11:00 AM - 1:00 PM
[PPS06-P02] Application of the Latest Classification Algorithms to Lunar Reflectance Spectral Data and its Evaluation
Keywords:moon
A geological map is an important information for understanding a terrestrial body. Multiple kinds of observation data concerning the surface are combined for the geologic interpretation. VIS-NIR reflectance spectra on the Moon include key information of the distributed minerals. The global lunar classification map of absorption spectra by K-means was proposed by Hareyama et al. (2019)[1], where the K value (classification number) was assumed to be 7 following the well-known geology. We attempt to classify the whole Moon with the optimal number of clusters. We aim to contribute to updating the classification map of the Moon in a data-oriented way.
In this study, we apply the latest unsupervised classification algorithms to the lunar reflection spectral data. There are two types of unsupervised learning clustering: non-hierarchical clustering and hierarchical clustering. We adopt both: Fuzzy K-means, a non-hierarchical clustering, and DBSCAN, a hierarchical and density-based spatial clustering. The results of applying both algorithms to the lunar reflectance spectral data are compared with existing geological maps [2] for evaluation, respectively. The used spectral data was obtained by the Spectral Profiler (SP) onboard the lunar orbiter KAGUYA [3]. We use the clustering tool RasterMiner [4], which implements the latest algorithms, for classification. Fuzzy K-means is called as soft clustering. It classifies clusters as "Fuzzy" without clearly determining the cluster affiliation. Fuzzy K-means is different from simple K-means, which is called as hard clustering where the clusters are classified clearly. Fuzzy K-means is applied after determining the number of classifications in advance by Elbow K-means.
Based on the results of Elbow K-means, Then we performed Fuzzy K-means with K = 4, 5, 6, 7, 8. With K = 4, the resultant global lunar classification map identified the major regions of the Moon (Highlands of 2 clusters, Mare as 1 cluster, Mare margins with SPA as 1 cluster), but the divisions were not so good. With K=5, the highlands were classified into 2 clusters. With K = 6, the area around Mare has been divided into smaller pieces. With K >= 7, a new cluster emerged just in the boundary between the two classified units around the Mare Orientale. A smaller K value results in insufficient division, and a larger K values may lead to over classification. DBSCAN needs no classification numbers in advance. Two parameters are used for the DBSCAN classification; E as the radius when a certain data is the center and MinPoints (MinPts) as the minimum number of points that exist within E. In DBSCAN, we examined the results and trends when the two parameters were changed. The resultant distribution of clusters based on DBSCAN (epsiron=0.9, MinPoints=2) seems to show some correspondence with the plagioclase abundance (wt%) in FHT (Farside Highlands Terrane). On the other hand, Mare did not become a cluster including a large number of elements (>5000) with any parameter set.
[1] Hareyama et al. (2019), "Global classification of lunar reflectance spectra obtained by Kaguya (SELENE): Implication for hidden basaltic materials," Icarus, 321, 407-425, https://doi.org/10.1016/j.icarus.2018.11.016.
[2] Fortezzo, C. M., Spudis, P. D. and Harrel, S. L. (2020), Release of the Digital Unified Global Geologic Map of the Moon At 1:5,000,000- Scale, Proceedings of the 51st Lunar and Planetary Science Conference, Houston, TX, LPI, #2710.
[3] JAXA KAGUYG (SELENE) web site, https://www.selene.jaxa.jp/index_e.htm (accessed on Feb. 2022).
[4] R. Uday Kiran (2021), Discovering Knowledge Hidden in Raster Images using RasterMiner, Proceedings of the 2021 Workshop on Intelligent Cross-Data Analysis and Retrieval (ICDAR ’21), Taipei, Taiwan, ACM, #1.
https://doi.org/10.1145/3463944.3472812.
In this study, we apply the latest unsupervised classification algorithms to the lunar reflection spectral data. There are two types of unsupervised learning clustering: non-hierarchical clustering and hierarchical clustering. We adopt both: Fuzzy K-means, a non-hierarchical clustering, and DBSCAN, a hierarchical and density-based spatial clustering. The results of applying both algorithms to the lunar reflectance spectral data are compared with existing geological maps [2] for evaluation, respectively. The used spectral data was obtained by the Spectral Profiler (SP) onboard the lunar orbiter KAGUYA [3]. We use the clustering tool RasterMiner [4], which implements the latest algorithms, for classification. Fuzzy K-means is called as soft clustering. It classifies clusters as "Fuzzy" without clearly determining the cluster affiliation. Fuzzy K-means is different from simple K-means, which is called as hard clustering where the clusters are classified clearly. Fuzzy K-means is applied after determining the number of classifications in advance by Elbow K-means.
Based on the results of Elbow K-means, Then we performed Fuzzy K-means with K = 4, 5, 6, 7, 8. With K = 4, the resultant global lunar classification map identified the major regions of the Moon (Highlands of 2 clusters, Mare as 1 cluster, Mare margins with SPA as 1 cluster), but the divisions were not so good. With K=5, the highlands were classified into 2 clusters. With K = 6, the area around Mare has been divided into smaller pieces. With K >= 7, a new cluster emerged just in the boundary between the two classified units around the Mare Orientale. A smaller K value results in insufficient division, and a larger K values may lead to over classification. DBSCAN needs no classification numbers in advance. Two parameters are used for the DBSCAN classification; E as the radius when a certain data is the center and MinPoints (MinPts) as the minimum number of points that exist within E. In DBSCAN, we examined the results and trends when the two parameters were changed. The resultant distribution of clusters based on DBSCAN (epsiron=0.9, MinPoints=2) seems to show some correspondence with the plagioclase abundance (wt%) in FHT (Farside Highlands Terrane). On the other hand, Mare did not become a cluster including a large number of elements (>5000) with any parameter set.
[1] Hareyama et al. (2019), "Global classification of lunar reflectance spectra obtained by Kaguya (SELENE): Implication for hidden basaltic materials," Icarus, 321, 407-425, https://doi.org/10.1016/j.icarus.2018.11.016.
[2] Fortezzo, C. M., Spudis, P. D. and Harrel, S. L. (2020), Release of the Digital Unified Global Geologic Map of the Moon At 1:5,000,000- Scale, Proceedings of the 51st Lunar and Planetary Science Conference, Houston, TX, LPI, #2710.
[3] JAXA KAGUYG (SELENE) web site, https://www.selene.jaxa.jp/index_e.htm (accessed on Feb. 2022).
[4] R. Uday Kiran (2021), Discovering Knowledge Hidden in Raster Images using RasterMiner, Proceedings of the 2021 Workshop on Intelligent Cross-Data Analysis and Retrieval (ICDAR ’21), Taipei, Taiwan, ACM, #1.
https://doi.org/10.1145/3463944.3472812.