Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

A Survey on Satellite Image Processing

Using Machine Learning
Dr. Suresh L1; Abhilash S2; Adithya M Patil3; Anirudh N A4; Chandan S M5
Department of Information Science
RNS Institute of Technology, Bengaluru

Abstract:- This survey examines the application of deep II. LITERATURE SURVEY
learning to satellite imagery for advanced Earth
observation analysis, focusing on land use, object The eight research papers examined in this survey
detection, and land mapping. The project employs present a range of approaches to satellite image processing,
Convolutional Neural Networks (CNNs) to achieve from foundational machine learning models to cutting-edge
accurate LULC classification, object detection (e.g., deep learning frameworks. Each study provides valuable
swimming pools, cars), and image segmentation tasks. perspectives on the strengths, challenges, and computational
Key applications include multi-class land cover trade-offs associated with different techniques for activities
mapping, binary segmentation for building such as LU classification, object detection, and
identification, and landslide detection, all of which play environmental monitoring. This section consolidates these
an essential role in accurately defining land features. By insights, comparing the methods on key criteria like
automating these analyses, the project facilitates rapid, accuracy, processing speed, scalability, and suitability for
scalable assessments that support timely and informed practical applications. By analyzing and contrasting these
decision-making in sectors like agriculture, forestry, and approaches, we not only identify current limitations and
disaster management, ultimately enhancing geographic gaps but also outline potential directions for enhancing
understanding and resource management. satellite image analysis, aiming to drive forward its
effectiveness and utility across fields including agriculture,
Keywords:- Satellite Image Processing, Machine Learning, urban development, and disaster response. practical use
Deep Learning, CNNs, LULC Classification, Object cases. The in-depth analysis aims not only to identify gaps in
Detection, Image Segmentation, Remote Sensing, Disaster current research but also propose future directions to
Management, Earth Observation, High-Resolution Imagery, improve image scaling methods.
Environmental Monitoring, Feature Extraction, GIS,
Landslide Detection, Land Mapping, Data Preprocessing,  Classification of Satellite Images Using Deep Learning:
Neural Networks, Satellite Imagery Analysis. This paper introduces a deep learning-based framework
designed for the automated classification and detection
I. INTRODUCTION of objects in high-resolution satellite imagery. The
framework utilizes a combination of CNNs, combined
The use of machine learning to satellite image with metadata from satellite data, to enhance accuracy
processing is transforming how we analyze and understand and adaptability across diverse classes. Experimental
Earth observation data. This rapidly evolving field blends outputs show that this ensemble approach achieves up to
remote sensing with advanced computational methods to 83% accuracy and performs reliably across multiple
transform vast volumes of high-resolution satellite images object types, achieving 95% accuracy in 15 classes. The
into actionable insights. By automating processes like system’s adaptability and accuracy make it particularly
classification, segmentation, object detection, and change suited for applications in large-scale monitoring tasks
detection, machine learning—especially deep learning—has such as environmental tracking, law enforcement, and
revolutionized traditional approaches, enabling the emergency response, where rapid, precise detection is
detection of complex patterns within large-scale datasets essential.
that were once challenging to uncover. Our project,
“Satellite Image Processing Using Machine Learning,”  Machine Learning Algorithms for Classifying Satellite
proposes a robust framework for the systematic processing Images Using Google Earth Engine and Landsat Data:
and analysis of satellite images. This framework is built This paper explores the application of six machine
around core components, including data pre-processing, learning algorithms to identify different types of LC
classification, model evaluation, and visualization. Utilizing types across Morocco using Landsat 8 satellite data
Convolutional Neural Networks (CNNs) trained on satellite within the Google Earth Engine (GEE) environment. The
data, our system performs tasks such LULC Classification, framework assesses algorithm performance based on
object detection (e.g., cars, swimming pools), image metrics like overall accuracy, Kappa coefficient, and
segmentation, and disaster event identification, such as specific class accuracy. Findings indicate that the MD
landslides and floods. Our results confirm the system’s classifier is the most effective, reaching a level of
effectiveness in accurately analyzing satellite imagery, accuracy of 93.85%, proving highly useful for mapping
enabling a wide range of applications vital to fields such as and analyzing remote and diverse landscapes. The
forestry, urban planning, and disaster management.

IJISRT24NOV1716 www.ijisrt.com 2589

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
addition of spectral indices-such as NDVI and MNDWI  Feature Extraction of Satellite Images Using Machine
further improves classification accuracy, with MD Learning: This paper explores a machine learning
demonstrating the greatest benefit from these approach, specifically Convolutional Neural Networks
enhancements. This study underscores GEE's value as a (CNNs), to extract meaningful features from high-
scalable solution for handling extensive satellite resolution satellite images. This method, suitable for
imagery, offering dependable data for environmental applications in environmental monitoring, LU planning,
monitoring andregional land management initiatives. and resource management, involves training CNN to
identify and categorize key features in satellite data, such
 Image Processing Techniques for Satellite Image as water bodies, vegetation, and urban areas. By
Analysis Using Machine Learning: This paper explores preserving spatial context and resisting changes in object
different machine learning approaches applied to position or orientation, CNNs improve accuracy over
satelliteimage analysis, focusing on methods for image traditional techniques. The paper highlights the CNN's
improvement, segmentation, extraction of key features, ability to process satellite images across multiple
fusion, and classification. The review explores the spectral bands, enabling detailed analysis and
effectiveness of each technique in managing complex classification. With the added advantage of GPU
and large satellite datasets, often affected by noise and acceleration, this method demonstrates efficient for
environmental interference. Through a comparison of large-scale satellite image analysis, making it useful and
methods, including CNN-driven segmentation and deep effective tool in applications requiring accurate and
feature extraction, the study underscores their relevance dependable feature extraction.
for applications in areas such as urban development
and environmental monitoring. The paper also  Predicting Land Use and Land Cover from Satellite
highlights the current challenges and areas for future Images Using Machine Learning Techniques: This paper
research, suggesting the potential for hybrid models and explores a machine learning approach, specifically
improvements in algorithmic resilience to meet the Convolutional Neural Networks (CNNs), to extract
demands of real-timesatellite image processing. meaningful features from high-resolution satellite
images. This method, suitable for applications in
 Convolutional Neural Network with Dilated Anchors for environmental monitoring, LU planning, and resource
Detecting Objects in Very High-Resolution Satellite management, involves training CNN to identify and
Images: This paper introduces an approach for detecting categorize key features in satellite data, such as water
objects tailored for very high-resolution satellite images. bodies, vegetation, and urban areas. By preserving
Utilizing CNN with expanded anchor boxes, the method spatial context and resisting changes in object position or
is relied on the YOLOv3 framework but modifies anchor orientation, CNNs improve accuracy over traditional
box sizes by 30–40% to better capture objects of varying techniques. The paper highlights the CNN's ability to
scales. The method demonstrates improved accuracy for process satellite images across multiple spectral bands,
detecting small, densely located objects, such as vehicles enabling detailed analysis and classification. With the
and sports courts, by using larger anchor boxes that added advantage of GPU acceleration, this method is
encompass both the object and its surroundings. efficient for large-scale satellite image analysis, making
Evaluations on the NWPU VHR-10 dataset reveal that it a useful tool in applications requiring precise feature
this enhancement boosts detection performance, making extraction.
the technique suitable for detailed satellite imagery tasks
that demand high precision.  Multiple Object Detection and Segmentation for Remote
Sensing Images: This paper introduces a CNN-based
 Designing and Evaluating a Deep Learning Model to approach for identifying and classifying features in high-
Count Objects in Satellite Images: This presents a resolution satellite images, which is important for urban
specialized deep learning model tailored for accurately planning and environmental monitoring, and resource
counting objects in satellite images, with applications in management. The approach uses CNNs to recognize
urban development, environmental monitoring, and elements such as land masses, bodies of water, and
defence. Using the YOLOv8 framework and trained on vegetation by preserving spatial information, even when
the DOTA dataset, which includes diverse aerial images objects vary in position or orientation. The methodology
spanning 15 object categories, the model incorporates includes preprocessing techniques like filtering to
advanced preprocessing techniques like Laplacian of enhance image clarity and reduce noise, optimizing
Gaussian (LOG) and Difference of Gaussians (DOG) to images for feature detection. By analyzing images
enhance image contrast and reduce noise. This setup across different spectral bands, the model achieves
enables the model to handle the unique challenges of refined classifications, while GPU acceleration speeds
satellite imagery, including small, densely packed up processing, making it feasible for large-scale datasets.
objects like vehicles and buildings. With an accuracy of This research highlights CNNs’ capabilities in satellite
80%, the model demonstrates efficiency and imagery analysis, offering a reliable solution for detailed
adaptability, offering a robust solution for extracting feature extraction in applications requiring high
essential data from high- resolution satellite images accuracy.
across various critical applications.

IJISRT24NOV1716 www.ijisrt.com 2590

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Application of Satellite Image in Disaster Detection: images for mapping and analysis.
This paper examines how satellite imagery is used to  To detect specific objects, such as cars and swimming
detect and assess the effects of natural disasters, pools, divide images into separate regions, and
emphasizing its crucial role in supporting rescue and streamline disaster detection for events like landslides
recovery operations. The research organizes disaster and floods.
detection techniques into four primary categories: rule-
based approaches, color distribution analysis, multi- III. PROPOSED SYSTEM
object detection, and end-to-end detection frameworks. It
highlights advancements in image quality achieved The proposed system uses CNNs to automate the
through Very High Resolution (VHR) imagery and extraction of complex features from satellite imagery, such
Synthetic Aperture Radar (SAR), which offer detailed, as distinguishing between land cover types, detecting
reliable data for disaster analysis. The study also structures like buildings, and identifying objects including
considers the difficulties of working with high- swimming pools and cars. This automated approach
resolution images and underscores the requirement for significantly reduces reliance on human analysis, thereby
effective feature extraction. By comparing a range of increasing the efficiency and speed of satellite data
algorithms, from traditional SVM to advanced interpretation. By training the model on a diverse range of
Convolutional Neural Networks (CNNs), the paper datasets, the system is designed to adapt across various
demonstrates how these technologies enhance disaster geographic regions and satellite image sources, ensuring
response and resourcemanagement, contributing to more reliable performance in differing environmental contexts.
timely and efficient responses to natural emergencies. This adaptability is essential for achieving high versatility
and resilience in real-world applications. Furthermore, by
 Objectives integrating deep learning techniques, the system scales to
process large volumes of images from satellites, which is
 To develop and implement CNNs tailored for tasks like crucial for timely analysis in fields such as environmental
LULC Classification, detecting objects, segmenting monitoring, urban planning, and infrastructure
images and disaster detection. development. With this capability, the system can handle
 To prepare and preprocess satellite imagery datasets for the demands of high-volume, large-scale image analysis,
deep learning applications, using tools like ArcGIS Pro making it a useful tool for rapidly responding to changes on
and the GeoTile Python library. a regional or global scale.
 To classify LULC and detect features within satellite

Fig 1 Expected Output

IJISRT24NOV1716 www.ijisrt.com 2591

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
IV. ADVANTAGES OF PROPOSED SYSTEM inconsistencies, or distortions—caused by things like
atmospheric conditions, sensor errors, or geographic shifts.
 Automated Feature Detection: By addressing these issues, we make sure the data is
Using CNNs, the system automatically identifies and consistent, reliable, and ready to support effective model
extracts complex features in satellite imagery, such as land building.
cover types, buildings, and various objects like cars and
swimming pools. This reduces the need for manual data  Model Training:
analysis, speeding up processes that would traditionally With the data prepared, we train the model by splitting
require human input. it into a training dataset for learning patterns. Techniques
like Convolutional Neural Networks (CNNs) are applied for
 Enhanced Accuracy: classifying images and segmentation, helping the model
CNNs are highly effective at capturing intricate details recognize and distinguish characteristics in the images.
and patterns, which allows the system to detect features
with high accuracy. This level of precision is essential for  Model Testing:
applications that require detailed insights, such as To see how well our model performs, we test it on new
environmental conservation and urban development. data it hasn’t seen before. By using this separate test set, we
can check whether the model can handle real-world
 Adaptability Across Diverse Environments: scenarios and make precise predictions based on new data.
The system is trained on varied datasets, which allows This step helps ensure the model to be reliable and effective
it to perform consistently across different geographic for the tasks it’s designed to do.
regions andwith imagery from different satellite sources.
 Satellite Image Processing and Prediction:
 Scalability for Large-Scale Data: Having the model trained, it’s ready to analyze new
By leveraging deep learning, the system can handle satellite images andmake predictions. Now, we can use it for
massive volumes of satellite imagery without a drop in a variety of tasks, such as classifying LULC, detecting
performance, which is critical for applications requiring objects, segmenting images, and even identifying signs of
analysis at scale, such as global environmental monitoring. disasters. This step brings the model's capabilities into real-
world applications, providing valuable insights from fresh
V. PROPOSED METHODOLOGY satellite data.

 Data Collection:  Satellite Image Processing Recommendations:

To build a reliable model, we gather satellite images Once the model analyzes and generates predictions
from a range of sources. For free, high- quality data, from satellite images, we can turn those insights into
platforms like NASA’s Landsat, ESA’s Sentinel-2, and actionable recommendations. These recommendations help
Google Earth Engine provide great options. For even higher- guide decision-making, offering practical steps based on the
resolution images, we can turn to paid services like Planet model’s findings.
Labs or Maxar. This diverse collection of data forms the
foundation for training and testing our model.  Web Interface:
A web interface makes the satellite image processing
 Data Preprocessing: system accessible and user-friendly. Through this platform,
Once we have satellite images, the following step is to various users-such as government bodies, researchers, and
clean and prepare them for use in machine learning. Satellite NGOs-can easily engage with the system, applying its
images can often come with some issues—such as noise, insights to support their unique projects and goals.

IJISRT24NOV1716 www.ijisrt.com 2592

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 2 System Design

VI. CONCLUSION [3]. N. Laban, B. Abdellatif, H. M. Ebeid, H. A.

Shedeed, and M. F. Tolba, "Convolutional Neural
 Enhanced Analysis for Earth Observation: Network with Dilated Anchors for Object Detection
Deep learning provides accurate analysis of satellite in Very High-Resolution Satellite Images," 2020 13th
images, enabling detailed Earth observation and valuable International Conference on Computer Engineering
insights from the data. and Systems (ICCES), pp. 21-26, 2020.
[4]. R. Raja Subramanian, N. Paidimarri, B. Navuluri, B.
 Broad Range of Applications: S. V. Oleti, and M. M. Boogavarapu, "Design and
These techniques support various uses like identifying Evaluation of a Deep Learning Model for Counting
land usage, detecting objects, segmenting buildings, and Objects in Satellite Images," in Proc. 4th IEEE Int.
mapping landslides or floods, making them highly useful in Conf. Data Eng. Commun. Syst. (ICDECS), 2024, pp.
areas like environmental tracking, urban development, and 1–6, doi: 10.1109/ICDECS59733.2023.10503432.
disaster management. [5]. M. Kamdi, P. Saraf, P. Lokulwar, C. Dhule, and R.
Agrawal, "Feature Extraction of Satellite Images
 Scalable and Accessible Technology: Using Machine Learning," in Proc. 2024 Int. Conf.
Through integration with platforms like Google Earth Innov. Challenges Emerg. Technol. (ICICET), 2024,
Engine, this system can handle large datasets in real time, pp. 1–7, doi: 10.1109/ICICET59348.2024.10616313.
making it scalable and accessible for advanced geospatial [6]. T. K. Das, D. K. Barik, and K. V. G. Raj Kumar,
analysis andenvironmental management. "Land-Use Land-Cover Prediction from Satellite
Images using MachineLearning Techniques," in 2022
REFERENCES International Conference on Machine Learning, Big
Data, Cloud and Parallel Computing (COM-IT-
[1]. M. Pritt and G. Chern, "Satellite Image Classification CON), Vellore, India, 2022, pp. 338-343. doi:
with Deep Learning," arXiv preprint 10.1109/COM-IT-CON54601.2022.9850602.
arXiv:2010.06497, 2020. [7]. A. Belangour, A. Erraissi, M. Wahbi, I. E. Bakali, B.
[2]. Y. Jiang, "Application of Satellite Image in Disaster Ez- Zahouani, R. Azmi, A. Moujahid, M. Zouiten, O.
Detection," 2021 2nd International Conference on Y. Alaoui, H. Boulaassal, M. Maatouk, and O. E.
Big Data & Artificial Intelligence & Software Kharki, " Applying Machine Learning Algorithms
Engineering (ICBASE), pp.524-529,2021, doi: for Classifying Satellite Images Using Google Earth
10.1109/ICBASE53849.2021.00103. Engine and Landsat Data: A Case Study of
Morocco," Remote Sensing Applications: Society and
Environment, vol. 11, pp. 71141-71142, 2023. doi:
10.1016/J.RSASE.2022.100898.

IJISRT24NOV1716 www.ijisrt.com 2593

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[8]. H. Kareemullah, D. Jose, and P. Nirmal Kumar,
"Multiple Object Detection and Segmentation for
Remote Sensing Images," in the Proceedings of the
2nd InternationalConference on Advances in
Electrical, Computing, Communication, and
Sustainable Technologies (ICAECT),
Chennai,India,2022,pp.1-5.doi:
10.1109/ICAECT54875.2022.9807848.
[9]. A. Ghorbanian, M. Kakooei, M. Amani, S. Mahdavi,
A. Mohammad Zadeh, and M. Hasanlou, “Improved
land cover map of Iran using Sentinel imagery within
Google Earth Engine and a novel automatic
workflow for land cover classification using migrated
training samples,” ISPRS Journal of Photogrammetry
and Remote Sensing, vol.167, pp.276-288,20

IJISRT24NOV1716 www.ijisrt.com 2594