Do image resolution and classifier choice impact island biogeographical parameters of terrestrial islands?

Island biogeography provides concepts for conservation management as fragmented habitats are comparable to ocean islands. Remote sensing can help to extract terrestrial habitat islands on the landscape scale. However, little is known about the effects of image resolution and classification algorithms on resulting island size and related parameters. We study the combined effect of three image resolutions (2, 10, and 30 m) and three classification algorithms (Artificial Neural Network, Random Forest, Support Vector Machine) by extracting quartz islands from WorldView-2 imagery using image segmentation. We compared four island parameters (i.e., area, distance, shape index, and perimeter-area ratio between resolutions and classifiers). We found that in all cases, image resolution and classification algorithms had a strong effect. However, image resolution was more important for area and shape. Artificial Neural Network always provided the best performance as a classifier (OA: 0.880, kappa: 0.801, F1: 0.912). Hence, conservation strategies could lead to different results when different pattern extraction strategies are applied. Future studies which aim at extracting terrestrial habitat islands from image datasets should aim for the highest possible resolution and compare the outcomes of different classifiers to ensure the best possible results.

Automated summary

This study investigates the effects of image resolution and classification algorithms on the extraction of terrestrial habitat islands. The results show that image resolution and classification algorithms significantly influence island size and shape parameters, with image resolution being more important for area and shape. Artificial Neural Network consistently performs the best as a classifier.