Abstract
ABSTRACT Islands, as a critical role in Earth’s ecosystems, serve as vital carriers for human activities and cultural heritage. Precisely identifying the quantity and area of islands can offer valuable scientific basis into pressing issues such as climate change, ecosystem fluctuations, and marine pollution. In this paper, we propose a novel method that integrates object-oriented Semantic Segmentation Networks (SSNs) with pixel-based Adaptive Thresholding (AT) for the purpose of island identification. Pixel-based accuracy evaluation of island area, the F1-score of SSNs (U-Net, PSPNet, DeepLabV3+, HRNetV2, SegFormer) + AT reach 99.61%, 99.59%, 99.59%, 99.13% and 99.60% respectively, showing an improvement of 0.59%, 1.45%, 1.44%, 0.79% and 0.24% compared to SSNs. Object-based accuracy evaluation of island quantity, the F1-score2 of SSNs (U-Net, PSPNet, DeepLabV3+, HRNetV2, SegFormer) + AT reach 90.48%, 89.60%, 90.07%, 90.28% and 90.65% respectively, showing an improvement of 10.51%, 41.98%, 14.25%, 9.06% and 5.34% compared to SSNs. This indicates that the method demonstrates notable advantages in both island area identification and island quantity identification. This method integrates object-based and pixel-based classification methods, making it more applicable for extracting different types of islands.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call