Modeling satellite-based open water fraction via flexible Beta regression: An application to wetlands in the north-western Pacific coast of Mexico

TL;DR

This paper introduces a flexible Beta regression model with Bayesian estimation for predicting open water fraction in wetlands using satellite imagery, demonstrating improved accuracy over standard models and providing accessible R software.

Contribution

The paper develops a novel Bayesian flexible Beta regression model for open water prediction from satellite data, with detailed estimation algorithms and open-source implementation.

Findings

FBR model outperforms standard models in predicting water fraction.

Bayesian estimation via Metropolis-Hastings and Gibbs sampling is effective.

Software implementation in R is publicly available.

Abstract

Carbon sequestration and water filtering are two examples of the several ecosystem services provided by wetlands. Open water mapping is an effective means to measure any wetland extension as these are comprised of many open water bodies. An economical, though indirect, approach towards mapping open water bodies is through applying geo-computational methods to satellite images. In this work we propose the flexible Beta regression (FBR) model to predict open water fraction from measurements of a water index. We focus on observations derived from two MODIS images acquired during the dry season of 2008 in Marismas Nacionales, a wetland located in the north-western Pacific coast of Mexico. A Bayesian estimation procedure is presented to estimate the FBR model; in particular, we provide details of a nested Metropolis-Hastings and Gibbs sampling algorithm to carry out parameter estimation. Our results show that the FBR model produces valid predictors of water fraction unlike the standard model. Our work is complemented by software developed in the R language and available through a GitHub repository.