Spatial validation reveals poor predictive performance of large-scale ecological mapping models.

Area of Science:

• Ecology
• Remote Sensing
• Forestry
• Geostatistics

Background:

• Accurate mapping of aboveground forest biomass is crucial for global carbon balance assessments.
• Existing large-scale biomass maps exhibit significant discrepancies, despite reported high validation statistics.
• Current validation methods often overlook spatial autocorrelation (SAC) in ecological data.

Purpose of the Study:

• To investigate the impact of spatial autocorrelation on the validation of forest biomass mapping models.
• To demonstrate how standard validation practices can lead to overoptimistic assessments of predictive power.
• To re-evaluate the predictive power of a random forest biomass model using spatial validation techniques.

Main Methods:

• Utilized a large forest inventory dataset (11.8 million trees) from central Africa.
• Trained and validated a random forest model using multispectral and environmental variables.
• Compared standard nonspatial validation with spatial validation methods that account for SAC.

Main Results:

• Nonspatial validation indicated the model explained over 50% of forest biomass variation.
• Spatial validation methods, accounting for SAC, revealed near-zero predictive power for the model.
• The study highlights how ignoring SAC inflates the apparent predictive accuracy of big data mapping models.

Conclusions:

• Common validation practices in large-scale ecological mapping studies may be fundamentally flawed due to ignored spatial autocorrelation.
• Apparent high predictive power can mask poor relationships between predictors and the ecological variable of interest.
• Erroneous forest biomass maps and interpretations may result from inadequate validation, impacting global carbon balance assessments.