Machine Learning for Future Hyperspectral Data Analysis with the Hyperscout-H instrument on the ESA’s Hera Mission

Introduction

            Hera is ESA’s first planetary defense mission designed to characterize the binary near-Earth asteroid (65803) Didymos [1]. Its secondary component, Dimorphos, was the target of the large-scale controlled collision with the NASA/DART spacecraft in September 2022 [2]. Thus, the Hera mission will provide an unprecedented opportunity to study the aftermath of the first ever artificial asteroid impact. This will be done with the help of several instruments, including the hyperspectral camera HyperScout-H (HS-H).

The goal of this work is to prepare the algorithms required to analyze the future data coming from HS-H. We aim to determine the surface morphology and its composition, including the characterization of space-weathering effects and the search for exogenous material. We are implementing machine learning (ML) approaches to classify the different surface patches using the asteroid spectral (taxonomic) classification. This abstract presents preliminary results on the spectral classification considering the nature of the data that will be acquired by HS-H. To perform several simulations, we use imaging data sourced from the JAXA/Hayabusa spacecraft equipped with the Asteroid Multi-band Imaging Camera (AMICA) [3, 4, 5, 6]. The reason for this choice is the expected similarity between the asteroid studied by JAXA/Hayabusa – (25143) Itokawa, and the target of ESA/Hera – (65803) Didymos.

Methods

            The reflectance spectrum of an asteroid, R(λ), is usually observed in the visible and near-infrared domains between 0.35 and 2.50 μm and can be classified into several types. One of the most used taxonomic classifications was defined in [7] using Principal Component Analysis (PCA) over a set of 371 spectra of asteroids. This uses letters to characterize the composition of the asteroids and it helps to perform various statistical studies over large populations of these celestial objects. Here, we develop a predictive model that accurately classifies the asteroid surface patches based on the reflectance spectrum observed by HS-H. The spectral window observed by HS-H spans the interval 0.657 – 0.949 μm. Each 5 x 5 macropixel on the mosaic filter on the HS-H detector records the signal within 25 narrow bands (25 pixels), each one with a different effective wavelength. Considering the spectral response of each band, it is challenging to accurately retrieve the true reflectance spectra due to the filter transmittances and spatial distribution.

To compare the HS-H spectrum with other spectra we need to consider the transmission functions of HS-H. Thus, we simulate an asteroid spectrum observed with HS-H. The converted reflectance is obtained by integrating over the transmission function domain in each band, i.e. ∫R(λ)τ(λ)dλ, where τ(λ) is the transmission function value at λ of one of the HS-H bands. An example is provided in Fig. 1 (a). Then, for each input data we get the equivalent HS-H observed spectrum containing 25 reflectance values computed for each one of the HS-H bands, and the spectral class that can be represented with a one-hot encoded vector. Thus, the learning problem is based on finding the best multi-class classifier. This is typically achieved by training a classifier (i.e. neural network) to predict the probability distribution over the classes by minimizing the cross-entropy loss, which is a measure of the dissimilarity between the predicted probability distribution and the true distribution.

Fig. 1 (a) Comparison between the HS-H simulated spectrum of Itokawa and the spectrum obtained from ground-based observations (b) Spectral classification map retrieved by applying the CNN on the simulated representation of Itokawa.

Data analysis and results

            We utilize the dataset described in [8] containing up to 3000 spectral samples from various asteroids. We select 11 classes that are the most relevant for our study case. We split the dataset such that 60% of the data is used for training, 20% for validation, and 20% for testing. We perform a stratified split, meaning that there are enough samples of each taxonomic class in each set. Each spectrum is interpolated by a spline curve to obtain a fit for the reflectance spectrum and to apply the transmission functions of HS-H. The data is augmented as needed by small shifts or rotations, or by adding noise like that produced by the HS-H instrument.

We implement traditional ML algorithms such as k-Nearest Neighbors, Decision Tree and Random Forest which we optimize via grid search. Also, we try different dense (DNN) and convolutional (CNN) neural network architectures aiming to get the best accuracy by width and depth scaling. A comprehensive analysis of these methods is done to choose the best classifier and obtain accurate classification maps, as the one shown in Fig. 1 (b), where we apply the CNN on the simulated data for asteroid Itokawa to determine the spectral class in each macropixel and to check the homogeneity.  

Our preliminary results on the spectral classification applied on simulated data of Itokawa are promising. A quick comparison with the images from [6] shows that we were able to identify various spectral patterns across Itokawa’s surface. We plan to simulate other test case scenarios by superimposing various patterns into the original data.

 

References
[1] P. Michel et al. The ESA Hera Mission: Detailed Characterization of the DART Impact Outcome and of the Binary Asteroid (65803) Didymos. , 3(7):160, July 2022.
[2] Andrew S. Rivkin et al.. The double asteroid redirection test (dart): Planetary defense investigations and requirements. The Planetary Science Journal, 2(5):173, aug 2021.
[3] J Saito et al. Detailed images of asteroid 25143 itokawa from hayabusa. Science (New York, N.Y.), 312:1341–4, 07 2006.
[4] Masateru Ishiguro et al. Global mapping of the degree of space weathering on asteroid 25143 Itokawa by Hayabusa/Amica observations. Meteoritics & Planetary Science, 42(10):1791–1800, 2007.
[5] Eri Tatsumi, Seiji Sugita. Cratering efficiency on coarse-grain targets: Implications for the dynamical evolution of asteroid 25143 Itokawa, Icarus, Volume 300, 2018, Pages 227-248.
[6] Sumire C. Koga et al. Spectral decomposition of asteroid Itokawa based on principal component analysis. Icarus, 299:386–395, 2018.
[7] Francesca E. DeMeo et al. An extension of the Bus asteroid taxonomy into the near-infrared. Icarus, 202(1):160–180, 2009.
[8] M. Mahlke et al. Asteroid taxonomy from cluster analysis of spectrometry and albedo. AA, 665:A26, September 2022.