Temporal RX-algorithm performance on Sentinel-2 images
(2024) In Master Thesis in Geographical Information Science GISM01 20241Dept of Physical Geography and Ecosystem Science
- Abstract
- The increasing availability of satellite images with seemingly ever increasing spatial, spectral, and temporal resolution is a treasure when searching for information about activity on the surface of the earth. However, the large amount of data is challenging for humans to search through. Anomaly detection in time series of satellite images could potentially help humans to search specific areas for interesting changes reduce the workload on humans.
The RX-algorithm is commonly used for spatial anomaly detection, but there are not that many publications concerning the RX-algorithm applied on temporal data. The studies which have been done are on desert areas, which is assumed to be less challenging than the area in this thesis. By... (More) - The increasing availability of satellite images with seemingly ever increasing spatial, spectral, and temporal resolution is a treasure when searching for information about activity on the surface of the earth. However, the large amount of data is challenging for humans to search through. Anomaly detection in time series of satellite images could potentially help humans to search specific areas for interesting changes reduce the workload on humans.
The RX-algorithm is commonly used for spatial anomaly detection, but there are not that many publications concerning the RX-algorithm applied on temporal data. The studies which have been done are on desert areas, which is assumed to be less challenging than the area in this thesis. By changing the direction of the data sampling from spatial, meaning sampling of several pixels in the x, y direction of one satellite image, to the temporal direction meaning pixels with the same x, y location from several images taken at different times, the algorithm can be used for temporal anomaly detection.
In this thesis the performance of the temporal RX-algorithm is assessed by applying it on a time series of Sentinel-2 images and comparing the algorithms classification against the classification made by a human. All the classifications are binary, in this assignment meaning that it is either no change detected or change detected.
Spectral analysis of pixels with 10x10 meter resolution is challenging, and the quality of the data is crucial for the results. Mixed pixels, co-registration errors and scattered irradiance are present to some extent in all satellite images. The areas selected for this study are chosen to test the performance in areas where these errors are likely to be present.
Confusion matrixes are used to interpret the performance by analyzing the number of true positives and negatives, and false positives and negatives. It does also keep the spatial location of the performance metrics, so that the position of the errors could be analyzed. This indicates which type of errors that the misclassification could be caused by. The temporal dimension is tracked, and when in time the anomalies are introduced is visible in figures which can be compared to the corresponding image.
This project aims to broaden the understanding of the temporal RX-algorithm´s performance on time series of satellite images where the environment is more complex than a desert. It will hopefully shed light on the limitations and possibilities for such automated approaches to anomaly detection that could possibly aid humans when interpreting images in the search of anomalies with unknown spectral signatures. (Less) - Popular Abstract
- The surge in satellite technology has made high-resolution images of Earth's surface more accessible than ever. These images, with a high level of spatial, spectral, and temporal details, hold immense potential for monitoring changes and activities on the ground. However, the sheer volume of data presents a significant challenge for human analysts. To address this, anomaly detection in time series of satellite images offers a new method to detect changes. This thesis tests temporal RX-algorithm performance on Sentinel-2 imagery.
A well-known method for anomaly detection is the RX-algorithm. However, its application to temporal data is less explored. Instead of examining different pixels within a single image (spatial sampling), this... (More) - The surge in satellite technology has made high-resolution images of Earth's surface more accessible than ever. These images, with a high level of spatial, spectral, and temporal details, hold immense potential for monitoring changes and activities on the ground. However, the sheer volume of data presents a significant challenge for human analysts. To address this, anomaly detection in time series of satellite images offers a new method to detect changes. This thesis tests temporal RX-algorithm performance on Sentinel-2 imagery.
A well-known method for anomaly detection is the RX-algorithm. However, its application to temporal data is less explored. Instead of examining different pixels within a single image (spatial sampling), this approach analyzes the same pixel across multiple images taken over time (temporal sampling). Previous research has primarily focused on desert landscapes. This thesis tests the temporal RX-algorithm in a more complex environment.
The study evaluates the temporal RX-algorithm using a series of Sentinel-2 satellite images. The algorithm's performance is compared to human classification, employing a binary system where changes are either detected or not detected. High-resolution spectral analysis at a 10x10 meter scale is challenging due to various image imperfections such as mixed pixels, co-registration errors, and scattered irradiance. The selected study areas were chosen specifically because they are prone to these errors, providing more challenging areas for the algorithm to classify.
The goal of this research is to expand the understanding of the temporal RX-algorithm's capabilities in more complex environments than those studied earlier. By highlighting both the limitations and potential of this approach, the study aims to enhance its utility in aiding human analysts searching for anomalous changes in satellite images. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9155694
- author
- Storsnes, Martin LU
- supervisor
- organization
- course
- GISM01 20241
- year
- 2024
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Geography, Geographical Information Systems, GIS, Change Detection, Anomaly Detection, Temporal RX
- publication/series
- Master Thesis in Geographical Information Science
- report number
- 177
- language
- English
- additional info
- External supervisor: Thomas Olsvik Opsahl, Norwegian Defence Research Establishment
- id
- 9155694
- date added to LUP
- 2024-06-03 20:44:54
- date last changed
- 2024-06-03 20:44:54
@misc{9155694,
abstract = {{The increasing availability of satellite images with seemingly ever increasing spatial, spectral, and temporal resolution is a treasure when searching for information about activity on the surface of the earth. However, the large amount of data is challenging for humans to search through. Anomaly detection in time series of satellite images could potentially help humans to search specific areas for interesting changes reduce the workload on humans.
The RX-algorithm is commonly used for spatial anomaly detection, but there are not that many publications concerning the RX-algorithm applied on temporal data. The studies which have been done are on desert areas, which is assumed to be less challenging than the area in this thesis. By changing the direction of the data sampling from spatial, meaning sampling of several pixels in the x, y direction of one satellite image, to the temporal direction meaning pixels with the same x, y location from several images taken at different times, the algorithm can be used for temporal anomaly detection.
In this thesis the performance of the temporal RX-algorithm is assessed by applying it on a time series of Sentinel-2 images and comparing the algorithms classification against the classification made by a human. All the classifications are binary, in this assignment meaning that it is either no change detected or change detected.
Spectral analysis of pixels with 10x10 meter resolution is challenging, and the quality of the data is crucial for the results. Mixed pixels, co-registration errors and scattered irradiance are present to some extent in all satellite images. The areas selected for this study are chosen to test the performance in areas where these errors are likely to be present.
Confusion matrixes are used to interpret the performance by analyzing the number of true positives and negatives, and false positives and negatives. It does also keep the spatial location of the performance metrics, so that the position of the errors could be analyzed. This indicates which type of errors that the misclassification could be caused by. The temporal dimension is tracked, and when in time the anomalies are introduced is visible in figures which can be compared to the corresponding image.
This project aims to broaden the understanding of the temporal RX-algorithm´s performance on time series of satellite images where the environment is more complex than a desert. It will hopefully shed light on the limitations and possibilities for such automated approaches to anomaly detection that could possibly aid humans when interpreting images in the search of anomalies with unknown spectral signatures.}},
author = {{Storsnes, Martin}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master Thesis in Geographical Information Science}},
title = {{Temporal RX-algorithm performance on Sentinel-2 images}},
year = {{2024}},
}