Road Extraction Using a Dual Attention Dilated-LinkNet Based on Satellite Images and Floating Vehicle Trajectory Data

Abstract

Automatic extraction of road from multi-source remote sensing data has always been a challenging task. Factors such as shadow occlusion and multi-source data alignment errors prevent current deep learning-based road extraction methods from acquiring road features with high complementarity, redundancy, and crossover. Unlike previous works that capture contexts by multi-scale feature fusion, we propose a dual attention dilated-LinkNet (DAD-LinkNet) to adaptively integrate local road features with their global dependencies by joint using satellite image and floating vehicle trajectory data. Firstly, a joint leastsquares feature matching-based floating vehicle trajectory correction model is used to correct the floating vehicle trajectory; then a convolutional network model DAD-LinkNet based on a dual-attention mechanism is proposed, and road features are extracted from the channel domain and spatial domain of the target image in turn by constructing a dual-attention module in the dilated convolutional layer and adopting a cascade connection; a weighted hyperparameter loss function is used as the loss function of the model; finally, the road extraction is completed based on the proposed DAD-LinkNet model. Experiments on three datasets show that the proposed DAD-LinkNet model outperforms the state-of-the-art methods in terms of accuracy and connectivity.

MS-AGAN: Road Extraction via Multi-Scale Information Fusion and Asymmetric Generative Adversarial Networks from High-Resolution Remote Sensing Images under Complex Backgrounds

“…However, complex backgrounds with high vegetation coverage and occlusions still remain a challenge. Gao et al propose the DAD-LinkNet and introduce trajectories into the integration with roads extracted from heterogeneous remote sensing images [107]. Trajectories are considered as local semantic data which are adaptively combined with diverse road features with a selfattention mechanism.…”

MultiScale Probability Map guided Index Pooling with Attention-based learning for Road and Building Segmentation

“…To improve accuracy and limit the model weight, D-Linknet [35] followed Linknet [12] by making a direct residual connection from the encoder to the decoder and used dilated convolutions to increase the receptive field. DAD-LinkNet [31] adaptively integrated the local and global road features by using floating vehicle trajectory and satellite data jointly. In [36], the authors incorporated a feature pyramid network in the generative adversarial networks to minimize the difference between the source and target domains for road segmentation.…”

“…In [36], the authors incorporated a feature pyramid network in the generative adversarial networks to minimize the difference between the source and target domains for road segmentation. A considerable performance from encoder-decoder-based framework are observed in [11,31,35] for road segmentation.…”

“…From another perspective, it is a fact that attention-based learning improves salient feature extraction while suppressing irrelevant features [27,28,29]. Attention mechanisms have also been infused with segmentation algorithms in [7,30,31,32], to extract better road and building segmentations, mostly by averaging or maximizing salient features over the channel dimensions of the feature maps. For example, CoANet [11] applies a connectivity attention module based on Squeezeand-excitation networks (SE-Net) [29].…”

“…For example, CoANet [11] applies a connectivity attention module based on Squeezeand-excitation networks (SE-Net) [29]. In [30,31], spatial and channel attention modules are utilized based on the scaleddot product attention of [33]. However, despite its usage, it is seldom found that attention learning highlights some domaindependent information like image backgrounds.…”