Forecast Coral Bleaching by Machine Learnings of Remotely Sensed Geospatial Data

Abstract

With the rapid changes in Earth climates, coral bleaching has been spreading worldwide and getting much severe. It is considered an imminent threat to marine animals as well as causing adverse impacts on fisheries and tourisms. Environmental agencies in affected regions have been made aware of the problem and hence starting to contain coral bleaching. Thus far, they often rely on conventional site survey to determine suitable sites to intervene and commence coral reef reviving process. With the recent advances in remote sensing technology, sea surface temperature (SST), acquired by satellites, has become a viable delegate to coral bleaching. Predicting coral bleaching based solely on SST is limited, as it is only one of many determinants. In addition, areas with different SST levels also exhibit different bleaching characteristics. Hence, area specific models are important for appropriately monitoring the events. Thus far, forecasting the bleaching based on SST alone has limited accuracy, because other disregarded factors are found equally influential. These are turbidity, salinity, and wind speed. Taken into account these geospatial factors, this paper evaluates different machine learning (ML) algorithms, on forecasting coral bleaching levels. Compared with official survey data, it was found that random forest (RF) gave the most accurate results, with accuracy and Kappa of 88.24% and 0.83, respectively. To further assist involved agencies in making data driven solutions to this problem, mapping forecasted by RF were visualized on a web application, implemented with Python and the most recent web frameworks and database systems. The proposed scheme could be extended to modelling coral bleaching in other areas, hence greatly reducing delayed in data acquisition and survey costs.