Potsdam Dataset Research Articles

A Semisupervised Convolution Neural Network for Partial Unlabeled Remote-Sensing Image Segmentation

Semantic segmentation methods for remote-sensing images based on the deep learning framework have achieved significant performance improvements. However, most of the existing work is based on fully supervised methods, which rely on a large number of manually annotated pixel-level labels. However, for remote-sensing images, labeling the ground-truth takes time and effort. To solve the problem in existing methods of overly relying on manual labeling, in this study, we propose a semisupervised convolution neural network based on contrastive loss for partial unlabeled remote-sensing image segmentation. In the design of the contrastive loss function, to capture the semantic relationship of pixels and improve the separability between different categories, we propose pixel-level and region-level contrastive loss. The pixel-level contrastive loss is designed to learn the correlation between different images, while region-level contrastive loss is designed to improve the quality of generated pseudo-labels. In addition, we designed a propagated self-training method that further guarantees the quality of the pseudo-labels and improves the richness of the labeled data. Experiments on POTSDAM and Vaihingen datasets demonstrate that the proposed method achieves the highest Mean Intersection over Union (mIOU) and significantly outperforms previous methods.

Geometry-Aware Segmentation of Remote Sensing Images via Joint Height Estimation

Recent studies have shown the benefits of using additional elevation data [e.g., digital surface model (DSM) or normalized DSM (nDSM)] for enhancing the performance of the semantic labeling of aerial images. However, previous methods mostly adopt 3-D elevation information as additional inputs, while, in many real-world applications, one does not have the corresponding DSM images at hand, and the spatial resolution of acquired DSM images usually does not match the aerial images. To alleviate this data constraint and also take advantage of 3-D elevation information, in this letter, a geometry-aware segmentation model is introduced to achieve accurate semantic labeling of aerial images via joint height estimation. Instead of using a single-stream encoder-decoder network for semantic labeling, we design a separate decoder branch to predict the height map and use the DSM images as side supervision to train this newly designed decoder branch. With the newly designed decoder branch, our model can distill the 3-D geometric features from 2-D appearance features under the supervision of ground-truth DSM images. Moreover, we develop a new geometry-aware convolution module that fuses the 3-D geometric features from the height decoder branch and the 2-D contextual features from the semantic segmentation branch. The fused feature embeddings can produce geometry-aware segmentation maps with enhanced performance. Our model is trained with DSM images as side supervision, while, in the inference stage, it does not require DSM data and directly predicts the semantic labels. Experiments on International Society for Photogrammetry and Remote Sensing (ISPRS) Vaihingen and Potsdam data sets demonstrate the effectiveness of the proposed method for the semantic segmentation of aerial images.

Feature-Selection High-Resolution Network With Hypersphere Embedding for Semantic Segmentation of VHR Remote Sensing Images

Very high resolution (VHR) remote sensing images contain various multi-scale objects, such as large-scale buildings and small-scale cars. However, these multi-scale objects cannot be considered simultaneously in the widely used backbones with large downsampling factor (e.g. VGG-like and ResNet-like), resulting in the appearance of various context aggregation approaches such as fusing low-level features and attention-based modules. To alleviate this problem caused by backbones with large downsampling factor, we propose a feature-selection high-resolution network (FSHRNet) based on an observation: if the features maintain high resolution throughout the network, a high precision segmentation result can be obtained by only using a 1×1 convolution layer with no need for complex context aggregation modules. Specifically, the backbone of FSHRNet is a multi-branch structure similar to HRNet where the high-resolution branch is the principal line. Then, a lightweight dynamic weight module, named feature-selection convolution layer (FSConv), is presented to fuse multi-resolution features, allowing adaptively feature selection based on the characteristic of objects. Finally, a specially designed 1×1 convolution layer derived from hypersphere embedding is used to produce the segmentation result. Experiments with other widely used methods show that the proposed FSHRNet obtains competitive performance on ISPRS Vaihingen dataset, ISPRS Potsdam dataset and iSAID dataset.

DHT: Deformable Hybrid Transformer for Aerial Image Segmentation

Due to the strong ability to model global information, the transformer-based methods have shown remarkable improvements in image segmentation tasks. However, the self-attention mechanism in the transformer is computationally expensive and relies on pre-trained parameters. Moreover, the transformer method is weak in modeling local information, which is unfavorable for accurately segmenting objects from high-resolution aerial images. To this end, an efficient deformable orientational self-attention (DoA) is proposed to simultaneously extract the global information and the local information. Besides, for parameter efficiency, we design a depthwise channel self-attention (DcA) to model the contextual information among channels. Combining with the DoA and DcA, we propose the deformable hybrid transformer (DHT) to perform high-quality object segmentation on aerial images. Experiments on ISPRS Potsdam dataset and WHU building dataset illustrate that the proposed DHT can not only achieve state-of-the-art (SOTA) results but also markedly reduce the dependence of the transformer on pre-trained parameters.

FALSE: False Negative Samples Aware Contrastive Learning for Semantic Segmentation of High-Resolution Remote Sensing Image

The existing SSCL of RSI is built based on constructing positive and negative sample pairs. However, due to the richness of RSI ground objects and the complexity of the RSI contextual semantics, the same RSI patches have the coexistence and imbalance of positive and negative samples, which causing the SSCL pushing negative samples far away while pushing positive samples far away, and vice versa. We call this the sample confounding issue (SCI). To solve this problem, we propose a False negAtive sampLes aware contraStive lEarning model (FALSE) for the semantic segmentation of high-resolution RSIs. Since the SSCL pretraining is unsupervised, the lack of definable criteria for false negative sample (FNS) leads to theoretical undecidability, we designed two steps to implement the FNS approximation determination: coarse determination of FNS and precise calibration of FNS. We achieve coarse determination of FNS by the FNS self-determination (FNSD) strategy and achieve calibration of FNS by the FNS confidence calibration (FNCC) loss function. Experimental results on three RSI semantic segmentation datasets demonstrated that the FALSE effectively improves the accuracy of the downstream RSI semantic segmentation task compared with the current three models, which represent three different types of SSCL models. The mean Intersection-over-Union on ISPRS Potsdam dataset is improved by 0.7\% on average; on CVPR DGLC dataset is improved by 12.28\% on average; and on Xiangtan dataset this is improved by 1.17\% on average. This indicates that the SSCL model has the ability to self-differentiate FNS and that the FALSE effectively mitigates the SCI in self-supervised contrastive learning. The source code is available at https://github.com/GeoX-Lab/FALSE.

Semantic Segmentation of Remote Sensing Images With Self-Supervised Semantic-Aware Inpainting

Semantic segmentation of remote sensing imageries plays a crucial role in resource exploration, urban planning, weather forecasting, etc. For this task, deep learning-based methods have shown significant achievement, typically trained with large-scale labeled data. However, these methods often suffer the performance deterioration facing limited labeled data in real-world applications. To address this problem, a novel self-supervised semantic segmentation framework is proposed for remote sensing imageries with limited labeled data. Specifically, image inpainting is acted as pixel-level pretext task for learning dense feature representations suitable for semantic segmentation. Further, rather than trivially leveraging the conventional random inpainting strategy, a novel adversarial training scheme is proposed to drive the pretext task to adaptively mask and restore salient local regions. The adversarial training scheme consists of instructor network and inpainting network, the instructor network increasingly predicts meaningful salient regions as erased regions, and meanwhile the inpainting network seeks for restoring the corrupted image as pretext task to learn its intrinsic representation. Moreover, the structural similarity (SSIM) is applied as a patch-level loss function for semantic segmentation considering that remote sensing images are highly structured. The experimental results on the ISPRS Potsdam dataset demonstrate that our method outperforms state-of-the-art self-supervised methods and the ImageNet pre-training methods. The source code is available at https://github.com/JasmineBJTU/self-supervised_RSSS.

Class-Guided Swin Transformer for Semantic Segmentation of Remote Sensing Imagery

Semantic segmentation of remote sensing images plays a crucial role in a wide variety of practical applications, including land cover mapping, environmental protection, and economic assessment. In the last decade, convolutional neural network (CNN) is the mainstream deep learning-based method of semantic segmentation. Compared with conventional methods, CNN-based methods learn semantic features automatically, thereby achieving strong representation capability. However, the local receptive field of the convolution operation limits CNN-based methods from capturing long-range dependencies. In contrast, Vision Transformer (ViT) demonstrates its great potential in modeling long-range dependencies and obtains superior results in semantic segmentation. Inspired by this, in this letter, we propose a class-guided Swin Transformer (CG-Swin) for semantic segmentation of remote sensing images. Specifically, we adopt a Transformer-based encoder–decoder structure, which introduces the Swin Transformer backbone as the encoder and designs a class-guided Transformer block to construct the decoder. The experimental results on ISPRS Vaihingen and Potsdam datasets demonstrate the significant breakthrough of the proposed method over ten benchmarks, outperforming both advanced CNN-based and recent Transformer-based approaches.

Global and Local Contrastive Self-Supervised Learning for Semantic Segmentation of HR Remote Sensing Images

Supervised learning for semantic segmentation requires a large number of labeled samples, which is difficult to obtain in the field of remote sensing. Self-supervised learning (SSL), can be used to solve such problems by pre-training a general model with a large number of unlabeled images and then fine-tuning it on a downstream task with very few labeled samples. Contrastive learning is a typical method of SSL that can learn general invariant features. However, most existing contrastive learning methods are designed for classification tasks to obtain an image-level representation, which may be suboptimal for semantic segmentation tasks requiring pixel-level discrimination. Therefore, we propose a global style and local matching contrastive learning network (GLCNet) for remote sensing image semantic segmentation. Specifically, 1) the global style contrastive learning module is used to better learn an image-level representation, as we consider that style features can better represent the overall image features. 2) The local features matching contrastive learning module is designed to learn representations of local regions, which is beneficial for semantic segmentation. The experimental results show that our method mostly outperforms SOTA self-supervised methods and the ImageNet pre-training method. Specifically, with 1\% annotation from the original dataset, our approach improves Kappa by 6\% on the ISPRS Potsdam dataset relative to the existing baseline. Moreover, our method outperforms supervised learning methods when there are some differences between the datasets of upstream tasks and downstream tasks. Since SSL could directly learn the essential characteristics of data from unlabeled data, which is easy to obtain in the remote sensing field, this may be of great significance for tasks such as global mapping. The source code is available at https://github.com/GeoX-Lab/G-RSIM.

High-Resolution Remote Sensing Image Semantic Segmentation via Multiscale Context and Linear Self-Attention

Remote sensing image semantic segmentation, which aims to realize pixel-level classification according to the content of remote sensing images, has broad applications in various fields. Thanks to the superiority of deep learning (DL), the semantic segmentation model based on the convolutional neural network (CNN) dramatically promotes the development of remote sensing image semantic segmentation. Due to the high resolution, comprehensive coverage, extensive data, and sizeable spectral difference of high-resolution remote sensing images (HRRSI), the existing GPU is not suitable for directly semantic segmentation of the whole image. Cutting the image into small patches will lead to the loss of context information, resulting in the decline of accuracy. To address this issue, we propose the multiscale context self-attention network (MSCSANet). It combines the benefits of the self-attention mechanism with CNN to improve the segmentation quality of various remote sensing images. The MSCSANet extracts multiscale features from multiscale context images to solve the problem of feature loss caused by image segmentation. In addition, in order to make use of the feature of large-scale context, the multiscale context patches are used to guide the local image patch to focus on different fine-grained objects to enhance the feature of the local image patch. Moreover, considering the limited computing resources, we designed a linear self-attention module to reduce the computational complexity. Compared with other DL models, our proposed model can enhance the ability of multiscale features in complex scenes, and realizes improvements of 1.56% mean intersection over union (MIoU) on the Gaofen Image Dataset and 1.93% MIoU on the ISPRS Potsdam Dataset, respectively.

Collaborative Network for Super-Resolution and Semantic Segmentation of Remote Sensing Images

In the past few years, multitask learning (MTL) has been widely used in a single model to solve the problems of multiple businesses. MTL enables each task to achieve high performance and greatly reduces computational resource overhead. In this work, we designed a collaborative network that simultaneously solves the super-resolution semantic segmentation and super-resolution image reconstruction. This algorithm can obtain high-resolution semantic segmentation and super-resolution reconstruction results by taking relatively low-resolution images as input when high-resolution data are inconvenient or computing resources are limited. The framework consists of three parts: the semantic segmentation branch (SSB), the super-resolution branch (SRB), and the structural affinity block (SAB). Specifically, the SSB, SRB, and SAB are responsible for completing super-resolution semantic segmentation, image super-resolution reconstruction, and associated features, respectively. Our proposed method is simple and efficient, and it can replace the different branches with most of the state-of-the-art models. The International Society for Photogrammetry and Remote Sensing (ISPRS) segmentation benchmarks were used to evaluate our models. In particular, super-resolution semantic segmentation on the Potsdam dataset reduced Intersection over Union (IoU) by only 1.8% when the resolution of the input image was reduced by a factor of two. The experimental results showed that our framework can obtain more accurate semantic segmentation and super-resolution reconstruction results than the single model.

Calibrated Focal Loss for Semantic Labeling of High-Resolution Remote Sensing Images

Currently, the most advanced high-resolution remote sensing image (HRRSI) semantic labeling methods rely on deep neural networks. However, HRRSIs naturally have a serious class imbalance problem which is not yet well solved by the current method. The cross-entropy (CE) loss is often used to guide the training of semantic labeling neural networks for HRRSIs, but it is essentially dominated by the major classes in the image, resulting in poor predictions for the minority class. Based on the prediction results, Focal Loss (FL) effectively suppresses the negative impact of class imbalance in dense object detection by redistributing the loss of each sample. In this paper, we thoroughly analyze the inadequacy of FL for semantic labeling, which inevitably introduces confusing-classified examples that are more difficult to classify while suppressing the loss of well-classified examples. Therefore, following the core idea of FL, we redefine the hard examples in semantic labeling of HRRSIs and propose the prediction confusion map (PCM) to measure the classification difficulty. Based on this, we further propose the Calibrated Focal Loss (CFL) for semantic labeling of HRRSIs. Finally, we conduct complete experiments on the International Society for Photogrammetry and Remote Sensing (ISPRS) Vaihingen and Potsdam datasets to analyze the semantic labeling performance, model uncertainty, and confidence calibration of different loss functions. Experimental results show that CFL can achieve outstanding results compared to other commonly used loss functions without increasing model parameters and training iterations, demonstrating the effectiveness of our method. In the end, combined with our previously proposed HCANet, we further verify the effectiveness of CFL on state-of-the-art network structures.

Semantic Segmentation for High-Resolution Remote-Sensing Images via Dynamic Graph Context Reasoning

Semantic segmentation for high-resolution remote-sensing (HRRS) images is one of the most challenging tasks in remote-sensing images understanding. Capturing long-range dependencies in feature representations is crucial for semantic segmentation. Recent graph-based global reasoning networks (<i>GloRe</i>) focus on modeling the global contextual relationship between latent nodes based on fully connected graph in interaction space. However, such a dense operation is susceptible to redundant features. Most importantly, it treats each node equally, ignoring the contextual relationship between nodes in graphs. In this work, we propose to explore more effective contextual representations in semantic segmentation by introducing dynamic graph contextual reasoning module over <i>GloRe</i>, dubbed DGCR. It incorporates local semantic information that represents the relationships between nodes to perform long-range contextual reasoning. More specifically, to provide effectively and flexible reasoning in graph-based reasoning approaches, we construct <inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>-nearest neighbor (KNN) graphs rather than fully connected graphs using only the <inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula> closest nodes depends on pairwise semantic distance. Extensive experiments on the International Society for Photogrammetry and Remote Sensing (ISPRS) Vaihingen and Potsdam datasets demonstrate the effectiveness and superiority of our proposed DGCR module over other state-of-the-art methods.

Top-Down Pyramid Fusion Network for High-Resolution Remote Sensing Semantic Segmentation

In recent years, high-resolution remote sensing semantic segmentation based on data fusion has gradually become a research focus in the field of land classification, which is an indispensable task of a smart city. However, the existing feature fusion methods with bottom-up structures can achieve limited fusion results. Alternatively, various auxiliary fusion modules significantly increase the complexity of the models and make the training process intolerably expensive. In this paper, we propose a new lightweight model called top-down pyramid fusion network (TdPFNet) including a multi-source feature extractor, a top-down pyramid fusion module and a decoder. It can deeply fuse features from different sources in a top-down structure using high-level semantic knowledge guiding the fusion of low-level texture information. Digital surface model (DSM) data and open street map (OSM) data are used as auxiliary inputs to the Potsdam dataset for the proposed model evaluation. Experimental results show that the network proposed in this paper not only notably improves the segmentation accuracy, but also reduces the complexity of the multi-source semantic segmentation model.

Boundary Extraction Based on Dual Stream Deep Learning Model in High Resolution Remote Sensing Images

Boundary extraction in remote sensing has an important task in studies such as environmental observa-tion, risk management and monitoring urban growth. Although significant progress has been made in the different calculation methods proposed, there are issues that need improvement, especially in terms of accuracy, efficiency and speed. In this study, dual stream network architecture of three different models that can obtain boundary extraction by using normalized Digital Surface Model (nDSM), Normalized Difference Vegetation Index (NDVI) and Near-Infrared (IR) band as the second stream, was explained. Model I is designed as the original HED, whereas the second stream of Model II, III, and IV use nDSM, nDSM + NDVI and nDSM + NDVI + IR, respectively. Thus, by comparing the models trained based on different data combinations, the contribution of different input data to the success of boundary extraction was revealed. For the training of the models, boundary maps produced from The International Society for Photogrammetry and Remote Sensing (ISPRS) Potsdam data set and input datasets augmented by rotation, mirroring and rotation were used. When the test results obtained from two-stream and multi-data-based models are evaluated, 11% better accuracy has achieved with Model IV compared to the original HED. The outcomes clearly revealed the importance of using multispectral band, height data and vegetation information as input data in boundary extraction beside commonly used RGB images.

Multi-Modality and Multi-Scale Attention Fusion Network for Land Cover Classification from VHR Remote Sensing Images

Land cover classification from very high-resolution (VHR) remote sensing images is a challenging task due to the complexity of geography scenes and the varying shape and size of ground targets. It is difficult to utilize the spectral data directly, or to use traditional multi-scale feature extraction methods, to improve VHR remote sensing image classification results. To address the problem, we proposed a multi-modality and multi-scale attention fusion network for land cover classification from VHR remote sensing images. First, based on the encoding-decoding network, we designed a multi-modality fusion module that can simultaneously fuse more useful features and avoid redundant features. This addresses the problem of low classification accuracy for some objects caused by the weak ability of feature representation from single modality data. Second, a novel multi-scale spatial context enhancement module was introduced to improve feature fusion, which solves the problem of a large-scale variation of objects in remote sensing images, and captures long-range spatial relationships between objects. The proposed network and comparative networks were evaluated on two public datasets—the Vaihingen and the Potsdam datasets. It was observed that the proposed network achieves better classification results, with a mean F1-score of 88.6% for the Vaihingen dataset and 92.3% for the Potsdam dataset. Experimental results show that our model is superior to the state-of-the-art network models.

DGFNet: Dual Gate Fusion Network for Land Cover Classification in Very High-Resolution Images

Deep convolutional neural networks (DCNNs) have been used to achieve state-of-the-art performance on land cover classification thanks to their outstanding nonlinear feature extraction ability. DCNNs are usually designed as an encoder–decoder architecture for the land cover classification in very high-resolution (VHR) remote sensing images. The encoder captures semantic representation by stacking convolution layers and shrinking image spatial resolution, while the decoder restores the spatial information by an upsampling operation and combines it with different level features through a summation or skip connection. However, there is still a semantic gap between different-level features; a simple summation or skip connection will reduce the performance of land-cover classification. To overcome this problem, we propose a novel end-to-end network named Dual Gate Fusion Network (DGFNet) to restrain the impact of the semantic gap. In detail, the key of DGFNet consists of two main components: Feature Enhancement Module (FEM) and Dual Gate Fusion Module (DGFM). Firstly, the FEM combines local information with global contents and strengthens the feature representation in the encoder. Secondly, the DGFM is proposed to reduce the semantic gap between different level features, effectively fusing low-level spatial information and high-level semantic information in the decoder. Extensive experiments conducted on the LandCover dataset and the ISPRS Potsdam dataset proved the effectiveness of the proposed network. The DGFNet achieves state-of-art performance 88.87% MIoU on the LandCover dataset and 72.25% MIoU on the ISPRS Potsdam dataset.

Appearance based deep domain adaptation for the classification of aerial images

This paper addresses appearance based domain adaptation for the pixel-wise classification of remotely sensed data using deep neural networks (DNN) as a strategy to reduce the requirements of DNN with respect to the availability of training data. We focus on the setting in which labelled data are only available in a source domain DS, but not in a target domain DT, known as unsupervised domain adaptation in Computer Vision. Our method is based on adversarial training of an appearance adaptation network (AAN) that transforms images from DS such that they look like images from DT. Together with the original label maps from DS, the transformed images are used to adapt a DNN to DT. The AAN has to change the appearance of objects of a certain class such that they resemble objects of the same class in DT. Many approaches try to achieve this goal by incorporating cycle consistency in the adaptation process, but such approaches tend to hallucinate structures that occur frequently in one of the domains. In contrast, we propose a joint training strategy of the AAN and the classifier, which constrains the AAN to transform the images such that they are correctly classified. To further improve the adaptation performance, we propose a new regularization loss for the discriminator network used in adversarial training. We also address the problem of finding the optimal values of the trained network parameters, proposing a new unsupervised entropy based parameter selection criterion, which compensates for the fact that there is no validation set in DT that could be monitored. As a minor contribution, we present a new weighting strategy for the cross-entropy loss, addressing the problem of imbalanced class distributions. Our method is evaluated in 42 adaptation scenarios using datasets from 7 cities, all consisting of high-resolution digital orthophotos and height data. It achieves a positive transfer in all cases, and on average it improves the performance in the target domain by 4.3% in overall accuracy. In adaptation scenarios between the Vaihingen and Potsdam datasets from the ISPRS semantic labelling benchmark our method outperforms those from recent publications by 10-20% with respect to the mean intersection over union.

Transformer Meets Convolution: A Bilateral Awareness Network for Semantic Segmentation of Very Fine Resolution Urban Scene Images

Semantic segmentation from very fine resolution (VFR) urban scene images plays a significant role in several application scenarios including autonomous driving, land cover classification, urban planning, etc. However, the tremendous details contained in the VFR image, especially the considerable variations in scale and appearance of objects, severely limit the potential of the existing deep learning approaches. Addressing such issues represents a promising research field in the remote sensing community, which paves the way for scene-level landscape pattern analysis and decision making. In this paper, we propose a Bilateral Awareness Network which contains a dependency path and a texture path to fully capture the long-range relationships and fine-grained details in VFR images. Specifically, the dependency path is conducted based on the ResT, a novel Transformer backbone with memory-efficient multi-head self-attention, while the texture path is built on the stacked convolution operation. In addition, using the linear attention mechanism, a feature aggregation module is designed to effectively fuse the dependency features and texture features. Extensive experiments conducted on the three large-scale urban scene image segmentation datasets, i.e., ISPRS Vaihingen dataset, ISPRS Potsdam dataset, and UAVid dataset, demonstrate the effectiveness of our BANet. Specifically, a 64.6% mIoU is achieved on the UAVid dataset.

Self-constructing graph neural networks to model long-range pixel dependencies for semantic segmentation of remote sensing images

ABSTRACT Capturing global contextual representations in remote sensing images by exploiting long-range pixel–pixel dependencies has been shown to improve segmentation performance. However, how to do this efficiently is an open question as current approaches of utilizing attention schemes, or very deep models to increase the field of view, increases complexity and memory consumption. Inspired by recent work on graph neural networks, we propose the Self-Constructing Graph (SCG) module that learns a long-range dependency graph directly from the image data and uses it to capture global contextual information efficiently to improve semantic segmentation. The SCG module provides a high degree of flexibility for constructing segmentation networks that seamlessly make use of the benefits of variants of graph neural networks (GNN) and convolutional neural networks (CNN). Our SCG-GCN model, a variant of SCG-Net built upon graph convolutional networks (GCN), performs semantic segmentation in an end-to-end manner with competitive performance on the publicly available ISPRS Potsdam and Vaihingen datasets, achieving a mean F1-scores of 92.0% and 89.8%, respectively. We conclude that the SCG-Net is an attractive architecture for semantic segmentation of remote sensing images since it achieves competitive performance with much fewer parameters and lower computational cost compared to related models based on convolutional neural networks.

Extraction of urban multi-class from high-resolution images using pyramid generative adversarial networks

The extraction of man-made structures from high-resolution images plays a vital role in various urban applications. This task is regularly complicated due to the heterogeneous appearance of the objects in the satellite images. In this work, we propose a multi-scale Generative Adversarial network to classify high-resolution images into urban classes (surface, building, tree, low-vegetation, car). We use uNet with EfficientNet B3 architecture as a generator and we use ResNet 18 architecture as a discriminator. We use a conditional generative loss based on the Dice coefficient and softmax functions. Experiments on the Vaihingen and Potsdam datasets were conducted to demonstrate the performance and we compare the results with other architectures. The results demonstrate the validity and higher performance of the proposed multi-scale network for extracting classes in urban areas with average F1-score 89.0% and 88.8%, and average accuracy 89.9% and 91.8% for Vaihingen and Potsdam datasets, respectively.