Computer Science – Learning
Scientific paper
Dec 2010
adsabs.harvard.edu/cgi-bin/nph-data_query?bibcode=2010agufm.p53a1501l&link_type=abstract
American Geophysical Union, Fall Meeting 2010, abstract #P53A-1501
Computer Science
Learning
[5464] Planetary Sciences: Solid Surface Planets / Remote Sensing, [6250] Planetary Sciences: Solar System Objects / Moon
Scientific paper
Our current work aims to develop the most reliable model for estimating lunar mineral abundances. In our previous studies, partial least squares (PLS) and genetic algorithm - partial least squares (GA-PLS) models have been used. PLS has two limitations: 1) redundant spectral bands cannot be removed effectively; and 2) non-linearity between lunar soil reflectance spectra and lunar mineral abundances cannot be accommodated. GA is an effective tool for selecting a set of spectral bands that are most sensitive to lunar minerals, and to some extent overcomes the first limitation. One objective of this study is to compare the PLS and GA-PLS performances in estimating lunar mineral abundances. Given the fact that GA-PLS is still subjected to the effect of nonlinearity, a hybrid PLS-BPNN model is developed and tested to determine how effective back propagation neural network (BPNN) is for overcoming the two limitations. BPNN accommodates nonlinearity with the sigmoid functions connecting BPNN layers of nodes, and the weights of redundant spectral bands are significantly decreased through the learning process. The lunar soil characterization consortium (LSCC) dataset is the only complete 'ground truth' data of the Moon consisting both soil reflectance spectra and mineral abundances. The LSCC dataset is composed of totally 19 Apollo samples and each sample has four particle size groups (< 10 µm, 10 - 20 µm, 20 - 45 µm, < 45 µm). Although the mineral abundances of the group < 45 µm are not measured, the mineral abundances of this group (validation set) are assumed to be the average of the mineral abundances of samples in other three groups (calibration set). PLS, GA-PLS and PLS-BPNN are assessed based on R-squares and relative root mean square error for the validation. The results indicates that GA-PLS performs better than PLS for retrieving the mineral abundance of agglutinate, plagioclase, pyroxene, and volcanic glasses (Figure 1), but PLS is preferred over GA-PLS for modeling the abundance of olivine and ilmenite (Figure 1). Comparison among the three models indicates that PLS-BPNN performs significantly better than both PLS and GA-PLS for estimating agglutinate, pyroxene, olivine, ilmenite, and volcanic glasses (Figure 1). For estimation of plagioclase, both PLS-BPNN and GA-PLS perform equally well. Overall PLS-BPNN is preferred over both PLS and GA-PLS for quantifying all lunar surface dominant minerals. Figure 1. Comparison of R- squares and relative root mean square errors resulting from the PLS, GA-PLS and PLS-BPNN validation.
Li Lexin
Li Shiheng
No associations
LandOfFree
Comparison of three different statistical methods for retrieving the lunar mineral abundance does not yet have a rating. At this time, there are no reviews or comments for this scientific paper.
If you have personal experience with Comparison of three different statistical methods for retrieving the lunar mineral abundance, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Comparison of three different statistical methods for retrieving the lunar mineral abundance will most certainly appreciate the feedback.
Profile ID: LFWR-SCP-O-1498912