Scale-Robust Deep-Supervision Network for Mapping Building Footprints From High-Resolution Remote Sensing Images

Abstract

Building footprint information is one of the key factors for sustainable urban planning and environmental monitoring. Mapping building footprints from remote sensing images is an important and challenging task in the earth observation field. Over the years, convolutional neural networks have shown outstanding improvements in the building extraction field due to their ability to automatically extract hierarchical features and make building predictions. However, as buildings are various in different sizes, scenes, and roofing materials, it is hard to precisely depict buildings of varied sizes, especially in large areas (e.g. nationwide). To tackle these limitations, we propose a novel deep-supervision convolutional neural network (denoted as DS-Net) for extracting building footprints from high-resolution remote sensing images. In the proposed network, we applied deep supervision with an extra lightweight encoder, which enables the network to learn representative building features of different scales. Furthermore, a scale attention module is designed to aggregate multiscale features and generate the final building prediction. Experiments on two publicly available building datasets, including the WHU Building Dataset and the Massachusetts Building Dataset, show the effectiveness of the proposed method. With only a 0.22M increment of parameters compared with U-Net, the proposed DS-Net achieved an IoU of 90.4% on the WHU Building Dataset and 73.8% on the Massachusetts Dataset. DS-Net also outperforms the state-of-theart building extraction methods on the two datasets, indicating the effectiveness of the proposed deep supervision and scale attention. Index Terms-Remote sensing image, deep learning, convolutional neural network, building footprint extraction.

ReA-Net: A Multiscale Region Attention Network With Neighborhood Consistency Supervision for Building Extraction From Remote Sensing Image

“…This is because the proposed multi-scale region attention mechanism and region consistency supervision strategy can better enhance the network's ability to extract semantic features from the remote sensing images, and by constraining and supervising the building regions and boundaries in the remote sensing images, it can make the network enrich the degree of attention to the local regions and contours of buildings, which improves the segmentation of the proposed network. VI, rows 1 to 10 are FCN [40], SegNet [41], DeeplabV3 [42], PSPNet [43], Unet [39], Res-Unet [44], HR-Net [45], Chen et al [46], BRRNet [20] and DS-Net [49] for semantic segmentation quantification, respectively. The experiments comparison results show that, compared with FCN, SegNet, DeeplabV3, PSPNet, Unet, Res-Unet, HR-Net, Chen et al, BRRNet and DS-Net, the Precision of the proposed algorithm has improved by 6.37%, 17.87% , 5.49%, 4.17%, 3.57%, 6.96%, 4.42%, 0.59% and -3.84%, respectively, and the Recall of the proposed algorithm has increased by 7.11%, 4.78%, 4.84%, 5.29%, 3.13%, 4.95%, 5.25%, 1.38% and 5.29%, respectively; the IoU of the proposed algorithm has improved by 9.53%, 16.7%, 7.39%, 6.73%, 4.79%, 8.56%, 6.88%, 1.28%, 0.31% and 0.98%, respectively; the F1-Score of the proposed algorithm has improved by 6.73%, 12.22%, 5.18%, 4.69%, 3.35%, 6.02%, 4.82%, 0.97%, 0.33% and 0.78%, respectively.…”

“…The proposed ReA-Net on the Aerial Imagery Dataset precision is 95.61%, Recall is 95.68%, IoU is 91.6%, MACs is 235.16G, and Params is 22.14M. Rows 1 to 12 are the semantic segmentation quantification of the comparisons algorithms such as FCN [40], SegNet [41], DeeplabV3 [42], PSPNet [43], Unet [39], Res-Unet [44] ,HR-Net [45], Deng et al [25], DR-Net [21], Chen et al [46], RSR-Net [48] and DS-Net [49] respectively. Compared with FCN, SegNet, DeeplabV3, PSPNet, Unet, Res-Unet, HR-Net, Deng et al , DR-Net , Chen et al, RSR-Net and DS-Net, the experiments results showed that the Precision of ReA-Net is improved by 4.86%, 3.13%, 2.1%, 1.15%, 1.26%, ,1.13%, 2.94%, 0.64% ,1.31%, 2.36%, 2.39% and 0.76% respectively; the Recall of ReA-Net is improved by 5.68%, 4.13%, 2.62%, 2.13%, 2.02%, 2.84%, 2.02%, 0.87% ,1.38%, 0.12%, 3.43% and 0.62% respectively; the IoU of ReA-Net is improved by 9.14%, 6.4%, 4.18%, 2.92%, 3.34%, 3.55% , 5.09%, 1.3% ,3.3%, 2.21%, 3.28% and 1.20% respectively; the F1-Score of ReA-Net is improved by 5.25%, 3.64%, 2.44%, 1.64%, 1.87%, 1.99%, 2.48%, 0.73% ,1.84.%, 1.24%, 2.91% and 0.68% respectively.…”

“…It can be seen from Table VI that the proposed ReA-Net algorithm on the Massachusetts dataset has a Precision of 83.93%, a Recall of 87.53%, an IoU of 74.77%, a MACs of 58.79G, and a Params of 22.14M. In TableVI, rows 1 to 10 are FCN[40], SegNet[41], DeeplabV3[42], PSPNet[43], Unet[39], Res-Unet[44], HR-Net[45], Chen et al[46], BRRNet[20] and DS-Net[49] for semantic segmentation quantification, respectively. The experiments comparison results show that, compared with…”

Building Type Classification with Incomplete Labels

“…The building footprints are either obtained from the satellite images, or, if available, from map data such as OSM. This paper focuses on the classification and post-processing part, as there are various other studies focusing on the building extraction task [10,11,12].…”

Uncertainty, Edge, and Reverse-Attention Guided Generative Adversarial Network for Automatic Building Detection in Remotely Sensed Images

“…In order to leverage large-scale contextual information and extract critical cues for identifying building pixels in the presence of complex background and when there is occlusion, researchers have proposed methods to capture local and long-range spatial dependencies among the ground entities in the aerial scene [55], [56]. Several researchers are also using transformers [60], attention modules [12], [61]- [63], and multiscale information [8], [43], [45], [46], [64]- [66] for this purpose. Recently, multiview satellite images [67], [68] are also being used to perform semantic segmentation of points on ground.…”