TL;DR
This paper introduces a novel machine learning framework (MC-EBM with API) to understand how to achieve a state where multiple ecosystem services are both highly supplied and balanced. By analyzing data from Fujian Province, China, they identify key environmental and human factors and their critical thresholds that influence these "ecosystem service bundles" (ESBs), offering practical land management insights.
What Problem Does This Paper Address?
Ecosystems provide a multitude of services (like clean water, food, and carbon sequestration) that are crucial for human well-being. However, these services often interact in complex ways, leading to trade-offs where improving one service might degrade another. Achieving a state where multiple services are simultaneously abundant and balanced – known as "high-supply and balanced ecosystem service bundles" (ESBs) – is a critical but poorly understood challenge. Traditional approaches often use linear models, which fail to capture the inherent non-linearity and threshold effects in ecosystems. This paper aims to bridge this gap by developing a method to identify the key drivers of ESB formation and, crucially, the specific ranges (thresholds) of these drivers that promote desirable ESB states.
Methods
The researchers developed a novel framework by integrating two key components:
- Multiclass Explainable Boosting Machine (MC-EBM): This is a machine learning model designed for multi-class classification problems, capable of learning complex, non-linear relationships from data. It uses a boosting technique where successive models learn from the errors of previous ones, iteratively improving predictions.
- Additive Post-Processing for Interpretability (API) algorithm: This algorithm is integrated with the MC-EBM to make its predictions interpretable. API allows for the visualization of how each driving factor influences the likelihood of different ESBs, revealing "shape functions" or contribution curves. This visualization is key to identifying threshold effects.
This combined approach allows them to not only predict ESB formation but also to understand why certain ESBs form and to pinpoint critical thresholds for influencing factors. They also employed a Self-Organizing Map (SOM) to initially cluster ecosystem services into bundles and used a hybrid sampling method (SMOTE and Tomek Links) to address class imbalance in their dataset.
Datasets
The study utilized a variety of datasets for Fujian Province, China, all standardized to a 1 km × 1 km grid for the year 2015:
- Digital Elevation Model (DEM): For topographic information.
- Land Use/Land Cover (LULC) map: To understand land cover types.
- Normalized Difference Vegetation Index (NDVI): A proxy for vegetation health.
- Chinese Vegetation Type map: For detailed vegetation classification.
- Gross Domestic Product (GDP) spatial distribution data: To represent economic activity.
- Population spatial distribution data: To represent human presence and activity.
- Meteorological data: Including temperature, precipitation, and sunshine hours.
- Total grain production data: For prefecture-level cities.
These datasets were sourced from various Chinese data repositories, including the Geographic Spatial Data Cloud Platform, China Meteorological Data Service Center, and the Resource and Environment Science and Data Center.
Key Insights and Ideas
- Non-linearity is Key: Ecosystems don't respond linearly to changes. Small shifts in factors can sometimes have minor effects, while crossing a certain threshold can lead to dramatic changes in ecosystem service bundles.
- Water is a Balancing Act: A higher proportion of water bodies is a strong positive influence, enhancing and balancing multiple ecosystem services. However, there's an optimal range; too much or too little can be detrimental.
- Human Activity is a Double-Edged Sword: While some human activities (like agriculture) are essential for certain services (food production), intensified human activities (indicated by proximity to construction land, high population/GDP density) generally reduce overall service supply and can lead to undesirable ESB states.
- Thresholds Matter for Management: Simply knowing a factor is important isn't enough. Identifying the specific "sweet spots" or optimal ranges for factors like land use proportions (cultivated land, woodland, water bodies) is crucial for effective land management.
- Interpretable AI for Ecology: The MC-EBM-API framework offers a powerful way to move beyond "black box" models in ecology, providing visual insights into complex ecological relationships.
Key Results
- Dominant Drivers Identified: The study identified several key factors influencing ESBs, including the proportion of water bodies, distance to construction land, annual solar radiation, annual precipitation, population density, and GDP density.
- Optimal Thresholds for Ideal ESBs:
- Cultivated land: Best between 16.5% and 36% (not exceeding 59%).
- Woodland: Optimal around 62% for balanced development, with a strong supply around 31.7%-39.6%.
- Water bodies: Optimal range of 3.7% to 11.4% for high supply and ecological diversity.
- Grassland: Optimal between 23.5% and 50%.
- Built-up land: Should not exceed 10.1%.
- Model Performance: The MC-EBM model significantly outperformed multinomial logistic regression (MLR) in accuracy, achieving much higher accuracy rates for all bundle classes and a substantially higher Balanced Accuracy (BACC) score (84.5% vs. 37.1%).
Limitations and Future Work
The authors acknowledge several limitations:
- Single-Year Data: The study relied on data from a single year (2015), which limits the ability to capture long-term temporal dynamics.
- Spatial Scale: The analysis was conducted at a 1 km × 1 km grid scale, and mechanisms might differ at other spatial scales.
Future work could involve:
- Multi-temporal Analysis: Incorporating data from multiple years to understand temporal trends and changes in ESBs.
- Cross-Scale Validation: Applying the framework to different regions and at multiple spatial scales to test its robustness and applicability.
- Further Refinement of Models: Exploring other advanced interpretable machine learning techniques.
Summary updated