Linking structural forest heterogeneity and ecological processes using Sentinel-2 and FAD-based zoning

Area of Science:

• Forest Ecology
• Remote Sensing
• Machine Learning

Background:

• Forest structural heterogeneity impacts ecological functions, but spectral diagnostics for specific zones are limited.
• Understanding these spectral-ecological relationships is crucial for effective forest management and monitoring, especially in fragmented landscapes.

Purpose of the Study:

• To assess if Sentinel-2 vegetation indices can detect ecological variations across distinct forest structural zones.
• To utilize interpretable machine-learning models for analyzing spectral-ecological relationships in fragmented forests.

Main Methods:

• Integrated Sentinel-2 imagery (2016, 2020, 2024) with field data across Foreground Area Density (FAD)-based structural zones.
• Applied correlation and cluster analysis to reduce multicollinearity among 17 vegetation indices.
• Employed Extra Trees (ET) and LightGBM (LGBM) for predictive modeling, retaining ET for interpretation.

Main Results:

• ET and LGBM showed comparable performance, with ET often yielding lower RMSE values.
• Predictive models accurately estimated degradation class, moisture, site type, and stand age across zones.
• Zone-specific spectral-ecological relationships were identified, with rare zones showing higher stress responses.

Conclusions:

• The framework successfully links spectral traits to ecological gradients in fragmented forests using open data.
• NDRE, MCARI, NDMI, and CVI proved valuable for monitoring diverse ecological attributes.
• This approach offers a reproducible method for ecological monitoring in fragmented forest landscapes.