Land-Cover Classification Using MaxEnt: Can We Trust in Model Quality Metrics for Estimating Classification Accuracy?

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Ecological Modeling

Background:

• Maximum Entropy (MaxEnt) is a widely used algorithm for species distribution modeling.
• Its application is expanding to land-cover classification using remote sensing data.
• The transferability of species distribution modeling best practices to land-cover classification remains underexplored.

Purpose of the Study:

• To assess the applicability of MaxEnt model quality metrics for single-class land-cover classification.
• To investigate if recommended procedures for species distribution models yield accurate land-cover classifications.
• To determine the suitability of various accuracy and quality metrics as proxies for land-cover classification outcomes.

Main Methods:

• Utilized MaxEnt for single-class land-cover classification with remote sensing imagery.
• Generated 1980 classification maps for four land cover types: built, grass, deciduous, and evergreen.
• Calculated discrimination accuracy coefficients (e.g., Kappa, Overall Accuracy) and model quality metrics.

Main Results:

• Correlation patterns between model quality metrics were consistent for some land covers but not all.
• The relationship between model quality metrics and land-cover classification accuracy was land-cover-dependent.
• Built cover showed no consistent correlation patterns; grass, evergreen, and deciduous showed consistent associations.

Conclusions:

• Model quality metrics are not universally reliable indicators of land-cover classification accuracy.
• Accuracy evaluation should prioritize discrimination coefficients (e.g., Kappa, Overall Accuracy).
• Over-reliance on model quality metrics for land-cover classification can be misleading, especially for certain land types.