Temporal comparison of coral reefs in the Galápagos using differential alignment of 3D models and Machine-Learning

Underwater photogrammetry is a popular tool for coral community surveys, however, comparing 3D maps over time requires a complicated set of operations. Primarily, it is difficult to align dynamic scenes in order to visualize and measure changes because different components in the map move in different directions and rates. Moreover, analyzing 3D maps for coral community structure involves image annotation which is laborious and requires expert knowledge. Here, we use differential alignment and Model Assisted Labeling to solve the aforementioned problems. Instead of aligning our full maps, we crop them into sub-parts and align them separately enabling us to copy the location of 3D survey points between years and conduct a direct pointwise comparison. To annotate the points, we used classifications from a deep neural net trained by us and reviewed by a human annotator, resulting in substantial time savings. Our results show that differential alignment is effective for 3D model registration over time, and that although our classifier reached precision values typically around 60 percent for key classes, this was efficient in expediting the annotation process. We perform a pointwise temporal comparison of benthic composition and examine key ecological indicators including percent coral cover, coral diversity, and structural complexity. The results show that while structural complexity decreased over time, coral cover showed site specific trends warranting further research in the region to determine how these changes may affect reef resilience and recovery.