Comparing multispectral and hyperspectral data for savannah tree species classification using machine learning
(2025) In Student thesis series INES NGEM01 20251Dept of Physical Geography and Ecosystem Science
- Abstract
- This thesis explores the classification of savannah trees using EnMAP hyperspectral and Sentinel-2 multispectral satellite data using a Random Forest classifier. The study aimed to assess if the increase in spectral information from the hyperspectral data gives an increase in classification performance, to find the advantages and limitations to hyperspectral data compared to multispectral data and if the most important bands for this tree species classification which are identified from the hyperspectral data are fully covered by the multispectral sensor. The classification found that using the EnMAP data produced an accuracy of 74% and an F1 score of 0.65, while using Sentinel-2 data produced a very similar result with an accuracy of 71%... (More)
- This thesis explores the classification of savannah trees using EnMAP hyperspectral and Sentinel-2 multispectral satellite data using a Random Forest classifier. The study aimed to assess if the increase in spectral information from the hyperspectral data gives an increase in classification performance, to find the advantages and limitations to hyperspectral data compared to multispectral data and if the most important bands for this tree species classification which are identified from the hyperspectral data are fully covered by the multispectral sensor. The classification found that using the EnMAP data produced an accuracy of 74% and an F1 score of 0.65, while using Sentinel-2 data produced a very similar result with an accuracy of 71% and an F1 score of 0.65, found to be statistically insignificant using a McNemar’s statistical test. The advantage of the hyperspectral data is the detailed spectral information with more bands in the most important wavelengths for tree species classification. The increased spatial resolution of the Sentinel-2 data gave a small advantage to the classification, as resampling the Sentinel-2 data to 30 m caused a marginal decrease in performance. The most important EnMAP bands were identified as 1749 nm, 1756 nm, 1996 nm, 2157 nm and 2330 nm. However, only 2157 nm is covered by Sentinel-2, inside band 12. Overall, the thesis found that multispectral data with fewer but more carefully positioned bands can produce very similar classification results to hyperspectral data, even in challenging areas such as savannah tree species classification. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9205427
- author
- Doherty, Joseph Patrick LU
- supervisor
-
- Lars Eklundh LU
- Zhanzhang Cai LU
- organization
- course
- NGEM01 20251
- year
- 2025
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Physical Geography and Ecosystem analysis, remote sensing, multispectral, hyperspectral, savannah
- publication/series
- Student thesis series INES
- report number
- 734
- language
- English
- id
- 9205427
- date added to LUP
- 2025-06-25 09:15:39
- date last changed
- 2025-06-25 09:15:39
@misc{9205427,
abstract = {{This thesis explores the classification of savannah trees using EnMAP hyperspectral and Sentinel-2 multispectral satellite data using a Random Forest classifier. The study aimed to assess if the increase in spectral information from the hyperspectral data gives an increase in classification performance, to find the advantages and limitations to hyperspectral data compared to multispectral data and if the most important bands for this tree species classification which are identified from the hyperspectral data are fully covered by the multispectral sensor. The classification found that using the EnMAP data produced an accuracy of 74% and an F1 score of 0.65, while using Sentinel-2 data produced a very similar result with an accuracy of 71% and an F1 score of 0.65, found to be statistically insignificant using a McNemar’s statistical test. The advantage of the hyperspectral data is the detailed spectral information with more bands in the most important wavelengths for tree species classification. The increased spatial resolution of the Sentinel-2 data gave a small advantage to the classification, as resampling the Sentinel-2 data to 30 m caused a marginal decrease in performance. The most important EnMAP bands were identified as 1749 nm, 1756 nm, 1996 nm, 2157 nm and 2330 nm. However, only 2157 nm is covered by Sentinel-2, inside band 12. Overall, the thesis found that multispectral data with fewer but more carefully positioned bands can produce very similar classification results to hyperspectral data, even in challenging areas such as savannah tree species classification.}},
author = {{Doherty, Joseph Patrick}},
language = {{eng}},
note = {{Student Paper}},
series = {{Student thesis series INES}},
title = {{Comparing multispectral and hyperspectral data for savannah tree species classification using machine learning}},
year = {{2025}},
}