DBNINet: A Decoder-Guided Boundary Enhancement and Node Interaction Network for Road Extraction From Remote-Sensing Images

Abstract

Roadextraction from remote-sensing images is a fundamental task in Earth observation applications. However, accurate extraction remains challenging due to complex backgrounds, weak boundaries, tree occlusion, and discontinuous road structures, particularly in forest and rural scenes. To address these issues, this article proposes DBNINet, a decoder-guided boundary and topology enhancement network for road extraction. Two complementary enhancement strategies are introduced. In the skip connection pathways, a Road-Boundary Enhancement Module (RBEM) and a Foreground-Adaptive Perception Module (FAPM) are designed. RBEM first leverages wavelet transform to enhance road edge features, while FAPM captures complementary foreground and background cues using decoder-guided road structure information. These features are then utilized to dynamically extract road regions. At the bottleneck, a Node Interaction Module is introduced to model global–local road structure relationships and improve structural connectivity. Experiments are conducted on a mixed dataset comprising forest-road images and the DeepGlobe road dataset. The proposed DBNINet achieves 53.05% IoU, 69.32% F1-score, 71.14% precision, and 69.32% recall, outperforming several state-of-the-art segmentation methods. Experimental results demonstrate that DBNINet effectively preserves thin road boundaries, reduces false detections, and improves road connectivity in complex remote-sensing scenes.