No abstract

No abstract

With the development of deep learning and remote sensing technologies in recent years, many semantic segmentation methods based on convolutional neural networks (CNNs) have been applied to road extraction. However, previous deep learning-based road extraction methods primarily used RGB imagery as an input and did not take advantage of the spectral information contained in hyperspectral imagery. These methods can produce discontinuous outputs caused by objects with similar spectral signatures to roads. In addition, the images obtained from different Earth remote sensing sensors may have different spatial resolutions, enhancing the difficulty of the joint analysis. This work proposes the Multiscale Fusion Attention Network (MSFANet) to overcome these problems. Compared to traditional road extraction frameworks, the proposed MSFANet fuses information from different spectra at multiple scales. In MSFANet, multispectral remote sensing data is used as an additional input to the network, in addition to RGB remote sensing data, to obtain richer spectral information. The Cross-source Feature Fusion Module (CFFM) is used to calibrate and fuse spectral features at different scales, reducing the impact of noise and redundant features from different inputs. The Multiscale Semantic Aggregation Decoder (MSAD) fuses multiscale features and global context information from the upsampling process layer by layer, reducing information loss during the multiscale feature fusion. The proposed MSFANet network was applied to the SpaceNet dataset and self-annotated images from Chongzhou, a representative city in China. Our MSFANet performs better over the baseline HRNet by a large margin of +6.38 IoU and +5.11 F1-score on the SpaceNet dataset, +3.61 IoU and +2.32 F1-score on the self-annotated dataset (Chongzhou dataset). Moreover, the effectiveness of MSFANet was also proven by comparative experiments with other studies.

Multi-object semantic segmentation from remote sensing images has gained significant attention in land resource surveying, global change monitoring, and disaster detection. Compared to other application scenarios, the objects in the remote sensing field are larger and have a wider range of distribution. In addition, some similar targets, such as roads and concrete-roofed buildings, are easily misjudged. However, existing convolutional neural networks operate only in the local receptive field, and this limits their capacity to represent the potential association between different objects and surrounding features. This paper develops a Multi-task Quadruple Attention Network (MQANet) to address the above-mentioned issues and increase segmentation accuracy. The MQANet contains four attention modules: position attention module (PAM), channel attention module (CAM), label attention module (LAM), and edge attention module (EAM). The quadruple attention modules obtain global features by expanding the receptive fields of the network and introducing spatial context information in the label. Then, a multi-tasking mechanism which splits a multi-category segmentation task into several binary-classification segmentation tasks is introduced to improve the ability to identify similar objects. The proposed MQANet network was applied to the Potsdam dataset, the Vaihingen dataset and self-annotated images from Chongzhou and Wuzhen (CZ-WZ), representative cities in China. Our MQANet performs better over the baseline net by a large margin of +6.33 OA and +7.05 Mean F1-score on the Vaihingen dataset, +3.57 OA and +2.83 Mean F1-score on the Potsdam dataset, and +3.88 OA and +8.65 Mean F1-score on the self-annotated dataset (CZ-WZ dataset). In addition, each image execution time of the MQANet model is reduced 66.6 ms compared to UNet. Moreover, the effectiveness of MQANet was also proven by comparative experiments with other studies.