Land Use Land Cover Classification from Satellite Imagery using mUnet: A Modified Unet Architecture Lakshya Garg 1 , Parul Shukla 2 , Sandeep Kumar Singh 2 , Vaishangi Bajpai 2 and Utkarsh Yadav 2 1 Electrical and Electronics Engineering, Delhi Technological University (DTU), Delhi, India 2 Strategic Operations and Research, RMSI Pvt Ltd Noida, Delhi, India utkarsh.yadav@india.rmsi.com Keywords: Satellite Imagery, Land-use-classification, Convolutional Networks, Remote Sensing, Deep Learning. Abstract: Land-use-land-cover classification(LULC) is used to automate the process of providing labels, describing the physical land type to represent how a land area is being used. Many sectors such as telecom, utility, hydrology etc need land use and land cover information from remote sensing images. This information pro- vides an insight into the type of geographical distribution of a region with providing low level features such as amount of vegetation, building area, and geometry etc as well as higher level concepts such as land use classes. This information is particularly useful for resource-starved rapidly developing cities for urban plan- ning and resource management. LULC also provides historical changes in land-use patterns over a period of time. In this paper, we analyze patterns of land use in urban and rural neighborhoods using high reso- lution satellite imagery, utilizing a state of the art deep convolutional neural network. The proposed LULC network, termed as mUnet is based on an encoder-decoder convolutional architecture for pixel-level semantic segmentation. We test our approach on 3 band, FCC satellite imagery covering 225 km 2 area of Karachi. Experimental results show the superiority of our proposed network architecture vis-` a-vis other state of the art networks. 1 INTRODUCTION Recent advancements in remote sensing have re- sulted in easy accessibility of satellite imagery of Earth. Among many applications of satellite ima- gery, Land-use-land-cover (LULC) forms an in- tegral part. LULC plays a pivotal role in ur- ban planning, land resource management and pro- vides a useful insight into the growth rate indexes of different population spectrum. LULC models also highlight the historical changes in an area by quantitatively showcasing the changes in habita- tion, vegetation, water areas over a span of years. Recent breakthroughs in deep learning have resul- ted in LULC’s application to various new versa- tile domains such as development of smart cities, urban planning, environmental monitoring and di- saster recovery. The remote sensing community benefited the most with the use of deep convo- lutional neural networks for various tasks such as automatic feature extraction for classifying road, building footprints, grasslands etc (Mnih, 2013). Land use classification refers to the consolidation of physical land attributes defining a region and what the cultural and socio-economic function that the land serves. In this paper, we address LULC classification by modifying the Unet (Ronneberger et al., 2015) ar- chitecture for pixel level segmentation. The propo- sed model is henceforth referred to as mUnet. mUnet consists of a ladder-like structure consisting of con- volutional encoder layers followed by a series of de- coding convolutional layers. Our method has the fol- lowing advantages as compared to the traditional seg- mentation methods. • mUnet provides classification at pixel level. • mUnet consists of less number of trainable para- meters as compared to original Unet. • mUnet outperforms other state of the art segmen- tation architectures. Our paper consist of the following sections: In section-II we review the studies which adopt deep learning methods to the problem of land-use- classification. In section-III, we present the dataset Garg, L., Shukla, P., Singh, S., Bajpai, V. and Yadav, U. Land Use Land Cover Classification from Satellite Imagery using mUnet: A Modified Unet Architecture. DOI: 10.5220/0007370603590365 In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 359-365 ISBN: 978-989-758-354-4 Copyright c  2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved 359