machine learning

Summary

Habitat mapping is crucial for understanding habitat connectivity and for spatial planning, environmental management, conservation, and targeted research, including long-term change monitoring. However, such information has been lacking for many Dutch Caribbean islands, especially regarding marine habitats. This study used 2144 georeferenced images from different surveys to develop habitat models predicting the distribution of habitat types within the Saba Bank National Park. The habitat models link environmental factors to species or habitat occurrence, enabling predictions in unsurveyed areas with known covariates. Machine learning techniques (Random Forests, Gradient Boosting, and weighted K Nearest Neighbor) were applied to interpret and predict ten habitat types over the Bank. Three models were created for each technique: 1) utilizing only geographic coordinates; 2) incorporating covariables such as depth, distance to the edge of the Bank, Topographic Position Index (TPI), and Terrain Ruggedness index (TRI); 3) a combination of the previous two models. All models performed well, accurately predicting habitat types between 67 and 74% of the georeferenced images. However, the most natural representation occurred with models combining geographic and covariate variables. Predicted habitats include coral reef, patch reef, gorgonian reef, sargassum fields, cyanobacteria-dominated fields, Lobophora fields, Neogoniolithon- Lyngbya habitat, other macroalgae fields, sand with a mix of species, and bare sand. Habitat distribution appears to be related to the main currents in the area and depth, with coral reefs occurring mainly along the southern and eastern edge of the Bank, with gorgonians and other soft corals dominating there the shallow areas. Macroalgae, including fields of Sargassum, dominate the back-reef area. Extensive sand plains dominate the center of the Bank, and along the north-western and northern edge of the Bank, between 40 and 60m depth Lobophora fields can occur. In the south-eastern back reef area a number of mounds built up by the coralline alga Neogoniolithon occur. The Luymes Bank, the northeastern part of the Saba Bank, was the only area that was not correctly predicted, indicating that additional field-based observations are needed to refine results in this area.
 

Studies on coral reefs increasingly combine aspects of science and technology to understand the complex dynamics and processes that shape these benthic ecosystems. Recently, the use of advanced computational algorithms has entered coral reef science as new powerful tools that help solve complex coral reef related questions, which were unsolvable just a decade earlier. Some of these advanced algorithms consist of Computational Intelligence (CI) approaches, a branch of Artificial Intelligence that uses intelligent systems to address complex real-world problems yielding more robust, tractable and simpler solutions than those obtained by conventional mathematical techniques. This paper describes the most commonly used CI techniques related to coral reefs and the main improvements obtained with these methods over classical algorithms in this field. Some recommendations are given for the application of CI techniques to complex coral reef related problems, and vice-versa, for the application of novel coral reef dynamics concepts to improve the Coral Reef Optimization (CRO) algorithm, an optimization method inspired by coral reef dynamics.

The natural beauty of coral reefs attracts millions of tourists worldwide resulting in substantial revenues for the adjoining economies. Although their visual appearance is a pivotal factor attracting humans to coral reefs current monitoring protocols exclusively target biogeochemical parameters, neglecting changes in their aesthetic appearance. Here we introduce a standardized computational approach to assess coral reef environments based on 109 visual features designed to evaluate the aesthetic appearance of art. The main feature groups include color intensity and diversity of the image, relative size, color, and distribution of discernable objects within the image, and texture. Speci c coral reef aesthetic values combining all 109 features were calibrated against an established biogeochemical assessment (NCEAS) using machine learning algorithms. These values were generated for ∼2,100 random photographic images collected from 9 coral reef locations exposed to varying levels of anthropogenic in uence across 2 ocean systems. Aesthetic values proved accurate predictors of the NCEAS scores (root mean square error < 5 for N ≥ 3) and signi cantly correlated to microbial abundance at each site. This shows that mathematical approaches designed to assess the aesthetic appearance of photographic images can be used as an inexpensive monitoring tool for coral reef ecosystems. It further suggests that human perception of aesthetics is not purely subjective but in uenced by inherent reactions towards measurable visual cues. By quantifying aesthetic features of coral reef systems this method provides a cost e cient monitoring tool that targets one of the most important socioeconomic values of coral reefs directly tied to revenue for its local population.