Land use Land Cover Classification using Deep Neural
Network
Purva Suryawanshi
1
, Suraj Sawant
2
and Amit Joshi
3
1
Post Graduate Student, Computer Engineering Department, College of Engineering, Pune, India
Email: suryawanshipd20.comp@coep.ac.in
2-3
Assistant Professor, Computer Engineering Department, College of Engineering, Pune, India
Email: {sts.comp, adj.comp}@coep.ac.in
Abstract—With remote sensor data, land use and land cover classification offers geographical
information. The land use land cover classification process considers several factors for urban
planning, natural resources management, environment management, and disaster monitoring.
The recent advances in deep learning models to classify land use land cover classes. The
objective of this research work is to use remote sensing and Geospatial Information Systems to
perform land use land cover classification. The customized Sentinel-2 dataset is generated using
QGIS to implement a Deep Neural network. In this study, pixel-based classification was
performed using the UNet model. Accuracy assessment was performed using Overall Accuracy
and F1-Score. The research showed overall classification accuracy of 95.5% and a value of F1-
Score is 0.66. The UNet model is suitable for customized datasets to present essential
information for sustainable environmental planning.
Index Terms— Remote sensing, Land use land cover, GIS, Deep Neural network, UNet,
Sentinel-2.
I. INTRODUCTION
Land Use Land Cover(LULC) classification is important for mapping and environmental applications, such as
urban planning [1], natural resources management, and crop yield prediction [2]. To provide useful land
information, remote sensing is essential and often used with GIS techniques. Remote sensing technology has a
high potential to perform pixel-based LULC classification. Remotely sensed imagery has traditionally been
classified into land cover (LC) and land use (LU). Pixel-based image classification is assigning labels to each
pixel of the image [3]. High-resolution remote sensing images provide spectral signatures, texture features, and
spatial structure of ground objects which are essential for LULC classification [4]. Sentinel-2 L2A satellite
images have been used to research various geographical applications. The effectiveness of Sentinel-2 has been
explored, particularly in LULC mapping, and it has demonstrated a high potential for application[5].
Several techniques have been developed for LULC classification, including traditional machine-learning
algorithms and advanced deep learning networks. Traditional machine learning algorithms such as the support
vector machine (SVM) and the maximum likelihood classifier (MLC) are used to classify images into several
classes. The image's low-level features, such as color, geometry characteristics, greyscale and spatial texture, are
extracted for classification. These methods are less efficient, costly, and time-consuming for high-resolution
satellite imagery[6]. The deep learning architectures allow the high-level feature extraction, and robust features
Grenze ID: 01.GIJET.9.1.562
© Grenze Scientific Society, 2023
Grenze International Journal of Engineering and Technology, Jan Issue