TAG-Para: Hybrid deep attentions and graphical representations for Roads' network topological error correction

Abstract

Modern deep-learning methods offer incredible tools for extracting road networks from remote sensing images. However, these methods often operate as black boxes, making it challenging to interpret the uncertainty associated with their predictions, particularly when aiming to predict road networks with minimal topological errors. Therefore, this study proposed "TAG-Para network learning" a hybrid and multi-scale topological attention graphical error correction network, which is a convolutional neural network (CNN) to extract local information in the form of road nodes in network receptive field, which are also actually road masks(receptive maps). CNN's branch preserves local ground details but suffers from insufficient global modeling of various nodes. Then we bonded the CNN ViT framework in parallel multi-scale features extraction, which consists of a down-sampling-free CNN classifier and a Transformer branch's classifier to jointly capture local and global features into a unified approach named Parallel Learning. TAG-Para also comprised of Attention Node Graphical (ANG) modules, utilizes a full CNN to enhance road mask connectivity by identifying road nodes and a Road Refinement Backbone (RRB) module based on node connectivity These nodes are then used to generate a road vector map. Bayesian optimization for hyperparameter tuning offered formalism to quantify the uncertainty associated.