Comparisons of regression and machine learning methods for estimating mangrove above-ground biomass using multiple remote sensing data in the red River Estuaries of Vietnam

Nguyen Hong Quang, Claire H. Quinn, Rachael Carrie, Lindsay C. Stringer, Le Thi Van Hue, Christopher R. Hackney, Dao Van Tan

Published in ‘Remote Sensing Applications: Society and Environment’

Abstract

Currently, remote sensing platforms provide state-of-the-art data for multiple purposes including applications related to coastal wetlands. Mangrove above-ground biomass (MAGB) together with its extent is considered well correlated with the habitats’ environmental and economic values. Above-ground biomass can be estimated by models that integrate remote sensing, field data and statistical information. However, it remains difficult to decide which model and which data offer the best performance for any one study location. Hence, this study aims to assess the spatial change in MAGB over a 45-year period and investigate different approaches to quantify this change through linear and multi linear regression models. Specifically, we test a non-linear model (Multivariate Adaptive Regression Splines; MARS), and non-parametric machine learning models, to predict MAGB using vegetation indices and biophysical variables derived from optical remote sensing data from Sentinel-2, Landsat-8, SPOT-7 and synthetic aperture radar remote sensing data from ALOS-2. The multi linear regression (MLR) and the MARS models were trained by field measured MAGB data to a good level of accuracy (R2 = 0.80 and RMSE = 5.56 Mg ha−1 for MLR and R2 = 0.89, RMSE = 5.42 Mg ha−1 for MARS). These models were subsequently applied to Landsat 2, 5 and 8 time-series images to assess changes in MAGB values and mangrove forest extent over the period 1975 to 2020. To ensure accurate training data for the models, we conducted field work to measure MAGB in 24 plots measured in May 2019. Findings showed that the MARS model generated MAGB values with higher accuracy than linear regression and multi linear regression models. Uses of vegetation indices (Normalized Differenced Vegetation Index, Soil-adjusted Vegetation Index, Green-Normalized Differenced Vegetation Index, Simple Ratio, and Red-edge Simple Ratio) generated MAGB values with accuracy slightly higher than using biophysical variables (Leaf area index, Fraction of Absorbed Radiation, Fractional vegetation cover, and Leaf chlorophyll content). Sentinel-2 and Landsat 8 were effective data sources for MAGB estimates, while SPOT-7 and ALOS-2 produced acceptable MAGB accuracy. Modelling the Landsat time series found an increase in both MAGB values and forest extent over the 1975–2020 period. The MARS model, Sentinel-2, Landsat 8 and vegetation indices are the recommended models and data to use to measure MAGB and could be used to understand changes in MAGB and forest extent at national and regional scales.