Monitoring Corals using AI

Corals are one of the most diverse marine species, that play a vital role in the marine ecosystem. Because of such a diverse colony, coral reefs are therefore known as the "rainforest of the ocean". According to NOAA, about 25% of the ocean's fish depend on coral reefs for food, shelter, and reproduction. Reefs are also a source of many medicines. They protect coasts from erosion and even help in their recreation. NOAA also claims that around half a billion people depend on corals and a separate study from United Nations stated that the corals are estimated to be worth more than ten billion per year. This clearly shows that corals are an important component of the ecosystem.

However, due to recent changes in the climate due to global warming and irregularities in the oceanic water temperature, the corals are at a great threat. The rising ocean temperature, increasing marine pollution, and uncontrolled fishing have severely damaged the coral ecosystem. Increasing events of coral bleaching have increased the death rates of corals. Areas rich with corals, like the Great Barrier Reef in Australia, have recorded a large number of corals affected due to such bleaching events. Reefs across several hundred miles were seen affected by the bleaching event. 

Hence, it is now extremely necessary to regularly undergo coral examinations to have a better understanding of affected reefs. Researchers are trying new ways to study coral beds with improved efficiency. Several manual methods are currently implemented by scientists to track coral ecosystems. But recently, AI has entered coral research. The addition of the Deep Learning based model to the cloud-based coral portal CoralNet will help researchers a lot. CoralNet is an open-source platform for the automatic and manual analysis of coral reefs. More than 1.7 images are uploaded to date by more than 3000 registered users. 

Corals are spread across the entire globe covering about 284,000 sq km of land. It is therefore practically impossible to analyze the entire reef cover. Hence the studies are mostly based on the samplings. Often the surveys are made by pointing the camera downward to a particular coral and taking picture of the defined sample just above the bottom. The goal of these surveys mostly includes analyzing the health of corals, the area covered by them, and the possibility of any disease among their colonies. 

CoralNet is a platform for coral reef image analysis that allows users to create sources, which are collections of images with associated labels. These labels are chosen from a pre-defined set of options and are used to classify specific regions within the images. CoralNet's machine learning engine uses these labeled regions to train a classifier that can then automatically label new images. When the classifier receives an image patch, it outputs a score for each label in the set, ranging from 0.0 to 1.0. The label with the highest score is considered the annotation for the pixel location in the image patch. 

The scores for all labels sum to 1.0, indicating the classifier's confidence in its decision. This classification process is important because it allows researchers to efficiently analyze large datasets of coral reef images. With CoralNet, researchers can quickly identify areas of interest within images and categorize them based on a defined set of labels. This can help identify trends and patterns within the data that may be missed by human analysis alone. Overall, CoralNet's ability to automatically label images based on user-defined labels provides a powerful tool for coral reef researchers to analyze large datasets and gain a better understanding of these complex ecosystems.

Marine scientists have uploaded 1.7 million images from over 2,040 ecological surveys around the world on CoralNet Alpha, a platform initially created for the annotating of coral reefs. The uploaded images cover a broader range of habitats and classes, including seagrasses, cold water rocky habitats, oil rigs, pier pilings, and autonomous reef monitoring structures. The sources are from different groups of marine scientists and there is no universally agreed set of labels or taxonomy. Therefore, users have defined a total of 4,489 labels, including duplicates. The duplicates were identified, and corresponding duplicates were merged into a common label by selecting 280 representative sources for training the deep learning engine. The selected sources were randomly divided into 254 sources for training and testing the backbone networks and 26 sources for training and testing the classifiers. A total of 1,279 labels were designated as V1, which are used in at least 3 sources, are used to annotate at least 100 points, and do not designate "unsure," "dark," or similar catch-all categories. In V2, an additional 50 sources were exported for training, and 4 catch-all type labels were removed.

The backbone networks are designed to extract visual features from the images, while the classifiers are used to classify the extracted features into different taxa and habitats. The researchers also developed a new deep learning architecture called CoralNet-ResNet, which is specifically designed for the CoralNet dataset and achieves state-of-the-art performance on the dataset.

The CoralNet model offers a valuable tool for marine scientists to annotate and classify ecological surveys, thereby contributing to a better understanding of marine ecosystems and their conservation. The accurate classification of different taxa and habitats can help researchers identify areas that are particularly vulnerable to human activities and prioritize conservation efforts accordingly. Moreover, the CoralNet model can be used to monitor changes in marine ecosystems over time, which can provide valuable information for policymakers and stakeholders. In conclusion, the CoralNet model represents a significant advance in the field of marine ecology and conservation. The development of the CoralNet model has also highlighted the importance of collaboration and data sharing in scientific research, as the model relies on a large dataset of annotated images collected from different sources around the world. As such, the CoralNet model serves as a model for how data sharing and collaboration can contribute to scientific progress and environmental conservation.