.

\begin{figure*}[t]
\centering
\includegraphics[width=0.9\textwidth]{figures/satellite_example.pdf}
\caption{\label{fig:satellite_example} Example of a satellite image from the ISPRS 2D Semantic Labeling dataset \cite{rottensteiner2012isprs}. The ground truth labels are shown in the rightmost panel.}
\end{figure*}

In this work, we focus on semantic segmentation, which is the task of assigning a class label to each pixel in an image. This task is important for many applications such as autonomous driving, medical imaging and remote sensing \cite{chen2018deeplab}. In particular, we focus on semantic segmentation of satellite images, which is a challenging problem due to the large variation in appearance between different classes and the small size of objects in satellite images (see Figure~\ref{fig:satellite_example}).

In this paper, we propose a novel approach for semantic segmentation of satellite images based on deep learning. Specifically, our approach consists of two components: 1) an encoder-decoder architecture for feature extraction and pixel-wise classification and 2) an attention mechanism for improved feature learning. The encoder-decoder architecture consists of a convolutional neural network (CNN) encoder that extracts features from the input image and a decoder that uses these features to classify each pixel in the image. The attention mechanism is used to improve feature learning by focusing on relevant regions in the input image. We evaluate our approach on two publicly available datasets for semantic segmentation of satellite images and show that it achieves state-of-the-art performance.






















% \section{Related Work} \label{sec:related_work}
% Semantic segmentation has been studied extensively in recent years with deep learning approaches being particularly successful \cite{long2015fully}. In particular, fully convolutional networks (FCNs) \cite{long2015fully}, which are based on CNNs with upsampling layers added at the end of the network, have achieved state-of-the-art results for many applications such as medical imaging \cite{ronneberger2015u}, autonomous driving \cite{chen2018deeplab}, and remote sensing \cite{marmanis2016semantic}. FCNs are well suited for semantic segmentation tasks since they can be used to directly produce pixel-wise classifications without requiring additional postprocessing steps.
% % FCNs have been used successfully for many applications such as medical imaging \cite{ronneberger2015u}, autonomous driving \cite{chen2018deeplab}, and remote sensing \cite{marmanis2016semantic}.
% % FCNs are well suited for semantic segmentation tasks since they can be used to directly produce pixel-wise classifications without requiring additional postprocessing steps.
% % FCNs have been used successfully for many applications such as medical imaging [], autonomous driving [], and remote sensing [].
% % FCNs are well suited for semantic segmentation tasks since they can be used to directly produce pixel-wise classifications without requiring additional postprocessing steps.
% % For example, Ronneberger et al. [] proposed U-Net [] which is an FCN based model designed specifically for biomedical image segmentation tasks. U-Net has achieved state-of-the art results on several datasets including ISBI cell tracking challenge [] and Kaggle data science bowl [] among others [].
% % Chen et al.[] proposed DeepLab [] which is an FCN based model designed specifically for object boundary detection tasks such as autonomous driving []. DeepLab has achieved state-of-the art results on several datasets including Cityscapes [] and PASCAL VOC 2012 [] among others [].
% % Marmanis et al.[] proposed SegNet [] which is an FCN based model designed specifically for semantic segmentation tasks such as remote sensing []. SegNet has achieved state-of-the art results on several datasets including ISPRS Vaihingen dataset [] among others [].