Deep Learning-Based Building Extraction from Remote Sensing Images: Comparison
View Latest Version
Please note this is a comparison between Version 2 by Peter Tang and Version 1 by Lin Luo.

Building extraction from remote sensing (RS) images is a fundamental task for geospatial applications, aiming to obtain morphology, location, and other information about buildings from RS images, which is significant for geographic monitoring and construction of human activity areas. In recent years, deep learning (DL) technology has made remarkable progress and breakthroughs in the field of RS and also become a central and state-of-the-art method for building extraction. 

  • deep learning
  • convolutional neural network
  • building extraction
  • high resolution
  • remote sensing

Deep Learning-Based Building Extraction from Remote Sensing Images: A Comprehensive Review

1. Introduction

With the rapid development of imaging technology, high-resolution remote sensing (RS) imagery is becoming more and more readily available. Therefore, research within the field of RS has flourished, and automatic building segmentation from high-resolution images has received widespread attention [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15]. The process of extracting buildings from RS images is shown in Figure 1, which is essentially a pixel-level classification of RS images to obtain binary images with contents of building or non-building, and this process can be modeled as a semantic segmentation problem [16][17][18][19][20][21][22][23][24][25][26][27][28][29].
Energies 14 07982 g001 550
Figure 1. Illustration of extracting buildings from remote sensing images. The white and black pixels in prediction denote buildings and background respectively.
Deep learning (DL), with convolutional neural networks (CNN) [30][31][32][33][34] as its representative, is an automated artificial intelligence technique that has emerged in recent years, specializing in learning general patterns from large amounts of data as well as exploiting the knowledge learned to solve unknown problems. It has been successfully applied and rapidly developed in areas such as image classification [35], target detection [36], boundary detection [37], semantic segmentation [16], and instance segmentation [38] in the field of computer vision. Proving to be a powerful tool for breakthroughs in many fields, DL techniques applied to building extraction in RS have emerged and become the mainstream technical tools. Although there are some reviews on RS image building extraction [39][40][41][42] or DL-based RS image processing [43][44][45], there is still a lack of a research that summarizes the latest results of RS image building extraction based on DL techniques. In this paper, we extensively review the DL-based building extraction from RS images, excluding the extraction of roads and other man-made features, in which the processing inputs include aerial images, satellite images, and other multi-source data such as light detection and ranging (LiDAR) point cloud data and elevation data.
As a fundamental task in the field of RS, automatic building extraction is of great significance in a wide range of application areas such as urban planning, change detection, map services and disaster management [46][47][48][49][50][51][52][53][54][55][56]. It is the basis for accomplishing these applications to have efficient and accurate building information. Building extraction has some unique features and challenges, which mainly include the following:
  • Building types are in general highly changeable. They differ in interior tones and textures and have a variety of spatial scales. In addition, their shapes and colors may vary from building to building.
  • Buildings generally stand in close proximity to features of similar materials such as roads, and can easily be confused with other elements. The segmentation quality of boundary contours is particularly important.
  • The long-distance association relationship between buildings and surrounding objects is an important concern due to a variety of complex factors that may cause foreground occlusions, such as shadows, artificial non-architectural features, and heterogeneity of building surfaces.
  • RS images have more complex and diverse backgrounds and scenes, and the shapes of buildings are more regular and well-defined than those of natural objects, rendering boundary issues particularly critical.
 

2. DL Techniques

Semantic segmentation, one of the research directions of DL most closely related to building extraction, is not an isolated research area, but a natural step in the process from coarse to refined inference. It is a downstream task of image classification, a fundamental computer vision task, for which image classification models provide feature extractors that extract rich semantic features from different layers. Here, firstly, we recall the classic deep CNNs and design inspirations used as deep semantic segmentation systems, and point out its enlightening role for subsequent segmentation networks. Then, transfer learning, an important means of training DL models, is introduced. Finally, we introduce the loss functions used to train segmentation networks.

2.1. Deep CNNs

As one of the most fundamental tasks in computer vision, the image classification task assigns labels based on the input image and predefined categories. CNN-based image classification methods have matured in recent years and have become an important part of the downstream task of semantic segmentation. Here, we briefly review some classical CNN architectures for image processing, which include VGG, GooLeNet, and ResNet.

2.1.1. VGG Networks

In 2014, Visual Geometry Group (VGG) [57] at the University of Oxford proposed a network with more than 10 layers with concise design principles to build deeper neural network models. The structure of the VGG network is shown in Figure 2, with the main components being a 3 × 3 convolution operation and a 2 × 2 max-pooling operation. The small-sized convolutional layer has a smaller number of parameters and computations than the convolutional layer with large-sized convolutional kernels (e.g., 5 × 5 or 7 × 7 convolution operations in AlexNet [35]) to obtain a similar perceptual field. In addition, a remarkable feature is to increase the number of feature maps after using the pooling layer, reducing the phenomenon of useful information loss in feature maps after downsampling.
Energies 14 07982 g003 550
Figure 2. VGG network architecture.
VGG network is one of the most influential CNN models because of its reinforcing the important idea in DL that CNNs with deeper architectures can facilitate hierarchical feature representation of visual data. It could be a guide to the structural design of subsequent deep CNN models. Meanwhile, VGG with 16 layers (VGG-16) has become one of the common feature extractors for downstream tasks.

2.1.2. GoogLeNet

GoogLeNet based on inception modules was proposed by Google in 2014 [58], which won the ImageNet competition that year. It has been improved several times in the following years, leading to InceptionV2 [59], InceptionV3 [60] and InceptionV4 [61]. The structure of the inception module is shown in Figure 3, presenting a net-in-network (NIN) architecture. The same network layer including large-size convolution, small-size convolution and pooling operations can capture feature information separately in a parallel manner. In addition, inception modules control the number of channels with 1×1
 
convolution and enhance the network representation by fusing information from different sensory domains or scales. Due to these modules, the number of parameters and operations is greatly reduced, while the network advances in terms of storage footprint and time consumption. The idea of inception provides a new way of stacking networks for CNN architecture design, rather than just sequential stacking, as well as it can be designed to be much wider. For the same number of parameters, inception-based networks are wider and more expressive, providing a fundamental direction for lightweight design of deep neural networks.
Energies 14 07982 g004 550
Figure 3. Inception module with dimensionality reduction from the GoogLeNet architecture.

2.1.3. ResNet

Presented in 2015, ResNet [62] is a landmark research result that pushed neural networks to deeper layers. 152-layer ResNet ranked in the top five at ILSVRC 2015 with an error rate of 3.6% and achieved a new record with respect to classification, detection, and localization in a single network architecture. Through experiments and analysis of several deep CNN models, it was found that deep networks experience network degradation during layer deepening and could not necessarily perform better than shallow networks. In response, a deep residual structure was proposed as shown in Figure 4, allowing the network to shift to learning of residuals. The residual network learns new information different from what was previously available, relieving the pressure on the deep network to learn feature representations and update parameters. It allows the DL model to move once again in a deeper and better direction.
Energies 14 07982 g005 550
Figure 4. Residual block from the ResNet architecture.

Abstract

2.2. Transfer Learning

Training a deep neural network from scratch is often difficult for two reasons. On the one hand, training a deep network from scratch requires enough dataset, and the dataset of the target task is not large enough. On the other hand, it takes a long time for the network to reach convergence. Even if a sufficiently large dataset is obtained and the network can reach convergence in a short time, it is much better to start the training process with weights from previous training results than with randomly initialized weights [63][64]. Yosinski et al. [65] demonstrated that even features learned by migration from less relevant tasks are better than those learned directly from random initialization. It also takes into account that the transferability will decrease as the difference between the source task and the target task increases.
However, the application of transfer learning techniques is not so straightforward. The use of pre-trained networks must satisfy the network architecture constraint of using existing network architectures or network components for transfer learning. Then, the training process itself in transfer learning is very small compared to the training process from scratch, so it can pave the way for fast convergence of downstream tasks. An important practice in transfer learning is to continue the training process from a previously trained network to fine-tune the weight values of the model. It is important to choose the layers for fine-tuning wisely, generally choosing the higher layers in the network, as the underlying layers generally tend to retain more general features.
ImageNet [66][67] is a large image classification dataset in the field of computer vision and is often used to train the feature extraction network part of segmentation networks. VGG-16 and ResNet pre-trained by ImageNet are easily available to be used as the encoder part of the segmentation network as well. In addition, a large collection of RS image segmentation data has also been collected and merged into a large dataset and used to pre-train the segmentation network [68].
Building extraction from remote sensing (RS) images is a fundamental task for geospatial applications, aiming to obtain morphology, location, and other information about buildings from RS images, which is significant for geographic monitoring and construction of human activity areas. In recent years, deep learning (DL) technology has made remarkable progress and breakthroughs in the field of RS and also become a central and state-of-the-art method for building extraction. This paper provides an overview over the developed DL-based building extraction methods from RS images. Firstly, we describe the DL technologies of this field as well as the loss function over semantic segmentation. Next, a description of important publicly available datasets and evaluation metrics directly related to the problem follows. Then, the main DL methods are reviewed, highlighting contributions and significance in the field. After that, comparative results on several publicly available datasets are given for the described methods, following up with a discussion. Finally, we point out a set of promising future works and draw our conclusions about building extraction based on DL techniques.

2.3. Loss Function

Deep neural network models are trained with the loss-gradient back-propagation algorithm, so that the design of the loss function is also directly related to the efficiency of the network training and the performance of the model on the target task. Then the rest of this section describes several commonly used loss functions in building segmentation networks. To facilitate the expression of the computational process, y and p denote the ground truth label and prediction result, respectively.

Contributions

DeconvNet-Fusion [2]

**

**

Multi-source data post-fusion

FCN [83]

*

*

Early CNNs

ConvNet [84]

*

*

Signed distance

Fused-FCN4s [85]

**

**

Multi-source data post-fusion

SegNet-Dist [86]

*

*

Signed distance

MC-FCN [87]

*

**

Multi-scale architecture

MFRN [88]

*

**

Multi-scale architecture

BR-Net [89]

**

**

Boundary extraction, multiple tasks

GMEDN [90]

***

**

NB, multi-scale architecture

ENRU-Net [91]

**

**

APNB

PISANet [92]

**

**

Pyramid self-attention module

ELU-FCN-CRFs [93]

*

*

ELU, CRFs

FC-DenseNet-FPCRF [

  • Cross entropy loss: Cross entropy loss (CE) is the most commonly used loss function in dense semantic annotation tasks. It can be described as:

  • Weighted Cross Entropy loss: Weighted cross entropy loss (WCE) is obtained by summing over all pixel losses and can not actively cope with application scenarios such as building extraction where the categories are unbalanced. Therefore, WCEs that consider category imbalance, such as median frequency balancing (MFB) [
  • 2
  • in East Asia with a ground resolution of 2.7 m. Images of different colors from different sensors and seasons constitute a challenging case for automated building extraction. The vector building map contains 29,085 buildings. The entire image is also seamlessly cropped into 17,388 slices for training and testing, processed in the same way as the aerial dataset. Of these, 21,556 buildings (13,662 tiles) were used for isolated training and the remaining 7529 buildings (3726 tiles) were used for testing. An example is shown in Figure 7.
  • 69][70] CEs, have emerged.

where w is the category balance weight in median form, expressed by the ratio of the median of the pixel frequencies of all the categories to the pixel frequencies of that category.
  • Dice loss: Dice loss is designed for the intersection over union (IoU), an important evaluation metric in semantic segmentation, and is designed to improve the performance of the model by increasing the value of this evaluation metric.

  • Focal loss: Focal loss (FL) is improved from CE loss. To address class imbalance, an intuitive idea is to use weighting coefficients to further reduce the loss of the easy classification category. FL can be expressed as:

where α is the weighting factor for the classes and γ≥0 is a tunable parameter.

3. Datasets and Evaluation Metrics

In the case of DL, data is an extremely important component, specially with the deepening of the network and the increasing number of parameters. The establishment of each new building extraction method for basic DL requires the validation of a dataset.

3.1. Open Datasets

The data sources used to validate the building extraction methods are numerous including datasets compiled by several research institutions and data obtained by literature authors from publicly available websites (e.g., Google Earth, OpenStreetMap, and United States Geological Survey [71][72][73][74][75][76][77]). The former is of higher quality, while the latter is relatively more confusing and less generalized.

71

]

**

**

FPCRFs, GCNs

CNN-RNN [94]

***

***

Iterative refinement of RNN architecture

EANet [95]

***

***

Boundary-aware networks

Networks with BP loss [96]

**

**

BP loss

BRRNet [97]

***

***

Residual refinement module

DSFE-GGCN [98]

**

***

Gated GCN, deep feature embedding

FCN with LFE [99]

*

*

Local feature extraction module

EU-Net [100]

*

**

DSPP, category balanced loss

ScasNet [101]

***

***

Multi-scale aggregation

SR-FCN [102]

*

**

Multiscale prediction, ASPP

Building-A-Nets [103]

**

***

GAN

P-LinkNet [

  • Massachusetts Buildings Dataset [78]: The datasets, available on the website of Toronto University (https://www.cs.toronto.edu/~vmnih/data/, 15 August 2021), consists of 151 high-resolution aerial images of Boston’s urban and suburban areas. The image size in Massachusetts Buildings Dataset is 1500 × 1500 pixels, and each image covers a widespread area of 2250 × 2250 m2
  • The dataset was randomly divided into three subsets: a training set of 137 images, a validation set of 4 images and a test set of 10 images. It is worth mentioning that these data are restricted to regions where the average missed noise level is about 5% or lower. An example is shown in Figure 5.
  • Inria Aerial Dataset [79]: This dataset, available on https://project.inria.fr/aerialimagelabeling/ (15 August 2021), consists of 360 high-resolution RGB aerial images covering different cities, including Austin, Chicago, Gitza, West/East Tyrol, Vienna, Bellingham, Bloomington, and San Francisco. The areas cover urban buildings with different characteristics. For example, most of the buildings in Chicago and San Francisco are densely distributed and usually smaller in shape, while the buildings in Kitsap are scattered. The images have a spatial resolution of 0.3 m and an image size of 5000 × 5000 pixels, each covering a widespread surface of 1500 × 1500 m2
  • . Only 180 images were provided with public pixel annotation (ground truth), and the remaining 180 images were reserved for testing, where users could submit predicted images and obtain scores on the official website. To test the performance of the segmentation method more easily and quickly, by convention, the first five images of each region from the training set can be selected for validation. It is worth mentioning that all image data are of high quality as they are derived from different aerially captured orthorectified images of the landscape that are officially available locally, ignoring data such as Open Street Maps (OSM). An example is shown in Figure 6.
  • WHU Building Dataset [80]: The whole dataset, available on the website of Photogrammetry and Computer Vision (GPCV) at Wuhan University (http://gpcv.whu.edu.cn/data/, 15 August 2021), contains both aerial image dataset and satellite image dataset. The WHU aerial dataset covers 18,700 buildings of diverse shapes and colors. The entire image and the corresponding vector shapefiles were seamlessly cropped into 8189 patches of 512 × 512 pixels with a ground resolution of 0.3 m. The WHU satellite dataset consists of six adjacent satellite images covering 550 km
Energies 14 07982 g006 550
Figure 5. An example of the Massachusetts Building Dataset. (a) Original image; (b) Ground truth label.
Energies 14 07982 g007 550
Figure 6. An example of the Inria dataset. (a) Original image; (b) Ground truth label.
Energies 14 07982 g008 550
Figure 7. An example of the WHU dataset. (a) Original image; (b) Ground truth label.

3.2. Evaluation Metrics

In order to evaluate the performance of segmentation methods, it is usually necessary to select some quantitative evaluation metrics to evaluate the accuracy of different methods. Here, we introduce the commonly used evaluation metrics, including pixel accuracy (PA), precision (Pre), recall (Rec), F1 score (F1), and IoU. In the building extraction task, the building is the positive case and the background category is the negative case. These four metrics are defined as:

where tp, tn, fp and fn are the number of true positives, true negatives, false positives and false negatives pixels, respectively.

4. Building Extraction Methods Based on DL

DL techniques represented by CNNs have been developed for a long time in the direction of building extraction under the field of RS, whose processing of the input and output can be shown in Figure 8. Various deep neural network architectures for solving building extraction problems have emerged one after another.
Energies 14 07982 g009 550
Figure 8. Processing of input and output on DL-based building extraction study. The dataset is partially enhanced to supply the model for back-propagation training until a certain termination condition is reached, such as iteration time and number of iterations. After that, the model can enter the application phase, where inference on unseen data produces predictions that match the requirements.
Patch-based annotation networks [81][82] are the key process of the adoption of DL into the building segmentation problem, with the main advantage of helping researchers to free themselves from complex manual feature design and perform automated building extraction for high- and even ultra-high-resolution RS images. The patch-based approach is essentially an image classification network that assigns a specified label to each patch, where the last layer of the network is usually a fully connected layer. The method cuts the image into a number of sub-images much smaller than the original size, i.e. patches, after which a CNN is applied to process the individual patches and give a single classification for each one, and finally stitch them together to form a complete image. The patch-level annotation method does not require high capacity of the network, and the network is usually uncomplicated in structure and easy to design. Saito et al. [81] designed a simple neural network containing three convolutional layers and two fully connected layers to accomplish automatic extraction of buildings, in which the feasibility and effectiveness of the method was confirmed by experiments. However, the patch-based classification method has two inherent defects that can not be avoided. On the one hand, the features of neighboring patches are similar and the proportion of overlapping regions is extremely large. Thus, there is a large amount of redundant computation, resulting in wasted resources and low efficiency. On the other hand, there is a lack of long-distance information exchange. As a result, the method can not fully exploit the contextual information in high-resolution RS images, with difficulty in completely and accurately extracting buildings from complex backgrounds. Fully convolutional network (FCN) [16] is a landmark pixel-based segmentation method proposed to provide new inspiration for applying CNNs to advance building extraction research. The core idea is to use existing CNNs as encoders to generate hierarchical features and use upsampling means such as deconvolution as decoders to reconstruct images and generate the semantic labels, eliminating the fully connected layer exclusively. A classical encoder-decoder structure is formed, which can theoretically accept images of different sizes as network inputs and output semantically labeled images at pixel level with the same resolution. There exists a common point in current segmentation networks, i.e., feature extraction is performed by the encoder in the process of performing multi-stage downsampling, and the decoder gradually recovers the size and structure of the image in the process of upsampling and generates semantic annotations. Based on this starting point consideration, a popular approach is to use the image classification network with the fully connected layer removed directly as a feature extraction network, i.e., encoder, such as VGG-16, GoogLeNet and ResNet; the decoder part is composed of upsampling modules such as deconvolution, which eventually generates dense pixel-level labels. However, in spite of a robust approach, there are limitations in the classical FCN model for building extraction from RS images:
1. RS images are usually high-resolution with rich contextual semantic information, while the classical classification network is not sufficient for mining global contextual information.
2. CNNs do well in mining local features, but not in modeling long-distance association information. It is difficult for the plain decoder structure to reconstruct the structured hierarchical detail information, such as building boundaries and contours, which is lost due to the decrease of feature map resolution caused by the encoder downsampling.
3. The RS images are informative, so the processing of building extraction problem should focus on the model operation efficiency while ensuring the segmentation accuracy.
Table 1 shows the main methods involved here, containing the architectures involved in the methods, their main contributions, and a hierarchy based on their task objectives: accuracy (ACC) and reusability (Reu) of the model structure. Specifically, Reu indicates whether the advanced network modules proposed in the literature can be reused relatively easily by other partitioned networks or studies. Each objective is divided into three grades, depending on the degree of focus of the corresponding work on that objective. From the perspective of accuracy, aggregating multi-scale contextual information, considering boundary information, iterative refinement and adopting appropriate post-processing strategies are aspects to be considered. Network components with good reusability are usually robust while not producing large changes in the size of the input and output feature maps, such as attention modules. Moreover, they are usually designed with the objective of aggregating certain elements that are useful for accomplishing the target task.
Table 1. Deep learning-based methods on building extraction.

Methods

Acc.

Reu.

104

]

*

**

Multi-scale structure LinkNet

MA-FCN

[

105

]

*

**

Boundary constraints, multiscale prediction

GAN-SCA [106]

**

***

SCA, GAN

HFSA-Unet [107]

***

***

Two-stage channel attention

ESFNet [108]

*

***

Separable factorized residual block

ACR-Net [109]

**

***

RBAC

SegNet-Dist-Fused [110]

*

*

Signed distance, multi-source data fusion

CFCN [111]

**

**

Boundary constraint networks

4.1. Baseline Methods

FCN, SegNet, and U-Net all employ an encoder-decoder architecture, but offer different aspects of design mindsets that are reflected in the encoder, upsampling, and skip connection, respectively.
  • The encoders for FCN and SegNet are usually obtained by removing fully connected layers using classification networks such as VGG-16 and ResNet, and the encoder for U-Net is designed to be symmetric with the decoder, allowing the depth of the network to be increased or decreased depending on the complexity of the task.
  • The decoder structure of FCN is the simplest and contains only one deconvolution operation, while U-Net and SegNet adopt multiple upsampling to organize the decoder structure.
  • There is a feature fusion by FCN with feature maps organized by pixel-by-pixel summing, U-Net with feature map stitching, and SegNet with pooling indices generated by pooling operation embedded in the decoder feature map to solve the problem of insufficient recovery information in the upsampling process.
These three types of basic methods have been applied to the building extraction problem since a few years ago [3][83][84][86][87][88][89], which have recently been used mainly as a baseline to motivate new methods and to compare their effectiveness.

4.2. Contextual Information Mining

The key points of building extraction are mining local information (short-distance contextual information around pixels such as building outline and boundary) and global information (long-distance contextual information between buildings and background and overall association relationship between buildings and buildings with other pixels in the image). Rich local information helps to improve the accuracy of pixel-level annotation, while complete global information is also essential to resolve local blur. It is the concern of all DL-based building extraction methods to balance and fuse these two aspects.

4.3. Lightweight Network Design

Building extraction is usually performed in high-resolution images so that the design of segmentation networks has to take into account the consumption of computational resources such as GPU memory and the inference speed of the prediction phase. However, most existing methods usually require a large number of parameters and floating-point operations to obtain high accuracy, which leads to high computational resource consumption and low inference speed.
In order to achieve a better balance between accuracy and efficiency, a common approach is to apply an existing lightweight network or adopt a more efficient convolutional module to develop a lightweight network as a feature extraction network [112][113][114]. Lin et al. [108] and Liu et al. [109] developed new feature extraction backbone networks with deep separable convolutional asymmetric convolution respectively, incorporating decoder networks to achieve segmentation results with accuracy no less than mainstream networks such as U-Net, SegNet, and earlier lightweight networks such as ENet [115], with significantly lower number of parameters and computational effort.

4.4. Multi-Source Data

Extracting buildings from RGB images is currently the most widely used method. However, DSM elevation data and LiDAR data are widely used as auxiliary data to correct the angle of the building and pinpoint the location of the building to improve the accuracy of building segmentation. In other words, RGB data provides extensive background color information and building shape information, while DSM elevation data and LIDAR data provide accurate relative position information and three-dimensional spatial information. The fusion of RGB images with data from other sources exists in two main stages, the pre-processing stage before input to the network and the post-processing stage of the network.
Data fusion in the previous stage typically involves attaching multi-source data such as DSM as an additional channel to RGB images to form multi-channel data [110][111]. However, the approach of direct fusion of data ignores the variability of different data sources. Huang et al. [2] utilizes independent FCNs to provide segmentation results based on data from different data sources and performs feature fusion at the final layer with confidence votes to obtain the results. Bittner et al. [85] employs multiple mutually independent encoder networks for feature extraction from multiple sources of data separately, and the segmentation results are derived with a decoder after fusing the features.

References

  1. Li, E.; Femiani, J.; Xu, S.; Zhang, X.; Wonka, P. Robust Rooftop Extraction From Visible Band Images Using Higher Order CRF. IEEE Trans. Geosci. Remote Sens. 2015, 53, 4483–4495.
  2. Huang, Z.; Cheng, G.; Wang, H.; Li, H.; Shi, L.; Pan, C. Building extraction from multi-source remote sensing images via deep deconvolution neural networks. In Proceedings of the 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Beijing, China, 10–16 June 2016; pp. 1835–1838.
  3. Bittner, K.; Cui, S.; Reinartz, P. Building Extraction from Remote Sensing Data Using Fully Convolutional Networks. In Proceedings of the International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Hannover, Germany, 6–9 June 2017; pp. 481–486.
  4. Xu, Y.; Wu, L.; Xie, Z.; Chen, Z. Building Extraction in Very High Resolution Remote Sensing Imagery Using Deep Learning and Guided Filters. Remote Sens. 2018, 10, 144.
  5. Maltezos, E.; Doulamis, A.; Doulamis, N.; Ioannidis, C. Building Extraction From LiDAR Data Applying Deep Convolutional Neural Networks. IEEE Geosci. Remote Sens. Lett. 2019, 16, 155–159.
  6. Sun, G.; Huang, H.; Zhang, A.; Li, F.; Zhao, H.; Fu, H. Fusion of Multiscale Convolutional Neural Networks for Building Extraction in Very High-Resolution Images. Remote Sens. 2019, 11, 227.
  7. Wu, G.; Guo, Z.; Shao, X.; Shibasaki, R. GEOSEG: A Computer Vision Package for Automatic Building Segmentation and Outline Extraction. In Proceedings of the 2019 IEEE International Geoscience and Remote Sensing Symposium, Yokohama, Japan, 28 July–2 August 2019; pp. 158–161.
  8. Huang, J.; Zhang, X.; Xin, Q.; Sun, Y.; Zhang, P. Automatic building extraction from high-resolution aerial images and LiDAR data using gated residual refinement network. ISPRS J. Photogramm. Remote Sens. 2019, 151, 91–105.
  9. Li, W.; He, C.; Fang, J.; Zheng, J.; Fu, H.; Yu, L. Semantic Segmentation-Based Building Footprint Extraction Using Very High-Resolution Satellite Images and Multi-Source GIS Data. Remote Sens. 2019, 11, 403.
  10. Zhang, Y.; Gong, W.; Sun, J.; Li, W. Web-Net: A Novel Nest Networks with Ultra-Hierarchical Sampling for Building Extraction from Aerial Imageries. Remote Sens. 2019, 11, 1897.
  11. Davari Majd, R.; Momeni, M.; Moallem, P. Transferable Object-Based Framework Based on Deep Convolutional Neural Networks for Building Extraction. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2019, 12, 2627–2635.
  12. Zhang, Z.; Wang, Y. JointNet: A Common Neural Network for Road and Building Extraction. Remote Sens. 2019, 11, 696.
  13. Chen, S.; Shi, W.; Zhou, M.; Zhang, M.; Chen, P. Automatic Building Extraction via Adaptive Iterative Segmentation With LiDAR Data and High Spatial Resolution Imagery Fusion. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2020, 13, 2081–2095.
  14. Erdem, F.; Avdan, U. Comparison of different U-net models for building extraction from high-resolution aerial imagery. Int. J. Environ. GeoInform. 2020, 7, 221–227.
  15. Milosavljevi, A. Automated Processing of Remote Sensing Imagery Using Deep Semantic Segmentation: A Building Footprint Extraction Case. ISPRS Int. J. Geo-Inf. 2020, 9, 486.
  16. Long, J.; Shelhamer, E.; Darrell, T. Fully Convolutional Networks for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Boston, MA, USA, 7–12 June 2015.
  17. Badrinarayanan, V.; Kendall, A.; Cipolla, R. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. IEEE Trans. Pattern Anal. Mach. Intell. 2017, 39, 2481–2495.
  18. Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Proceedings of the Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015, Munich, Germany, 5–9 October 2015; Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F., Eds.; Springer International Publishing: Cham, Switzerland, 2015; pp. 234–241.
  19. Zheng, S.; Jayasumana, S.; Romera-Paredes, B.; Vineet, V.; Su, Z.; Du, D.; Huang, C.; Torr, P.H.S. Conditional Random Fields as Recurrent Neural Networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, 10–13 December 2015.
  20. Peng, C.; Zhang, X.; Yu, G.; Luo, G.; Sun, J. Large Kernel Matters—Improve Semantic Segmentation by Global Convolutional Network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017.
  21. Chen, L.C.; Papandreou, G.; Kokkinos, I.; Murphy, K.; Yuille, A.L. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv 2014, arXiv:1412.7062.
  22. Chen, L.C.; Papandreou, G.; Kokkinos, I.; Murphy, K.; Yuille, A.L. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE Trans. Pattern Anal. Mach. Intell. 2018, 40, 834–848.
  23. Chen, L.C.; Papandreou, G.; Schroff, F.; Adam, H. Rethinking atrous convolution for semantic image segmentation. arXiv 2017, arXiv:1706.05587.
  24. Chen, L.C.; Zhu, Y.; Papandreou, G.; Schroff, F.; Adam, H. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany, 8–14 September 2018.
  25. Wang, P.; Chen, P.; Yuan, Y.; Liu, D.; Huang, Z.; Hou, X.; Cottrell, G. Understanding Convolution for Semantic Segmentation. In Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Lake Tahoe, NV, USA, 12–15 March 2018; pp. 1451–1460.
  26. Lin, G.; Milan, A.; Shen, C.; Reid, I. RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017.
  27. Chen, J.; Lu, Y.; Yu, Q.; Luo, X.; Adeli, E.; Wang, Y.; Lu, L.; Yuille, A.L.; Zhou, Y. Transunet: Transformers make strong encoders for medical image segmentation. arXiv 2021, arXiv:2102.04306.
  28. Zheng, S.; Lu, J.; Zhao, H.; Zhu, X.; Luo, Z.; Wang, Y.; Fu, Y.; Feng, J.; Xiang, T.; Torr, P.H.; et al. Rethinking Semantic Segmentation From a Sequence-to-Sequence Perspective With Transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, TN, USA, 19–25 June 2021; pp. 6881–6890.
  29. Strudel, R.; Garcia, R.; Laptev, I.; Schmid, C. Segmenter: Transformer for Semantic Segmentation. arXiv 2021, arXiv:2105.05633.
  30. Ning, F.; Delhomme, D.; LeCun, Y.; Piano, F.; Bottou, L.; Barbano, P. Toward automatic phenotyping of developing embryos from videos. IEEE Trans. Image Process. 2005, 14, 1360–1371.
  31. Ciresan, D.; Giusti, A.; Gambardella, L.; Schmidhuber, J. Deep neural networks segment neuronal membranes in electron microscopy images. Adv. Neural Inf. Process. Syst. 2012, 25, 2843–2851.
  32. Farabet, C.; Couprie, C.; Najman, L.; LeCun, Y. Learning Hierarchical Features for Scene Labeling. IEEE Trans. Pattern Anal. Mach. Intell. 2013, 35, 1915–1929.
  33. Hariharan, B.; Arbeláez, P.; Girshick, R.; Malik, J. Simultaneous Detection and Segmentation. In Proceedings of the Computer Vision—ECCV 2014, Zurich, Switzerland, 6–12 September 2014; Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T., Eds.; Springer International Publishing: Cham, Switzerland, 2014; pp. 297–312.
  34. Gupta, S.; Girshick, R.; Arbeláez, P.; Malik, J. Learning Rich Features from RGB-D Images for Object Detection and Segmentation. In Proceedings of the Computer Vision—ECCV 2014, Zurich, Switzerland, 6–12 September 2014; Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T., Eds.; Springer International Publishing: Cham, Switzerland, 2014; pp. 345–360.
  35. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. Adv. Neural Inf. Process. Syst. 2012, 25, 1097–1105.
  36. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. Adv. Neural Inf. Process. Syst. 2015, 28, 91–99.
  37. Xie, S.; Tu, Z. Holistically-nested edge detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, 10–13 December 2015; pp. 1395–1403.
  38. Li, Y.; Qi, H.; Dai, J.; Ji, X.; Wei, Y. Fully convolutional instance-aware semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017; pp. 2359–2367.
  39. Feng, T.; Zhao, J. Review and Comparison: Building Extraction Methods Using High-Resolution Images. In Proceedings of the 2009 Second International Symposium on Information Science and Engineering, Shanghai, China, 26–28 December 2009; pp. 419–422.
  40. Jozdani, S.; Chen, D. On the versatility of popular and recently proposed supervised evaluation metrics for segmentation quality of remotely sensed images: An experimental case study of building extraction. ISPRS J. Photogramm. Remote Sens. 2020, 160, 275–290.
  41. Bo, Z.; Chao, W.; Hong, Z.; Fan, W. A review on building extraction and Reconstruction from SAR image. Remote Sens. Technol. Appl. 2012, 4.
  42. Mishra, A.; Pandey, A.; Baghel, A.S. Building detection and extraction techniques: A review. In Proceedings of the 2016 3rd International Conference on Computing for Sustainable Global Development (INDIACom), New Delhi, India, 16–18 May 2016; pp. 3816–3821.
  43. Zhu, X.X.; Tuia, D.; Mou, L.; Xia, G.S.; Zhang, L.; Xu, F.; Fraundorfer, F. Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources. IEEE Geosci. Remote Sens. Mag. 2017, 5, 8–36.
  44. Zhang, L.; Zhang, L.; Du, B. Deep Learning for Remote Sensing Data: A Technical Tutorial on the State of the Art. IEEE Geosci. Remote Sens. Mag. 2016, 4, 22–40.
  45. Huang, B.; Reichman, D.; Collins, L.M.; Bradbury, K.; Malof, J.M. Dense labeling of large remote sensing imagery with convolutional neural networks: A simple and faster alternative to stitching output label maps. arXiv 2018, arXiv:1805.12219.
  46. Clark, R.N.; Roush, T.L. Reflectance spectroscopy: Quantitative analysis techniques for remote sensing applications. J. Geophys. Res. Solid Earth 1984, 89, 6329–6340.
  47. Zhou, G.; Song, C.; Simmers, J.; Cheng, P. Urban 3D GIS From LiDAR and digital aerial images. Comput. Geosci. 2004, 30, 345–353.
  48. Tang, J.; Wang, L.; Yao, Z. Analyzing urban sprawl spatial fragmentation using multi-temporal satellite images. Giscience Remote Sens. 2006, 43, 218–232.
  49. Wu, S.s.; Qiu, X.; Wang, L. Population estimation methods in GIS and remote sensing: A review. Giscience Remote Sens. 2005, 42, 80–96.
  50. Tian, J.; Cui, S.; Reinartz, P. Building Change Detection Based on Satellite Stereo Imagery and Digital Surface Models. IEEE Trans. Geosci. Remote Sens. 2014, 52, 406–417.
  51. Montoya-Zegarra, J.A.; Wegner, J.D.; Schindler, K. Semantic segmentation of aerial images in urban areas with class-specific higher-order cliques. ISPRS Ann. Photogramm. Remote Sens. Spat. Inf. Sci. 2015, 2, 127–133.
  52. Grinias, I.; Panagiotakis, C.; Tziritas, G. MRF-based segmentation and unsupervised classification for building and road detection in peri-urban areas of high-resolution satellite images. ISPRS J. Photogramm. Remote Sens. 2016, 122, 145–166.
  53. Liu, Y.; Li, Z.; Wei, B.; Li, X.; Fu, B. Seismic vulnerability assessment at urban scale using data mining and GIScience technology: Application to Urumqi (China). Geomat. Nat. Hazards Risk 2019, 10, 958–985.
  54. Li, X.; Li, Z.; Yang, J.; Liu, Y.; Fu, B.; Qi, W.; Fan, X. Spatiotemporal characteristics of earthquake disaster losses in China from 1993 to 2016. Nat. Hazards 2018, 94, 843–865.
  55. Zhang, B.; Chen, Z.; Peng, D.; Benediktsson, J.A.; Liu, B.; Zou, L.; Li, J.; Plaza, A. Remotely sensed big data: Evolution in model development for information extraction . Proc. IEEE 2019, 107, 2294–2301.
  56. Liu, Y.; So, E.; Li, Z.; Su, G.; Gross, L.; Li, X.; Qi, W.; Yang, F.; Fu, B.; Yalikun, A.; et al. Scenario-based seismic vulnerability and hazard analyses to help direct disaster risk reduction in rural Weinan, China. Int. J. Disaster Risk Reduct. 2020, 48, 101577.
  57. Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv 2014, arXiv:1409.1556.
  58. Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.E.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; Rabinovich, A. Going Deeper with Convolutions. arXiv 2014, arXiv:1409.4842.
  59. Ioffe, S.; Szegedy, C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In Proceedings of the 32nd International Conference on Machine Learning, Lille, France, 7–9 July 2015; Bach, F., Blei, D., Eds.; PMLR: Lille, France, 2015; Volume 37, pp. 448–456.
  60. Szegedy, C.; Vanhoucke, V.; Ioffe, S.; Shlens, J.; Wojna, Z. Rethinking the Inception Architecture for Computer Vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 26–30 June 2016.
  61. Szegedy, C.; Ioffe, S.; Vanhoucke, V.; Alemi, A.A. Inception-v4, inception-resnet and the impact of residual connections on learning. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, San Francisco, CA, USA, 4–9 February 2017.
  62. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 26–30 June 2016.
  63. Ahmed, A.; Yu, K.; Xu, W.; Gong, Y.; Xing, E. Training Hierarchical Feed-Forward Visual Recognition Models Using Transfer Learning from Pseudo-Tasks. In Proceedings of the Computer Vision—ECCV 2008, Marseille, France, 12–18 October 2008; Forsyth, D., Torr, P., Zisserman, A., Eds.; Springer: Berlin/Heidelberg, Germany, 2008; pp. 69–82.
  64. Oquab, M.; Bottou, L.; Laptev, I.; Sivic, J. Learning and Transferring Mid-Level Image Representations using Convolutional Neural Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Columbus, OH, USA, 23–28 June 2014.
  65. Yosinski, J.; Clune, J.; Bengio, Y.; Lipson, H. How transferable are features in deep neural networks? arXiv 2014, arXiv:1411.1792.
  66. Deng, J.; Dong, W.; Socher, R.; Li, L.J.; Li, K.; Fei-Fei, L. ImageNet: A large-scale hierarchical image database. In Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 20–25 June 2009; pp. 248–255.
  67. Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; et al. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 2015, 115, 211–252.
  68. Kaiser, P.; Wegner, J.D.; Lucchi, A.; Jaggi, M.; Hofmann, T.; Schindler, K. Learning Aerial Image Segmentation From Online Maps. IEEE Trans. Geosci. Remote Sens. 2017, 55, 6054–6068.
  69. Kampffmeyer, M.; Salberg, A.B.; Jenssen, R. Semantic Segmentation of Small Objects and Modeling of Uncertainty in Urban Remote Sensing Images Using Deep Convolutional Neural Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, Las Vegas, NV, USA, 26 June–1 July 2016.
  70. Eigen, D.; Fergus, R. Predicting Depth, Surface Normals and Semantic Labels With a Common Multi-Scale Convolutional Architecture. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Sydney, Australia, 1–8 December 2016.
  71. Li, Q.; Shi, Y.; Huang, X.; Zhu, X.X. Building Footprint Generation by Integrating Convolution Neural Network With Feature Pairwise Conditional Random Field (FPCRF). IEEE Trans. Geosci. Remote Sens. 2020, 58, 7502–7519.
  72. Hosseinpoor, H.; Samadzadegan, F. Convolutional Neural Network for Building Extraction from High-Resolution Remote Sensing Images. In Proceedings of the 2020 International Conference on Machine Vision and Image Processing (MVIP), Qom, Iran, 18–20 February 2020; pp. 1–5.
  73. Yu, Y.; Ren, Y.; Guan, H.; Li, D.; Yu, C.; Jin, S.; Wang, L. Capsule feature pyramid network for building footprint extraction from high-resolution aerial imagery. IEEE Geosci. Remote Sens. Lett. 2020, 18, 895–899.
  74. Deng, W.; Shi, Q.; Li, J. Attention-Gate-Based Encoder–Decoder Network for Automatical Building Extraction. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 2611–2620.
  75. Guo, H.; Shi, Q.; Du, B.; Zhang, L.; Wang, D.; Ding, H. Scene-driven multitask parallel attention network for building extraction in high-resolution remote sensing images. IEEE Trans. Geosci. Remote Sens. 2020, 59, 4287–4306.
  76. Zhu, Q.; Liao, C.; Hu, H.; Mei, X.; Li, H. MAP-Net: Multiple Attending Path Neural Network for Building Footprint Extraction From Remote Sensed Imagery. IEEE Trans. Geosci. Remote Sens. 2021, 59, 6169–6181.
  77. Hu, Q.; Zhen, L.; Mao, Y.; Zhou, X.; Zhou, G. Automated building extraction using satellite remote sensing imagery. Autom. Constr. 2021, 123, 103509.
  78. Mnih, V. Machine Learning for Aerial Image Labeling. Ph.D. Thesis, University of Toronto, Toronto, ON, Canada, 2013.
  79. Maggiori, E.; Tarabalka, Y.; Charpiat, G.; Alliez, P. Can semantic labeling methods generalize to any city? The inria aerial image labeling benchmark. In Proceedings of the 2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Worth, TX, USA, 23–28 June 2017; pp. 3226–3229.
  80. Ji, S.; Wei, S.; Lu, M. Fully Convolutional Networks for Multisource Building Extraction From an Open Aerial and Satellite Imagery Data Set. IEEE Trans. Geosci. Remote Sens. 2019, 57, 574–586.
  81. Saito, S.; Yamashita, T.; Aoki, Y. Multiple object extraction from aerial imagery with convolutional neural networks. Electron. Imaging 2016, 2016, 1–9.
  82. Alshehhi, R.; Marpu, P.R.; Woon, W.L.; Mura, M.D. Simultaneous extraction of roads and buildings in remote sensing imagery with convolutional neural networks. ISPRS J. Photogramm. Remote Sens. 2017, 130, 139–149.
  83. Maggiori, E.; Tarabalka, Y.; Charpiat, G.; Alliez, P. Fully convolutional neural networks for remote sensing image classification. In Proceedings of the 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Beijing, China, 10–16 June 2016; pp. 5071–5074.
  84. Yuan, J. Automatic Building Extraction in Aerial Scenes Using Convolutional Networks. arXiv 2016, arXiv:1602.06564.
  85. Bittner, K.; Adam, F.; Cui, S.; K?rner, M.; Reinartz, P. Building Footprint Extraction From VHR Remote Sensing Images Combined With Normalized DSMs Using Fused Fully Convolutional Networks. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2018, 11, 2615–2629.
  86. Yang, H.L.; Lunga, D.; Yuan, J. Toward country scale building detection with convolutional neural network using aerial images. In Proceedings of the 2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Worth, TX, USA, 23–28 June 2017; pp. 870–873.
  87. Wu, G.; Shao, X.; Guo, Z.; Chen, Q.; Yuan, W.; Shi, X.; Xu, Y.; Shibasaki, R. Automatic building segmentation of aerial imagery using multi-constraint fully convolutional networks. Remote Sens. 2018, 10, 407.
  88. Li, L.; Liang, J.; Weng, M.; Zhu, H. A multiple-feature reuse network to extract buildings from remote sensing imagery. Remote Sens. 2018, 10, 1350.
  89. Wu, G.; Guo, Z.; Shi, X.; Chen, Q.; Xu, Y.; Shibasaki, R.; Shao, X. A boundary regulated network for accurate roof segmentation and outline extraction. Remote Sens. 2018, 10, 1195.
  90. Ma, J.; Wu, L.; Tang, X.; Liu, F.; Zhang, X.; Jiao, L. Building Extraction of Aerial Images by a Global and Multi-Scale Encoder-Decoder Network. Remote Sens. 2020, 12, 2350.
  91. Wang, S.; Hou, X.; Zhao, X. Automatic Building Extraction From High-Resolution Aerial Imagery via Fully Convolutional Encoder-Decoder Network With Non-Local Block. IEEE Access 2020, 8, 7313–7322.
  92. Zhou, D.; Wang, G.; He, G.; Long, T.; Yin, R.; Zhang, Z.; Chen, S.; Luo, B. Robust Building Extraction for High Spatial Resolution Remote Sensing Images with Self-Attention Network. Sensors 2020, 20, 7241.
  93. Shrestha, S.; Vanneschi, L. Improved Fully Convolutional Network with Conditional Random Fields for Building Extraction. Remote Sens. 2018, 10, 1135.
  94. Maggiori, E.; Charpiat, G.; Tarabalka, Y.; Alliez, P. Recurrent Neural Networks to Correct Satellite Image Classification Maps. IEEE Trans. Geosci. Remote Sens. 2017, 55, 4962–4971.
  95. Yang, G.; Zhang, Q.; Zhang, G. EANet: Edge-Aware Network for the Extraction of Buildings from Aerial Images. Remote Sens. 2020, 12, 2161.
  96. Zhang, Y.; Li, W.; Gong, W.; Wang, Z.; Sun, J. An Improved Boundary-Aware Perceptual Loss for Building Extraction from VHR Images. Remote Sens. 2020, 12, 1195.
  97. Shao, Z.; Tang, P.; Wang, Z.; Saleem, N.; Yam, S.; Sommai, C. BRRNet: A Fully Convolutional Neural Network for Automatic Building Extraction From High-Resolution Remote Sensing Images. Remote Sens. 2020, 12, 1050.
  98. Shi, Y.; Li, Q.; Zhu, X.X. Building segmentation through a gated graph convolutional neural network with deep structured feature embedding. ISPRS J. Photogramm. Remote Sens. 2020, 159, 184–197.
  99. Hamaguchi, R.; Fujita, A.; Nemoto, K.; Imaizumi, T.; Hikosaka, S. Effective Use of Dilated Convolutions for Segmenting Small Object Instances in Remote Sensing Imagery. In Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Lake Tahoe, NV, USA, 12–15 March 2018; pp. 1442–1450.
  100. Kang, W.; Xiang, Y.; Wang, F.; You, H. EU-Net: An Efficient Fully Convolutional Network for Building Extraction from Optical Remote Sensing Images. Remote Sens. 2019, 11, 2813.
  101. Liu, Y.; Fan, B.; Wang, L.; Bai, J.; Xiang, S.; Pan, C. Semantic labeling in very high resolution images via a self-cascaded convolutional neural network. ISPRS J. Photogramm. Remote Sens. 2018, 145, 78–95.
  102. Ji, S.; Wei, S.; Lu, M. A scale robust convolutional neural network for automatic building extraction from aerial and satellite imagery. Int. J. Remote Sens. 2019, 40, 3308–3322.
  103. Li, X.; Yao, X.; Fang, Y. Building-A-Nets: Robust Building Extraction From High-Resolution Remote Sensing Images With Adversarial Networks. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2018, 11, 3680–3687.
  104. Ding, Y.; Wu, M.; Xu, Y.; Duan, S. P-linknet: Linknet with spatial pyramid pooling for high-resolution satellite imagery. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2020, 43, 35–40.
  105. Wei, S.; Ji, S.; Lu, M. Toward Automatic Building Footprint Delineation From Aerial Images Using CNN and Regularization. IEEE Trans. Geosci. Remote Sens. 2020, 58, 2178–2189.
  106. Pan, X.; Yang, F.; Gao, L.; Chen, Z.; Zhang, B.; Fan, H.; Ren, J. Building Extraction from High-Resolution Aerial Imagery Using a Generative Adversarial Network with Spatial and Channel Attention Mechanisms. Remote Sens. 2019, 11, 917.
  107. He, N.; Fang, L.; Plaza, A. Hybrid first and second order attention Unet for building segmentation in remote sensing images. Sci. China Inf. Sci. 2020, 63, 1–12.
  108. Lin, J.; Jing, W.; Song, H.; Chen, G. ESFNet: Efficient Network for Building Extraction From High-Resolution Aerial Images. IEEE Access 2019, 7, 54285–54294.
  109. Liu, Y.; Zhou, J.; Qi, W.; Li, X.; Gross, L.; Shao, Q.; Zhao, Z.; Ni, L.; Fan, X.; Li, Z. ARC-Net: An Efficient Network for Building Extraction From High-Resolution Aerial Images. IEEE Access 2020, 8, 154997–155010.
  110. Yang, H.L.; Yuan, J.; Lunga, D.; Laverdiere, M.; Rose, A.; Bhaduri, B. Building Extraction at Scale Using Convolutional Neural Network: Mapping of the United States. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2018, 11, 2600–2614.
  111. Liu, W.; Yang, M.; Xie, M.; Guo, Z.; Li, E.; Zhang, L.; Pei, T.; Wang, D. Accurate Building Extraction from Fused DSM and UAV Images Using a Chain Fully Convolutional Neural Network. Remote Sens. 2019, 11, 2912.
  112. Anwar, S.; Hwang, K.; Sung, W. Structured Pruning of Deep Convolutional Neural Networks. ACM J. Emerg. Technol. Comput. Syst. 2015, 13.
  113. Han, S.; Mao, H.; Dally, W.J. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv 2015, arXiv:1510.00149.
  114. Molchanov, P.; Tyree, S.; Karras, T.; Aila, T.; Kautz, J. Pruning convolutional neural networks for resource efficient transfer learning. arXiv 2016, arXiv:1611.06440.
  115. Paszke, A.; Chaurasia, A.; Kim, S.; Culurciello, E. Enet: A deep neural network architecture for real-time semantic segmentation. arXiv 2016, arXiv:1606.02147.
More
© Text is available under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/)
ScholarVision Creations
Feedback