ME-Net: A Multi-Scale Erosion Network for Crisp Building Edge Detection from Very High Resolution Remote Sensing Imagery

Abstract

The detection of building edges from very high resolution (VHR) remote sensing imagery is essential to various geo-related applications, including surveying and mapping, urban management, etc. Recently, the rapid development of deep convolutional neural networks (DCNNs) has achieved remarkable progress in edge detection; however, there has always been the problem of edge thickness due to the large receptive field of DCNNs. In this paper, we proposed a multi-scale erosion network (ME-Net) for building edge detection to crisp the building edge through two innovative approaches: (1) embedding an erosion module (EM) in the network to crisp the edge and (2) adding the Dice coefficient and local cross entropy of edge neighbors into the loss function to increase its sensitivity to the receptive field. In addition, a new metric, Ene, to measure the crispness of the predicted building edge was proposed. The experiment results show that ME-Net not only detects the clearest and crispest building edges, but also achieves the best OA of 98.75%, 95.00% and 95.51% on three building edge datasets, and exceeds other edge detection networks 3.17% and 0.44% at least in strict F1-score and Ene. In a word, the proposed ME-Net is an effective and practical approach for detecting crisp building edges from VHR remote sensing imagery.

A Building Shape Vectorization Hierarchy From VHR Remote Sensing Imagery Combined DCNNs-Based Edge Detection and PCA-Based Corner Extraction

“…The network selection is very significant because the crisper building edges will directly improve the performance of the building corner extraction and shape vectorization in later steps. We noticed that ME-Net [10] produced relatively crisp and accurate building edges. In Section IV-A, we have demonstrated that ME-Net is more suitable in our hierarchy than the other state-of-the-art edge detection networks.…”

“…3: Set the classification threshold to 0.5 for all the pixels in P, and store the edge pixels in B. which are same as the parameters of the published study [10]. Since the ME-Net produced the building edge probability map with values between 0 and 1, we set a threshold [28], [29], [30] for classifying both the edge and nonedge pixels, and the threshold is generally 0.5 [31].…”

“…The Jiangbei New Area Buildings dataset presented by Wen et al [10] covers 53.67 km 2 of the Jiangbei New Area, Nanjing City, Jiangsu Province, China. As shown in Fig.…”

“…As suggested by Wen et al [10], to avoid the issue of overfitting the convolution neural networks, the images and label maps were clipped and augmented through rotation. The augmentation of training set is used to extract more complex and diverse buildings.…”

“…We conducted relevant experiments, and proved that the DCNNs-based edge detection networks are feasible for building edge extraction from VHR remote sensing images. ME-Net [10] achieved the highest accuracy in detecting the clearest and crispest building edges. The detection results of the building edges are shown in columns 3−7 of Fig.…”