Ecological Dissimilarity Matters More Than Geographical Distance When Predicting Land Surface Indicators Using...

Area of Science:

• Environmental science
• Machine learning
• Geospatial analysis

Background:

• Supervised machine learning models require extensive in situ data for training to predict Earth's surface conditions.
• Using training data from a different geographic region presents challenges due to potential variations in environmental characteristics.

Purpose of the Study:

• To investigate the conditions under which training data from one ecoregion can be substituted for another in supervised machine learning for Earth surface predictions.
• To determine the role of ecological dissimilarity versus geographic distance in predicting model performance across different ecoregions.

Main Methods:

• Trained machine learning models using in situ data from level IV ecoregions across the western United States.
• Tested model predictive performance on different ecoregions.
• Quantified ecoregion differences using geographical distance (centroid-to-centroid) and ecological dissimilarity (multivariate indicator space and temporal behavior from remote sensing data).

Main Results:

• Prediction error generally increased with geographical distance between training and testing ecoregions.
• Ecological dissimilarity was found to be a significant predictor of the expected error when applying a model trained in one ecoregion to another.
• The study demonstrates that ecological dissimilarity is a key factor in assessing the transferability of machine learning models.

Conclusions:

• Ecological dissimilarity is a more robust metric than geographic distance for predicting the accuracy of Earth surface models trained in different regions.
• Understanding ecological dissimilarity is essential for selecting appropriate training data and ensuring the reliability of machine learning predictions in novel environments.
• This research provides a framework for improving the transferability and accuracy of geospatial machine learning models.