Quantification of sawgrass marsh aboveground biomass in the coastal Everglades using object-based ensemble analysis and Landsat data

• Published in: Remote sensing of environment 2018-01, Vol.204, p.366-379
• Main Authors: Zhang, Caiyun, Denka, Sara, Cooper, Hannah, Mishra, Deepak R.
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial intelligence; Artificial neural networks; Biomass; Carbon cycle; Cladium; Coastal marshes; Correlation coefficient; Correlation coefficients; Data processing; Ensemble analysis for biomass prediction; Landsat; Landsat satellites; Learning algorithms; Learning theory; Machine learning; Machine learning regression algorithms; Mathematical models; Modelling; Neural networks; Object-based biomass modeling; Pixels; Regression analysis; Remote sensing; Sea level rise; Studies; Support vector machines
• ISSN: 0034-4257; 1879-0704
• DOI: 10.1016/j.rse.2017.10.018

Large-scale biomass quantification of sawgrass (Cladium jamaicense) marsh is critical to understand the carbon and energy cycle in the Florida Everglades. There is also a need to monitor biomass changes in the coastal Everglades due to continuing sea level rise. Previous research in biomass estimation of coastal marshes has focused on pixel-based parametric modeling methods. In this study, an object-based ensemble analysis approach was developed to map sawgrass biomass at multiple scales using Landsat data. Four machine learning regression algorithms including Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbor (k-NN), and Artificial Neural Network (ANN) were evaluated and compared to the commonly used Multiple Linear Regression (MLR) method for both live and total sawgrass biomass estimation. A weighted combining scheme was developed to integrate predictions from comparable models for ensemble analysis. Nonparametric machine learning models had better performance than the parametric approach. ANN and SVM produced similar results in live biomass estimation with the correlation coefficient (r) larger than 0.9, while ANN achieved the best result for the total biomass estimation (r=0.94). Sawgrass biomass maps were produced for two harvest seasons in 2014 and 2016 at three detail levels, which successfully revealed the spatial and temporal (seasonal and interannual) sawgrass biomass variations. Ensemble analysis of the ANN and SVM predictions of live sawgrass biomass not only made the estimation more reliable, but also generated an uncertainty map to identify the regions with a robust biomass prediction, as well as challenging areas for biomass quantification. It is concluded that the object-based ensemble analysis is a promising alternative to the commonly used pixel-based biomass modeling techniques. •An object-based ensemble approach is developed for sawgrass biomass modeling.•Model uncertainty analysis is designed from the ensemble approach.•Techniques are developed for wetlands but can be used for other ecosystems.•Machine learning models are powerful in biomass mapping.•Landsat is appropriate for biomass monitoring in the Everglades.