This web application uses cookies and other tracking technologies to ensure you get the best experience.
Resource ID
10674
Access
Open
Full Title
Spatial density modelling using statistical and machine learning methods and its application to the Antarctic krill
Author
Fang Lu, Christian Reiss, George Watters, Hiroto Murase and Toshihide Kitakado
Abstract
As a key species linking primary producers to the higher tropic levels in the Antarctic ecosystem, the Antarctic krill (Euphasia superba) plays an important role in the Antarctic ecosystem. Therefore, knowing the plausible spatial distribution of the krill will be useful for the resource management and conservation in this area. Species distribution models (SDMs) can help predict the spatial species density by quantifying the relationship between the observed species distribution and its influencing factors. In general, although both statistical models and machine learning methods can be applied as SDMs, there is a still open question of how the estimation performance of those SDMs for the Antarctic krill is. To address this question, we conducted simulation studies for six different SDMs under two different survey-designs with zig-zag-shaped and tooth-shaped track lines, in order to assess the estimation performance of each model for krill distribution and mean density. As the procedure, we first conditioned two different density distribution of krill in this region by using actual krill density observation data. Using the assumed true spatial density surfaces, we repeatedly generated simulation data under the two designs, and then applied six SDMs, two statistical models and four machine learning methods, to the data. As performance measures, the mean squared error of predicted surface (MSE), relative bias and root mean squared error (RMSE) for the mean density were used. As a result, machine leaning methods were proven to have higher and more reliable prediction abilities than traditional statistical models, especially random forests (RF) and boosted regression trees (BRT) were revealed to be the most reliable methods in this study. In addition, the zigzag-shaped and tooth-shaped designs are found to have comparable performances, and both of them can be applied in the krill field survey.