Application of Random Forest in Spatial Prediction and Environmental Modeling: From Theory to Practice

Original Author and Source:

• Original Author/Maintainer: reZa9343
• Source Platform: GitHub
• Original Project Title: Machine-Learning-Based-Spatial-Modeling-Using-Random-Forest
• Original Link: https://github.com/reZa9343/Machine-Learning-Based-Spatial-Modeling-Using-Random-Forest
• Publication Time: August 11, 2026

Core Viewpoint: This article provides an in-depth analysis of the application of the random forest algorithm in spatial prediction and environmental modeling, exploring its principles, advantages, and practical application scenarios, and offers a practical guide for geospatial data analysis.

In the fields of Geographic Information Systems (GIS) and environmental science, spatial prediction has always been a core challenge. Traditional spatial interpolation methods such as Kriging have solid theoretical foundations, but they struggle to handle high-dimensional features, nonlinear relationships, and complex environmental data. With the development of machine learning, random forest has become an important tool for spatial prediction and environmental modeling due to its excellent prediction performance and robustness.

Spatial data has characteristics of spatial autocorrelation, heterogeneity, and multi-scale features. Standard machine learning models need special adjustments and optimizations. As an ensemble learning method, random forest synthesizes prediction results through multiple decision trees, which can effectively capture complex patterns in spatial data. At the same time, it provides feature importance evaluation to help understand the key factors affecting spatial distribution.

The Idea of Ensemble Learning

Random forest belongs to the Bagging method in ensemble learning. Its core is to combine multiple weak learners (decision trees) to build a strong learner. Each tree is trained on a random subset of the original dataset, and only a random subset of features is considered when splitting nodes. This double randomness reduces the risk of overfitting.

Decision Tree Construction and Voting Mechanism

Regression trees are often used in spatial prediction tasks to predict continuous environmental variables (such as soil moisture, pollutant concentration). After training, the prediction result for a new spatial location is the average of the predictions from all trees (for regression) or majority voting (for classification).

Out-of-Bag Error and Model Evaluation

The out-of-bag (OOB) error of random forest uses about one-third of the data not involved in training as a validation set. It has a built-in cross-validation mechanism, which can estimate generalization performance without an additional validation set, making it suitable for scenarios with limited spatial data.

Design of Spatial Features

Feature engineering is crucial in spatial prediction, and features reflecting spatial relationships need to be constructed:

• Coordinate features: Latitude and longitude or projected coordinates
• Terrain features: Elevation, slope, aspect, curvature, etc.
• Distance features: Distance to rivers, roads, city centers
• Neighborhood statistics: Mean, standard deviation of surrounding areas, etc.
• Remote sensing indices: NDVI, NDWI, etc.

Handling Spatial Autocorrelation

Spatial autocorrelation leads to similar observations at nearby locations. It is necessary to avoid data leakage (e.g., do not use the neighborhood average of the target variable as a feature) and only use features that are independent in time or space.

Feature Importance Analysis

Random forest evaluates feature importance by calculating the contribution of features to reducing the OOB error, helping to identify key environmental factors affecting the spatial distribution of the target variable.

Spatial Cross-Validation

Traditional random cross-validation assumes that samples are independent, but spatial data violates this assumption. Spatial cross-validation is needed to ensure spatial separation between the training set and test set, which truly reflects generalization ability.

Hyperparameter Tuning

Main hyperparameters include:

• Number of trees (n_estimators): More trees are better, but marginal benefits decrease
• Maximum depth (max_depth): Controls complexity to prevent overfitting
• Minimum samples split (min_samples_split): Threshold for node splitting
• Maximum number of features (max_features): Number of features considered per split Use grid search or random search combined with spatial cross-validation to find optimal parameters.

Uncertainty Quantification

Estimate uncertainty through the variance of predictions among trees, guiding supplementary sampling (increasing observations in high-uncertainty areas) to improve prediction accuracy.

Soil Property Mapping

Using limited sampling points combined with auxiliary data such as DEM, remote sensing images, and geological maps to achieve high-resolution mapping of soil properties (organic matter content, pH value, texture), reducing the cost of traditional surveys.

Species Distribution Modeling

Processing species presence/absence data, integrating variables such as climate, terrain, and land use to generate suitability distribution maps, supporting biodiversity conservation and invasive species monitoring.

Air Quality Prediction

Fusing ground monitoring stations, meteorological, and remote sensing inversion data to build prediction models for PM2.5 and O3 concentrations, better capturing the complex interactions between pollution sources, meteorology, and geographic factors.

Geological Disaster Risk Assessment

Integrating multi-source data such as topography, geology, hydrology, and vegetation to build disaster susceptibility assessment models, identifying high-risk areas to support disaster prevention and mitigation decisions.

Practical Recommendations

• Prioritize data quality: Focus on the spatial representativeness of sampling points, the accuracy and timeliness of auxiliary data, the causal relationship between features and target variables, data preprocessing, and outlier handling.
• Improve interpretability: Use SHAP values (quantify feature contributions), partial dependence plots (show marginal relationships between features and predictions), and surrogate models (approximate complex models with simple ones) to enhance model credibility.

Future Outlook

Combine random forest with deep learning, such as using CNN to automatically extract features from remote sensing images before inputting them into random forest, leveraging the advantages of both.

As a mature and powerful machine learning algorithm, random forest has great potential in spatial prediction and environmental modeling. It can handle high-dimensional features and nonlinear relationships, provide feature importance evaluation and uncertainty quantification, and serve as a powerful tool for geospatial data analysis.

Successful spatial prediction requires combining algorithm knowledge with in-depth understanding of the research area's geographic background, data characteristics, and scientific problems. Technology is a means; solving practical problems and supporting scientific decision-making are the ultimate goals. With the improvement of data acquisition capabilities and the development of computing technology, random forest and its derivative methods will play a more important role in Earth system science.