Obtaining Road information from high-resolution remote sensing images (HRSI) is gaining attention in intelligent transportation systems. Existing road extraction methods tend to improve road connectivity with graph convolution or global attention, however, ignore the damage of introduced excessive effective receptive field (ERF) to multi-scale road details. In this study, we propose an Auxiliary-Decoding Road Extraction Network named AD-RoadNet which decouples multi-scale road representation and connectivity improvement based on two modules; the Hybrid Receptive Field Module (HRFM) and the Topological Feature Representation Module (TFRM). The HRFM is introduced in the encoder to emphasize target road features through adaptively matching the receptive field (RF) size for various scale roads, thus beneficial for multi-scale road representation. The TFRM is introduced in an auxiliary decoder to represent topological features with the position information encoded in the shared encoder and then helps the main decoder reason occluded roads, thus improving connectivity. Between the encoder and main decoder. The proposed model has a similar parameter scale as HRNetV2 and outperforms the state-of-the-art ResUnet, D-LinkNet, and HRNetV2 by 3.34%, 2.03%, and 1.53% in the mean Intersection of Union (mIoU) on DeepGlobe Road Dataset. Ablation analysis, inference size matter, and the robustness for unseen occlusion scenarios, low-quality labels, and various quality inference images are further presented to evaluate the proposed AD-RoadNet. All codes are available on https://github.com/zqluo-buaa/ADRoadnet
Land cover (LC) products play a crucial role in various fields such as change detection, resource management, and urban planning. The diversity in methods and principles used to create different products poses a challenge for researchers in choosing the most suitable one for research needs. Mainstream evaluation methods typically consider only a portion of the accuracy information from the product and require a significant effort in creating validation datasets. Here, we propose a relative accuracy assessment method for LC products based on optimal transport theory, which provides a comprehensive evaluation by utilizing a broader range of accuracy information within the product. The method can directly compute the similarity between the target product and the reference truth at a global scale, addressing the issue of quantitatively assessing product accuracy in the absence of a validation dataset. To validate the effectiveness of the method, we select Beijing as the study area to assess the accuracy of four LC products. The results suggest that the method allows for precise quantification of product accuracy, aligning closely with validation outcomes, which can provide valuable guidance to researchers in both product creation and selection.
Remote sensing products, such as land cover data products, are essential for a wide range of scientific studies and applications, and their quality evaluation and relative comparison have become a major issue that needs to be studied. Traditional methods, such as error matrices, are not effective in describing spatial distribution because they are based on a pixel-by-pixel comparison. In this paper, the relative quality comparison of two remote sensing products is turned into the difference measurement between the spatial distribution of pixels by proposing a max-sliced Wasserstein distance-based similarity index. According to optimal transport theory, the mathematical expression of the proposed similarity index is firstly clarified, and then its rationality is illustrated, and finally, experiments on three open land cover products (GLCFCS30, FROMGLC, CNLUCC) are conducted. Results show that based on this proposed similarity index-based relative quality comparison method, the spatial difference, including geometric shapes and spatial locations between two different remote sensing products in raster form, can be quantified. The method is particularly useful in cases where there exists misregistration between datasets, while pixel-based methods will lose their robustness.