Satellite-based Mangrove species abundance estimate using Machine Learning ensemble

Abstract

The mangrove ecosystem is a vital feature in a coastal area, playing a critical role in carbon sequestration beneath the soil. Carbon preservation capacity varies among different species of mangrove. Thus, by quantifying the number of mangrove species in a given area, the volume of carbon sequestered can be estimated. Satellite imagery is highly effective for gathering such data across vast territories. In this study, we present an evaluation of mangrove species abundance across a large coastal area using Landsat satellite imagery. We employed machine learning algorithms to classify species based on spectral field observation data to achieve this. These algorithms were trained individually and ensembled to enhance prediction performance. There are 466 models generated in a two-hour training phase. After assessing these models, we identified that a stacked ensemble consisting of Deep Learning, two Distributed Random Forests, a Generalized Boosting Model, a Generalized Linear Model, and Extreme Gradient Boosting algorithms has the most superior predictive accuracy. The model achieved a mean accuracy value of 95% when tested on observation data. After applying the best model to the satellite data, our results indicate that Rhizophora Apiculata and Excoecaria Agallocha are the two most abundant mangrove species in the study area, covering 17.71% (19502.37 Ha) and 10.49% (11549.79 Ha), respectively.