Species habitat modeling based on image semantic segmentation

Abstract

Habitat monitoring has emerged as a crucial practice for preserving ecological environments and ensuring species reproduction. Traditional habitat modeling often relies on the “lasagna model”—a McHarg-style approach that focuses on the ecological niche formed by the combined effect of multiple geographical factors at a single location. This model, however, overlooks the influence of the broader surrounding environment on habitat suitability. In this study, we propose a habitat modeling framework that integrates surrounding environmental conditions by employing kernel density analysis and a semantic segmentation method. The results demonstrate that kernel density analysis is effective in expanding the presence-only data into presence-absence data for habitat modeling. The semantic segmentation method, Segformer, outperforms the traditional MaxEnt in mapping the habitat of the Sandpiper family in Taiwan, achieving a higher Area Under the Curve (AUC) score (0.76 vs. 0.69). Another case study of the Swallow family indicates the limitations of the proposed method. This study highlights the potential of applying deep learning methods to habitat modeling, contributing to more comprehensive biodiversity assessments and conservation planning.