Quantifying mangrove leaf area index from Sentinel-2 imagery using hybrid models and active learning

Area of Science:

• Ecology
• Remote Sensing
• Forestry

Background:

• Mangrove forests provide essential ecosystem services but face increasing threats.
• Monitoring mangrove health requires accurate vegetation characteristic quantification, like Leaf Area Index (LAI).
• Satellite remote sensing offers a scalable solution for monitoring these vital ecosystems.

Purpose of the Study:

• To investigate the potential of radiative transfer models (RTM) combined with active learning (AL) for estimating mangrove LAI.
• To assess the performance of this approach using Sentinel-2 spectral reflectance data.
• To compare the developed method against existing techniques like the Sentinel Application Platform (SNAP) and red-edge NDVI.

Main Methods:

• Utilized radiative transfer models (RTM), specifically PROSAIL, for simulating canopy spectral reflectance.
• Employed active learning (AL) strategies to optimize model training data selection.
• Developed a hybrid Gaussian Processes Regression (GPR) model integrating RTM simulations and AL for LAI estimation.
• Validated LAI estimates against in-situ hemispherical photography measurements.

Main Results:

• The AL-driven hybrid GPR model achieved high accuracy in estimating mangrove LAI (R² = 0.77, RMSE = 0.13 m²/m², NRMSE = 9.57%).
• This approach significantly outperformed the SNAP biophysical processor (R² = 0.44) and red-edge NDVI methods.
• PROSAIL RTM yielded the highest accuracy among the tested canopy RTMs.

Conclusions:

• Physics-based machine learning, particularly AL-driven GPR, offers a robust and scalable method for estimating mangrove LAI from Sentinel-2 data.
• This approach reduces reliance on extensive in-situ data and adapts RTMs to specific ecosystems like mangroves.
• The findings support the potential for retrieving diverse vegetation variables to quantify large-scale mangrove dynamics.