Ecological prediction at macroscales using big data: Does sampling design matter?

Area of Science:

• Ecology
• Environmental Science
• Data Science

Background:

• Ecosystems respond to global change at macroscales.
• Ecosystem models often use data from limited, targeted monitoring sites.
• Sampling strategy's impact on model prediction is not fully understood.

Purpose of the Study:

• To investigate how different sampling strategies influence the predictive performance of ecosystem models.
• To assess the effectiveness of random, stratified random, and targeted sampling designs.
• To understand the implications for macroscale ecological predictions.

Main Methods:

• Subsampled a large dataset of 6,784 lakes across 1.8 million km².
• Mimicked three common sampling strategies: random, stratified random, and targeted.
• Estimated and compared model predictive performance for each strategy.

Main Results:

• Stratified random sampling did not outperform simple random sampling.
• Targeted sampling scenarios showed varied results; one performed poorly, others were similar to random sampling.
• Model predictive performance was influenced by the sampling strategy employed.

Conclusions:

• While some targeted sampling may introduce bias, it doesn't always degrade model performance.
• Compiling spatially extensive, existing datasets can produce robust models.
• Effective models can inform science and policy for global change issues.