2023 7th International Conference On Computing, Communication, Control And Automation (ICCUBEA)

Pimpri Chinchwad College of Engineering (PCCOE), Pune, India. Aug 18-19, 2023

High-Resolution Satellite Imagery Analysis for

Terrain and Surface Data Extraction: Techniques
2023 7th International Conference On Computing, Communication, Control And Automation (ICCUBEA) | 979-8-3503-0426-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICCUBEA58933.2023.10392073

and Applications
1st Sameeksha Sukate 2nd Sakshi Wajage 3rd Yogesh Verma
Computer Science and Engineering Computer Science and Engineering Computer Science and Engineering
MIT ADT University MIT ADT University MIT ADT University
Pune, India Pune, India Pune, India
[email protected] [email protected] [email protected]

4th Parth Tawde 5th Jayashree Prasad

Computer Science and Engineering Computer Science and Engineering
MIT ADT University MIT ADT University
Pune, India Pune, India
[email protected] [email protected]

Abstract—With multiple applications and use cases of Digital I. I NTRODUCTION

Elevation Models (DEMs), focusing significantly on forming 3D
models. It is used in various research areas such as flood The availability of many Satellite images has given rise to
modeling, agriculture, satellite navigation, farming, and forestry.
This research paper presents a methodology for extracting Digital applications in various fields, increasing the potential of DEM
Elevation Models (DEMs) from satellite images using U-Net. The extraction. The DEM is a 3D representation that provides
U-NET is a Deep Learning Model architecture that is commonly information about the height of the ground, vegetation, or any
used for image segmentation tasks. It is a fully convolutional other object above the ground level. The extraction of digital
neural network that uses an encoder-decoder architecture to elevation models (DEMs) from satellite images is an essential
extract features from the input image and generate a pixel wise
output. The U-Net model is ideal for DEM extraction from task in various fields, including environmental monitoring,
satellite images as it can effectively distinguish between the geological exploration, and urban planning. However, this
different features present in the images and accurately segment process can be time-consuming and complex due to the vast
the ground elevation ranges. The creation of the dataset was a amount of data that needs to be processed. Recent advances
crucial step in the project as it ensured that the U-Net model in deep learning techniques have shown promising results in
was trained on a diverse set of images. The SWISSIMAGE
dataset provided high-resolution RGB imagery, while the SRTM automating this process, making it faster and more efficient.
dataset provided the corresponding mask images with ground One such technique is the use of U-Net, a convolutional neural
elevation ranges. The results obtained from the project indicate network architecture that has been widely used in image seg-
that the model can perform reasonably well even with limited mentation tasks. The project was divided into two main parts:
data and without any significant pre-processing of the input dataset creation and U-Net training. In the first part, a dataset
dataset. However, there are still some challenges and limitations
corresponding to the use of U-Net for DEM extraction. In of 2000 images was created from the “SWISSIMAGE 10 cm
conclusion, this study provides a valuable contribution to the RGB imagery” dataset and 2000 corresponding mask images
field of remote sensing and computer vision, highlighting the were taken from the “SRTM Digital Elevation Data Version
potential of U-Net for DEM extraction from satellite images. In 4” dataset. Both datasets were available on Google Earth
addition, the latest improvements in both software and hardware Engine. The SWISSIMAGE dataset was not pre-processed,
have made it possible to carry the processing with remarkable
speed, even on moderately powerful hardware. We anticipate while the SRTM dataset underwent a pre-processing step,
that our research will act as a catalyst for further advancements where it was segmented into 9 labels, each representing a
and will facilitate the integration of this research in generating, distinct ground elevation range. In the second part of the
updating, and analysis of DEMs. However, further research is project, a UNET model was trained on these 2000 images. The
needed to address the limitations and challenges associated with model achieved a training accuracy of 70per and a validation
this approach and to develop more accurate and reliable models
for DEM extraction accuracy of 62 percent. The results indicate that the UNET
model can be used effectively to extract DEMs from satellite
images. This research paper aims to explore the use of U-
Index Terms—Digital Elevation Models, Remote Sensing, U- Net for DEM extraction from satellite images. The paper
Net, Deep learning, Image segmentation, Encoder-decoder
will begin by providing a brief overview of the current state

979-8-3503-0426-8/23/$31.00 ©2023 IEEE 1

Authorized licensed use limited to: Somaiya University. Downloaded on April 17,2025 at 17:20:24 UTC from IEEE Xplore. Restrictions apply.
of DEM extraction techniques and their limitations. It will depicts a DEM counter map generated from satellite images
then introduce the U-Net architecture and explain how it can using UNET, with 9 different colors representing various
be used for DEM extraction. Our research demonstrates the elevations ranging from 0 to 5000 meters. The map features
effectiveness of deep learning techniques for high-resolution shades of orange representing lower elevations and shades of
satellite imagery analysis for DEM extraction. Our approach purple representing higher elevations, providing a clear visual
has the potential to improve the accuracy and efficiency of representation of the terrain. DEM contour maps are a type of
DEM extraction from satellite images, which can have signif- topographical map that displays the three-dimensional shape of
icant practical applications in terrain mapping, environmental a terrain in a two-dimensional format. Contour lines are drawn
monitoring, disaster management, and urban planning. on the map to show areas of equal elevation, with each line
representing a specific height above sea level. These maps are
II. METHODS AND METHODOLOGIES useful for a variety of applications, including land surveying,
A. Overview and Literature Survey urban planning, and outdoor recreational activities such as
There has been a significant amount of research con- hiking and camping. DEM contour maps are a powerful tool
ducted by various researchers in the field of DEM extraction for understanding and visualizing the shape and topography of
from satellite imagery. Some of the key research areas and the land, and they are widely used in many different industries
approaches are summarized below. DEM (Digital Elevation and fields. Fig-2 depicts the process of extracting a digital
Model) data is generated using a variety of techniques, which
include: Light Detection and Ranging(LIDAR): This method
uses a laser scanner to measure the distance between the
sensor and the terrain, and then generates a 3D model of the
terrain based on these measurements. Photogrammetry: This
method involves capturing images of the terrain from various
angles and using them to generate a 3D model of the terrain.
Radar Interferometry: Radar interferometry, or InSAR, uses
two or more radar images of the same area to generate a
3D model of the terrain. By comparing the phase differences
between the images, InSAR can calculate the height differ-
ences between different parts of the landscape. Traditionally,
DEM extraction is a complex and time-consuming process that
requires skilled personnel, specialized software, and high-end
hardware. With the recent advancements in Deep Learning Fig. 2. Satellite-based DEM Contour Map Extraction Using U-Net (9
Elevation Ranges).
techniques, now it is possible to automate this process and
obtain accurate DEMs from satellite images with less manual elevation model (DEM) contour map from satellite images
intervention. One popular deep learning architecture used for using a neural network architecture called U-Net. The U-Net
image segmentation tasks, including DEM extraction, is the U- takes as input a dataset of satellite images along with their
Net. We have trained the UNET model using satellite images corresponding DEM mask and produces a contour map of the
and corresponding elevation maps. Once the model is trained, terrain with 9 different elevation ranges represented by distinct
it is used to segment the satellite images into 9 different colors.
height/elevation classes. Each class is assigned a specific color,
which is then used to create the DEM counter map. The III. TOOLS AND TECHNOLOGIES
resulting map shows the elevation or height of each pixel
in the satellite image, represented by a specific color. Fig-1 A. Dataset - Study Area

Fig. 3. Switzerland’s topography captured by SWISSIMAGE and SRTM

Fig. 1. Elevation Contour Map: Nine-Colour Representation. imagery

2
Authorized licensed use limited to: Somaiya University. Downloaded on April 17,2025 at 17:20:24 UTC from IEEE Xplore. Restrictions apply.
 In recent years, remote sensing technology has made it pos- for the handling of large datasets and GPU computation.
sible to extract valuable information from satellite images. One Google Earth Engine (GEE) allows individuals to visualize
such application is the generation of digital elevation models and examine satellite images of Earth for various purposes,
(DEMs) using satellite images and Shuttle Radar Topography including remote sensing - [4].
Mission (SRTM) DEM data - [1]. As show in fig-3, we used
the SWISSIMAGE - [3] and SRTM data catalog, both based on
the area of Switzerland, from Google Earth Engine to generate
images that we used to train our U-Net model. The U-Net
model was able to learn from these images and accurately
extract DEM information. The SWISSIMAGE catalog consists
of a mosaic of orthophotos with a ground resolution of 10
cm in flat regions, created from new color digital images
that encompass the entirety of Switzerland in flat areas and
main alpine valleys, and 25 cm in the Alps. These images
are updated every 3 years and are provided by Topography
Swisstopo. Orthophotos are distortion-free images with a uni-
form scale throughout, and they have numerous applications
in fields such as geology, forestry, environmental protection, Fig. 4. SWISSIMAGE and SRTM datasets, 2000 images each, Switzerland.
spatial planning, and natural hazards. SWISSIMAGE provides
the latest orthophotos by default, while older photos are stored It is a Google initiative aimed at archiving all spatial
and easily accessible. This allows for a detailed and up-to- images captured by satellites for research purposes. It has
date background map for various purposes. Along with this, been collecting such data since 1972 and was made available
we have a Shuttle radar topography mission (SRTM) a global for public research in 2008. The Earth Engine utilizes cloud
project that provides us with a high-resolution digital elevation technology to store massive datasets that necessitate substan-
model (DEM) of the Earth, which is of high quality and tial computational resources to handle vast geospatial data.
available for free. The project was a collaboration between The Earth Engine’s data catalog stores and pre-processes vast
the National Aeronautics and Space Administration (NASA), datasets. Moreover, the datasets have pre-built code snippets
the National Geospatial-Intelligence Agency (NGA), and the that users can utilize in the Earth Engine’s code editor to
space agencies of Germany and Italy. The SRTM technology explore and analyze the data. Additionally, the GEE offers an
was used to create the high-resolution digital-topographic application programming interface (API) for prototyping and
radar data of the Earth, which was used to generate the first visualizing the results. Google Colab is a cloud-based platform
worldwide set of Earth’s topographic data. Additionally, the that provides free access to a Jupyter Notebook environment
SRTM data is used to fill in areas with missing data to provide for writing and executing Python code. - [5]. It provides a
accurate digital elevation data. The SRTM was mounted on virtual machine with access to powerful GPUs, TPUs, and
a space shuttle and used remote sensing to acquire Earth’s other resources, enabling users to train machine learning
surface data, which was then converted into height data and models, perform data analysis, and more. Google Colab also
used to generate a three-dimensional map of larger areas of the supports popular libraries like TensorFlow and PyTorch. The
Earth. This high-resolution digital topography map can be used project’s success was demonstrated by the accuracy of the
in various fields and is publicly available. With the assistance extracted DEM, which was verified against reference data. The
of Google Collab’s Earth-Engine library, we have modified results showed that the project achieved high accuracy and
the SRTM data into 9 distinct colours in our project. The demonstrated the power of using Google Earth Engine and
SWISSIMAGE dataset comprises a total of 2000 images with Google Colab for large-scale satellite image analysis projects.
various tile sizes, covering over 70 percent of Switzerland’s The combination of these tools offers a cost-effective and
total area. To correspond with these 2000 images, we have efficient solution for researchers and data scientists working
generated 2000 masks (SRTM data) of the corresponding on similar projects.
Swiss Image using Google Earth-Engine library. These masks
are known as DEM tiles and are displayed in fig-4 Google B. U-Net
Earth Engine and Google Colab are powerful tools that enable The U-Net model architecture, introduced in 2015 by the
data scientists and researchers to work on large-scale satellite Computer Science Department of the University of Freiburg,
imagery analysis projects. In this project, Google Earth Engine Germany - [2], is characterized by its ”U” shape. Primarily
was used to collect the data, and SWISSIMAGE and SRTM designed for biomedical image segmentation, U-Net employs
data catalogs were utilized to extract information. The data an encoder-decoder architecture with skip connections. The
from Google Earth Engine was pre-processed on Google encoder section of U-Net compresses the spatial dimensions
Colab, which enabled the team to write code and perform and augments the channel count in each layer, while the
necessary tasks on the data. Google Colab was chosen due to decoder section expands the spatial dimensions and reduces
its ability to leverage Google’s infrastructure, which allowed the channel count in each layer. Neural Networks can “forget”

3
Authorized licensed use limited to: Somaiya University. Downloaded on April 17,2025 at 17:20:24 UTC from IEEE Xplore. Restrictions apply.
certain features during the training process, and skip connec- C. CNN (Convolutional Nural Network)
tions are used in U-Net to make sure details are not lost. CNNs (Convolutional Neural Networks) have a broad range
Unlike CNN, we do not have fully connected layers in the of uses, from voice recognition and facial recognition to image
end, since the desired output should be of the desired size. processing and nearly all computer vision tasks [8]. These
The encoder section, also referred to as the contracting path, networks function by mapping features, using a technique
consists of conventional convolutional layers. On the other that involves backpropagating results through various layers,
hand, the decoder section, also known as the expanding path, including pooling layers, convolutional layers, and fully con-
is comprised of transposed 2D convolutional layers, which nected layers. The three key ideas of CNN as identified by
are commonly known as deconvolutional layers. By layer- Goodfellow et al. [7] One approach involves utilizing small
wise concatenation of the encoder’s output with the decoder’s feature detectors to identify edges within an image, particularly
input, every deconvolutional layer in the decoder is connected in cases where the image is large, such as satellite imagery.
to a convolutional layer of an encoder. During training, the This technique, referred to as ”sparse interactions,” aims to
model learns to segment the input image into distinct regions, overcome the challenge of detecting edges in such images.
such as terrain and non-terrain areas, based on the learned Another strategy is to implement ”parameter sharing,” which
features. This makes the U-Net architecture an effective tool helps manage the number of weights required for processing
for DEM extraction from satellite images. In the context of an input image. By reducing the number of parameters, this
DEM extraction from satellite images, the U-Net architecture technique can significantly decrease the computational power
can be used to identify and extract the elevation information needed. Lastly, “Equivariant representation” implies that ob-
of the terrain from the input image. The encoder part of ject detection remains unaffected by variations in illumination
the network can be trained to identify various features of and changes in position, while the internal representation of
the image, such as vegetation, water bodies, and buildings, the detected object remains equivariant to such alterations.
that can affect the elevation of the terrain. The decoder part By utilizing these points, we can overcome the challenge of
of the network can then use the features identified by the processing large satellite images that are difficult to manage
encoder to reconstruct the elevation map of the terrain. By in their original dimensions. This operation involves only
incorporating skip connections between the encoder and de- a minimal number of parameters, streamlining the training
coder sections, the U-Net architecture effectively captures both process and accelerating the network’s performance. CNN
types of information. These connections enable the decoder utilizes multiple convolutional layers to extract characteristics
to access the features extracted by the encoder at various from input images. Each convolutional layer incorporates a
spatial resolutions, facilitating the preservation of both local collection of adaptable filters that are employed on localized
and global information in the ultimate output. To summarize, sections of the input image. Through convolution, these filters
the U-Net architecture stands as a robust and adaptable CNN interact with the input image, generating a collection of
architecture, well-suited for satellite image-based Digital El- feature maps that encapsulate various attributes of the original
evation Model (DEM) extraction. Its capacity to handle both image. Following each convolutional layer, a pooling layer
local and global information within the input image renders it is employed to downsize the feature maps while retaining
suitable for a diverse array of image segmentation tasks. With vital information. This serves to enhance network efficiency
its high accuracy and efficiency, the U-Net architecture has and prevent overfitting. The ultimate outcome of a CNN is
the potential to revolutionize the field of remote sensing and typically directed through one or more fully connected layers,
satellite image analysis. which employ the features derived from the convolutional
layers to classify the input image. CNNs have exhibited
remarkable success in a broad spectrum of applications, in-
cluding image classification, object detection, and image seg-
mentation. They have revolutionized the domain of computer
vision, establishing themselves as an indispensable tool across
various domains of machine learning and artificial intelligence.
To summarize, convolutional neural networks (CNNs) have
emerged as a critical component in the realm of deep learning,
particularly within the field of computer vision. By effectively
extracting image features and discerning patterns, CNNs have
found applications ranging from object recognition to medical
image analysis. However, there is still much to be explored
and improved upon in the realm of CNNs, such as optimizing
the architecture and improving interpretability.
D. Flow Chart
Fig. 5. U-Net for DEM extraction via satellite imagery: a model architecture
In Fig-6, we propose a flowchart for the extraction of the
Digital Elevation Model (DEM) from satellite images using

4
Authorized licensed use limited to: Somaiya University. Downloaded on April 17,2025 at 17:20:24 UTC from IEEE Xplore. Restrictions apply.
UNET. The first step in our proposed flowchart is to perform IV. RESULTS
preprocessing on Google Earth Engine to reduce the number The application of U-net in the extraction of digital elevation
of labels in the mask image from 256 labels to 9 labels. Next, models (DEM) from a dataset of 2000 satellite images has
a dataset consisting of 2000 images and their corresponding produced encouraging results. Through the use of pixel-level
masks is extracted from the preprocessed data. The dataset classification, U-net was able to accurately identify and differ-
is initially divided into training, validation, and test sets. entiate object borders, resulting in the successful extraction of
Subsequently, the UNET model undergoes training using the elevation data from the images. The encoder-decoder architec-
training set and evaluation using the validation set. Finally, ture of U-net effectively utilized the feature maps generated
the test set is utilized to generate the outcome. This proposed by the encoder network to produce highly detailed outputs,
flowchart presents a sequential guide for extracting Digital leading to a high degree of accuracy in the resulting DEM.
Elevation Models (DEMs) from satellite images employing Overall, the application of U-net in the processing of a large
UNET. In the last stage of the flowchart, the result is produced, dataset of satellite imagery has demonstrated its potential as a
encompassing the extracted DEM from the satellite images. reliable and effective tool for DEM extraction. In this project, a
This outcome is then compared to the ground truth DEM to U-Net model was trained on a dataset of 2000 satellite images
assess the accuracy of the model. The accompanying research and their corresponding digital elevation models (DEMs) to
paper will furnish a comprehensive explanation of each step
depicted in the flowchart, along with the methodology em-
ployed for DEM extraction from satellite images using UNET.

Fig. 7. Model’s results on unseen Test data

extract DEM contour maps, with 9 different colors repre-

senting various elevations. The presented fig-7 showcases a
satellite image of landscapes, accompanied by two additional
images. The first additional image is a ground truth Digital
Elevation Model (DEM), providing an accurate representation
of the elevation data of the terrain. The second image displays
the predicted DEM created by a U-Net model. The U-Net
model has been trained to accurately predict elevation data
from satellite imagery. The comparison of the two DEM
images provides valuable insight into the accuracy of the U-
Net model’s predictions.
The graphs, in fig-8, show the training and validation accu-
racy and loss of a UNET model trained on satellite images to
generate DEM contour maps as output. The training accuracy
and validation accuracy graphs show an upward trend with
Fig. 6. Data Flow diagram increasing epochs, indicating that the model is learning the
features and patterns in the data. The graphs illustrating the

5
Authorized licensed use limited to: Somaiya University. Downloaded on April 17,2025 at 17:20:24 UTC from IEEE Xplore. Restrictions apply.
 elevation data from satellite imagery. However, these methods
can be time-consuming and expensive. Nevertheless, the use of
advanced techniques, such as U-Net, has shown great potential
for improving the accuracy and efficiency of DEM extraction
from satellite imagery. U-Net has shown great potential in
identifying and segmenting terrain features from the back-
ground, which is a critical step in DEM extraction. The derived
Digital Elevation Models (DEMs) possess versatile utility in
a range of applications, encompassing environmental monitor-
ing, land use planning, and hazard mitigation. Therefore, the
continued development and use of advanced techniques for
DEM extraction from satellite imagery is essential for improv-
ing our understanding of the Earth’s surface and for making
informed decisions about its management and conservation.
The current processing time for extracting DEMs using U-
Net models can be quite time-consuming, especially for large-
scale projects. Future research can focus on developing faster
processing methods to allow for real-time DEM extraction.
This would open up a range of new applications, such as
monitoring of changes in terrain in real time. DEM extraction
using U-Net models can be combined with other technologies
to create more powerful applications. For example, the inte-
gration of DEMs with Geographic Information Systems (GIS)
could enable the creation of highly accurate and detailed maps.
Overall, the extraction of DEMs from satellite imagery using
U-Net models is a rapidly evolving field with many exciting
future possibilities.
R EFERENCES
[1] Jarvis, A., H.I. Reuter, A. Nelson, E. Guevara. 2008. Hole-filled SRTM
Fig. 8. Graphs showing U-Net’s accuracy and loss in DEM extraction from for the globe Version 4, available from the CGIAR-CSI SRTM 90m
satellites. Database: https://srtm.csi.cgiar.org.
[2] Ronneberger, O., Fischer, P. and Brox, T. 2015. U-Net: Convolutional
Networks for Biomedical Image Segmentation. Medical Image Comput-
ing and Computer-Assisted Intervention – MICCAI 2015 (Cham, 2015),
training loss and validation loss exhibit a consistent decline, 234–241
signifying the model’s progressive reduction in prediction er- [3] Swisstopo SWISSIMAGE10cm. Available online:
rors. The UNET model attained a training accuracy of 0.90 and https://www.swisstopo.admin.ch/en/geodata/images/-
ortho/swissimage10.html (accessed on 26 April 2023).
a validation accuracy of 0.89. Furthermore, the training loss [4] Noel Gorelick, Matt Hancher, Mike Dixon, Simon Ilyushchenko,
was measured at 0.92, while the validation loss was recorded David Thau, Rebecca Moore, Google Earth Engine: Planetary-
at 0.922. These outcomes serve as evidence that the UNET scale geospatial analysis for everyone, Remote Sensing of
Environment, Volume 202, 2017, Pages 18-27, ISSN0034-4257,
model demonstrates a high degree of accuracy in accurately https://doi.org/10.1016/j.rse.2017.06.031.
predicting Digital Elevation Model (DEM) contour maps based [5] Google Colaboratory [Computer software]. (2023). Retrieved from
on satellite images. We evaluated our approach on a dataset https://colab.research.google.com/
[6] Y. Lecun, L. Bottou, Y. Bengio and P. Haffner, ”Gradient-based learning
of satellite images and compared the resulting DEMs to those applied to document recognition,” in Proceedings of the IEEE, vol. 86,
obtained using traditional methods. Our experimental results no. 11, pp. 2278-2324, Nov. 1998, doi: 10.1109/5.726791.
demonstrate that the UNET-based approach can produce high- [7] Goodfellow, I., Bengio, Y., Courville, A. (2016). Deep learning (Vol. 1).
MIT Press.
quality DEMs that are comparable to those obtained using [8] Cohen, C.J., 2000. Early history of remote sensing, in: Proceedings 29th
traditional methods. Applied Imagery Pattern Recognition Workshop. Washington, DC, USA,
pp. 3–9. doi:10.1109/AIPRW.2000.953595
V. CONCLUSION AND FUTURE SCOPE
In conclusion, DEM extraction from satellite imagery is
a crucial process for creating highly accurate digital eleva-
tion models that can be used in a variety of applications.
Advances in satellite technology have made it possible to
collect vast amounts of high-resolution imagery, making it
easier to create detailed DEMs. Various methods, such as
stereo-photogrammetry and LiDAR, have been used to extract

6
Authorized licensed use limited to: Somaiya University. Downloaded on April 17,2025 at 17:20:24 UTC from IEEE Xplore. Restrictions apply.