easy is it to haul logs in and lumber out?
• Urban Areas: Is there a nearby town for workers to live in and buy groceries?
• Slope: Is the ground flat enough to build a factory, or is it a steep mountain?
• Rain: Is it too wet? Too much rain makes logging dangerous and expensive.
• Labor: Are there enough people available to work there?
• The Secret Sauce (Supply-Demand Ratio): This is the paper's biggest innovation. Imagine a pie representing all the wood in a county. If there are already five sawmills eating that pie, there's not much left for a new one. The AI calculates a "Supply-Demand Ratio" to see if a new sawmill would be starving for wood or if there's plenty to go around.

3. The Training Camp (Machine Learning)

The researchers didn't just use one AI; they trained five different "detectives" (algorithms like Random Forest, XGBoost, etc.) on the data.

• They asked: "Which detective is the best at predicting if a spot is 'Highly Suitable' or 'Not Suitable'?"
• The Winner: The Random Forest classifier won the competition. It was the most accurate, getting about 86% of the predictions right.

4. The "Why" (SHAP Analysis)

One of the coolest parts of this paper is that the AI doesn't just give an answer; it explains why. They used a tool called SHAP (which sounds like a friendly robot explaining its logic).

• The Revelation: The AI revealed that the Supply-Demand Ratio (the "Secret Sauce" mentioned above) was the single most important factor. It mattered more than the roads or the rain!
• The Runner-ups: Being close to roads, rail lines, and cities came in second place.
• The Surprises: Surprisingly, the steepness of the land (slope) and the type of trees (land cover) mattered very little in Mississippi because the state is mostly flat and full of forests. The AI learned that in this specific place, those factors weren't the deal-breakers.

5. The Result: A "Heat Map" of Success

Instead of giving a boring list of addresses, the model generated a color-coded map of Mississippi.

• Green/Red Zones: It shows exactly which 10-11% of the state is "Highly Suitable" (the prime real estate) and which areas are a "No-Go."
• Validation: They checked their map against where the actual sawmills are today. They found that 70-80% of existing sawmills are sitting right in the "Highly Suitable" zones the AI predicted. This proves the model works in the real world.

Why This Matters

Think of this framework as a GPS for business decisions.

1. It removes bias: It doesn't care who you know; it cares what the data says.
2. It saves money: It helps companies avoid building factories in bad spots.
3. It's dynamic: If a new sawmill opens or a road gets built, you can update the map instantly, and the AI will re-rank the best spots.

In a nutshell: The authors built a smart computer system that learned from history to tell us exactly where to build the next sawmill. It found that the most important thing isn't just being close to trees, but making sure there's enough wood to go around without too much competition. It's a recipe for success that anyone can follow, not just the experts.

Technical Summary

1. Problem Statement

The selection of optimal locations for sawmills is a complex Multi-Criteria Decision-Making (MCDM) problem. Traditional approaches face two primary limitations:

• Subjectivity: Conventional MCDM methods (e.g., AHP, Fuzzy Logic) rely heavily on expert opinion to assign weights to criteria, introducing bias and reducing replicability.
• Oversimplification: Exact and heuristic optimization models often prioritize proximity measures (transportation costs) while neglecting critical socio-economic and environmental factors like labor availability, market competition, and weather.

The specific challenge addressed is how to objectively determine the relative importance of diverse spatial (GIS) and non-spatial factors to assess and rank candidate sawmill locations without relying on subjective weighting, while handling large datasets efficiently.

2. Methodology: The LB-MCDM Framework

The authors propose a Learning-Based Multi-Criteria Decision-Making (LB-MCDM) framework that integrates Machine Learning (ML), Geographic Information Systems (GIS), and MCDM. The process consists of four phases:

A. Data Collection and Preprocessing

• Data Sources: The study utilizes raster data (land cover, slope, proximity to roads/rails/urban areas) and tabular data (labor statistics, timber supply/demand, market revenue, precipitation).
• Feature Engineering: A novel composite metric, the Supply-Demand Ratio (SDR), was introduced. It calculates the ratio of available timber supply to the total demand (existing mills + a hypothetical new mill) within a 75-mile radius, capturing dynamic market competition.
• Preprocessing: Data was cleaned, missing values imputed, and features scaled. Tabular data was converted to raster format for spatial integration in ArcGIS Pro.

B. Initial Suitability Mapping

• An initial suitability map was generated using a weighted sum overlay in GIS.
• Equal Weighting: To avoid initial bias, all features were assigned equal weights ($w_i = 1/K$).
• Sampling: 11,467 random candidate locations were sampled from this map, labeled with their feature values and an estimated suitability score (target variable).
• Classification: Locations were categorized into four classes: Highly Suitable, Suitable, Somewhat Suitable, and Not Suitable.

C. Feature Weight Tuning and Model Training

• Imbalance Handling: The dataset exhibited class imbalance. The SMOTE-ENN (Synthetic Minority Over-sampling Technique - Edited Nearest Neighbors) algorithm was applied to the training set to balance classes and remove noise.
• Feature Selection: Correlation analysis and Variance Inflation Factor (VIF) tests confirmed no severe multicollinearity. However, Terrain Slope and National Land Cover were removed due to low predictive importance in the specific context of Mississippi (flat terrain, uniform forest cover).
• Model Training: Five ML classifiers were trained on the remaining seven features:
  1. Random Forest (RF)
  2. XGBoost (XGB)
  3. Support Vector Classifier (SVC)
  4. Logistic Regression (LR)
  5. K-Nearest Neighbors (KNN)
• Weight Derivation: SHAP (SHapley Additive exPlanations) values were computed for each model to determine the objective, data-driven importance of each feature. These SHAP values were normalized to serve as new, optimized weights for the GIS layers.

D. Suitability Map Reconstruction and Validation

• The GIS system was re-run using the SHAP-derived weights to generate refined suitability maps.
• Validation: The models were validated against the actual spatial distribution of existing sawmills in Mississippi and through consultation with industry experts.

3. Key Results

Model Performance

• Best Performer: The Random Forest (RF) classifier achieved the highest performance with an Accuracy of 86.48% and an AUC score of 0.9656.
• Other Models: XGBoost (84.95% accuracy) and SVC (81.86% accuracy) performed well. Logistic Regression and KNN showed lower performance.
• Feature Reduction: Removing the two least important features (Slope and Land Cover) actually improved the RF model's performance slightly, confirming the efficiency of the 7-feature model.

Feature Importance (via SHAP)

• Most Influential Factor: The Supply-Demand Ratio (SDR) was identified as the most critical predictor across all models, highlighting the dominance of market dynamics over simple physical proximity.
• Secondary Factors: Road Distance, Rail Line Distance, and Urban Area Distance were the next most significant factors.
• Least Influential: Precipitation and Market Revenue showed the lowest impact in this specific context.

Spatial Analysis

• Suitability Distribution: The model predicts that approximately 10–11% of Mississippi's landscape is "Highly Suitable" for new sawmills.
• Validation: The model successfully validated against reality: 70–80% of existing sawmills are located in areas classified as "Highly Suitable" or "Suitable" by the RF model. Specifically, 74.9% of existing sawmills fell into these top two categories according to the RF model.

4. Key Contributions

Methodological

• Objective Weighting: The framework eliminates subjective expert weighting by using ML to derive feature importance (via SHAP) directly from data.
• Hybrid Integration: It successfully bridges the gap between GIS-based spatial analysis, MCDM, and Machine Learning.
• Novel Metric: The introduction of the Supply-Demand Ratio (SDR) provides a dynamic way to model market competition and resource availability simultaneously.
• Explainability: The use of SHAP analysis ensures the "black box" nature of ML is mitigated, providing transparent insights into why a location is suitable.

Practical

• Dynamic Suitability Maps: Unlike static lists, the model generates dynamic suitability maps that can be updated as new facilities open or close, providing real-time decision support.
• Problem Size Reduction: The framework can filter large candidate sets (thousands of locations) down to a high-potential subset, making subsequent NP-hard optimization problems computationally tractable.
• Replicability: The approach is data-driven and can be adapted to other industrial facility location problems beyond the timber sector.

5. Significance and Managerial Insights

• Context Matters: The study demonstrates that feature importance is context-dependent. While slope and land cover were irrelevant in flat, forest-rich Mississippi, they would likely be critical in mountainous regions.
• Market Dynamics > Proximity: Contrary to traditional logistics models that prioritize distance, this study shows that market competition and supply-demand balance (SDR) are the primary drivers for sawmill success.
• Trust through Transparency: By using SHAP, stakeholders can understand the specific drivers of site suitability, fostering trust in data-driven decisions.
• Investment Potential: The identification of 10-11% of the state as highly suitable offers a clear, data-backed target for investors and policymakers looking to expand the timber industry sustainably.

6. Limitations and Future Work

• Context Specificity: Findings are specific to Mississippi's geography and may not generalize directly to regions with different topographies or forest densities.
• Future Directions: The authors suggest applying the framework to other regions (states/countries), incorporating additional variables (environmental regulations, land costs, water transport), and testing more complex algorithms like Artificial Neural Networks (though this may reduce interpretability).