A Generalized Machine Learning Classifier for Spatiotemporal Analysis of Coral Reefs in the Red Sea

Despite the increase in available remote sensing data and associated interest in conducting change detection analysis of coral reefs, few studies have been conducted to evaluate the spatial generalization of this data when a machine learning classifier is trained on multiple datasets. This study proposes a method for developing a robust machine learning classifier, trained using data from one location and deployed to predict coral cover in a separate location, for coral reef change detection analysis of multiple locations using Landsat 7 and Landsat 8 imagery. The process of creating a robust classifier for coral detection builds upon previous, in-situ type analysis which are simply focused on change evaluation of a single location. The method proposed in this research expands these limited scope studies to include a more generalized approach that can be applied to multiple sites. The proposed framework includes image calibration, temporal image segmentation and differencing, support vector machine (SVM) training and tuning, statistical assessment of model accuracy, and pixel based change detection. Validation of the method was performed using three northern Red Sea locations known to contain an abundance of reefs. The reefs in this study have never been the subject of a change detection analysis at this spatial or temporal scale. The trained and tuned SVM attained an accuracy of 78% on the ground truth data from the Gulf of Aqaba location. The classifier was then deployed to ground truth data from a second site, Umluj, and attained an accuracy of 73%. A generalized, robust machine learning classifier was then trained using the combined information of both sites and correctly classified 71% of ground truth observations from both sites. This robust classifier was then deployed to identify ground cover in a third location, Al Waih. Change detection analysis was then conducted for all three locations. The analysis showed a decrease in coral cover between 2000 and 2018 of 11.4%, 3.4%, and 13.6% in the Gulf of Aqaba, Umluj, and Al Waih locations, respectively.

This study leveraged a robust machine learning classifier trained using the combined information of two large areas of interest in order to conduct a change analysis of the coral cover in the Red Sea. The spatial and temporal extent undertaken by this study has never been accomplished before. This research builds upon the previous in-situ methodology to produce a classifier that is robust to location specific biases. While removing these localized biases reduces the performance within the training location, it enables the classifier to generalize more effectively to new locations. In this way, the generalized classifier is robust to overfitting site specific conditions as is the case with an in-situ approach. In addition to expanding the boundaries of spatial constraints, this research aims to evaluate a longer time period than previous research. Very little research has been conducted to evaluate the change in coral reefs over an 18 year period using remote sensing data. Furthermore, given the frequency and severity of coral bleaching events over the time period, this research is a critical check point for identifying how coral reefs in the Red Sea are surviving under the threat of climate change.