WARE-Net: A Deep Learning Framework for Arctic Wilderness Road Extraction Using Google Satellite Embeddings

As a primary driver of anthropogenic disturbance, roads significantly impact the fragile ecological balance and serve as a critical indicator for quantifying human expansion in the Arctic wilderness. However, current geospatial datasets in these high-latitude regions suffer from severe fragmentation and limited coverage, creating a blind spot that impedes precise environmental monitoring and sustainable development planning. To address these deficiencies, this study explores the potential of utilizing Google Satellite Embeddings combined with deep learning methods to extract roads in the Arctic wilderness. Specifically, we propose the Wilderness Area Road Extraction Network (WARE-Net), a novel road extraction model based on a U-shaped architecture. The model integrates an encoder adapted for these Embeddings to enhance feature representation. To identify road morphological characteristics, a Linear Feature Enhancement Module is developed to effectively capture multi-directional linear features. Furthermore, in the decoding phase, a detection head fusing the outputs of three decoding modules is designed to improve road extraction performance. Experimental results demonstrate that WARE-Net achieves satisfactory performance on the test set, with an F1 score of 76.17% and an IoU of 63.58%. Moreover, road extraction experiments conducted in the Khanty-Mansiysky District (Russia, covering 46,400 km²) further validate the effectiveness and generalization capability of the proposed method. In conclusion, our approach holds significant promise for achieving large-scale, rapid, and accurate wilderness road extraction, thereby providing vital technical support for sustainable development assessment in the Arctic.