Machine Learning for Planetary Science

Machine Learning for Planetary Science presents planetary scientists with a way to introduce machine learning into the research workflow as increasingly large nonlinear datasets are acquired from planetary exploration missions. The book explores research that leverages machine learning methods to enhance our scientific understanding of planetary data and serves as a guide for selecting the right methods and tools for solving a variety of everyday problems in planetary science using machine learning. Illustrating ways to employ machine learning in practice with case studies, the book is clearly organized into four parts to provide thorough context and easy navigation.

The book covers a range of issues, from data analysis on the ground to data analysis onboard a spacecraft, and from prioritization of novel or interesting observations to enhanced missions planning. This book is therefore a key resource for planetary scientists working in data analysis, missions planning, and scientific observation.

• Includes links to a code repository for sharing codes and examples, some of which include executable Jupyter notebook files that can serve as tutorials
• Presents methods applicable to everyday problems faced by planetary scientists and sufficient for analyzing large datasets
• Serves as a guide for selecting the right method and tools for applying machine learning to particular analysis problems
• Utilizes case studies to illustrate how machine learning methods can be employed in practice

• Language: English
• Publisher: Elsevier Science
• Release date: Mar 22, 2022
• ISBN: 9780128187227

Book preview

Chapter 1: Introduction to machine learning

Hannah Kernera; Joseph Campbellb; Mark Stricklandc    aUniversity of Maryland, College Park, MD, United States

bCarnegie Mellon University, Pittsburgh, PA, United States

cArizona State University, Tempe, AZ, United States

Abstract

Planetary science missions and investigations are collecting increasingly large and complex datasets. Autonomy and machine intelligence are playing an increasingly important role in interpreting data from planetary missions. There is a need for machine learning and data science methods that can rapidly and intelligently extract information from these datasets in a useful manner for scientific analysis. This chapter aims to facilitate the development and adoption of machine learning methods for planetary science by providing an understanding of key machine learning concepts and methods. We focus on techniques that are most relevant for planetary science and include domain-specific examples to illustrate the methods.

Keywords

machine learning; planetary science; supervised learning; semisupervised learning; unsupervised learning; active learning

1.1 Overview of machine learning methods

A widely used definition of machine learning (ML) comes from Carnegie Mellon University professor Tom Mitchell: "A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E, or more simply, machine learning is the study of computer algorithms that improve automatically through experience" [21]. In other words, machine learning involves using computer programs to learn some task based on experience (in the form of data or observations) where how well that task is learned is measured by some performance metric. For example, your email spam filter is a machine learning system: a computer program learns to classify emails as spam or not spam by observing millions of example emails that were known to be either spam or not spam, and how well it has learned to do this is measured by the percentage of correctly classified emails (the accuracy). Much of machine learning is considered in the optimization context, where the learning task is defined as learning the parameters θ (weights or coefficients) of a function (where x is the input feature or features) by optimizing an objective function where y is the correct value or label (e.g., spam or not spam) and is the predicted value or label [21]. The objective function (also called a loss function) compares y to and quantifies how close the prediction is to the correct value.

Machine learning describes a subset of methods within the broader field of artificial intelligence (AI). Computer vision is also a subarea of AI that describes methods for perception (including recognition and understanding) of images, videos, or other visual content. There is substantial overlap between computer vision and machine learning since machine learning methods (such as convolutional neural networks) are often employed for computer vision tasks (such as detecting craters in a satellite image of Mars), though computer vision may also draw from nonML methods (e.g., SIFT, or scale-invariant feature transform [20]). The focus of this book is on the broader category of machine learning methods applied to planetary science, which includes but is not limited to computer vision tasks.

Machine learning methods are typically divided into the following subtypes:

(1)  Supervised learning

(2)  Unsupervised learning

(3)  Semisupervised learning

Each of these methods represents a different type of learning based on what data is available as input to the learning model. In supervised learning, the data provided includes both a feature vector and a label. The feature vector is a value (or set of values) intended to represent the observable feature(s), for a particular data sample. The label, in turn, is intended to represent a correct output for that sample. For example, one sample may consist of a feature vector in the form of an image (represented as a vector of pixel values) along with a label in the form of an integer that represents the category or class that describes the image. Fig. 1.1 illustrates an example input for classifying whether an image contains a crater or not: the input features are the pixels from the image and the class label is 1 if the image contains a crater and 0 if it does not. This is an example of a type of supervised learning called classification, in which the machine learning model maps input features to discrete output classes. The second type supervised learning is called regression, in which the model maps input features to a continuous output value (i.e., a real number). As an example, a model could be trained to predict the eccentricity of a crater in an image: the input features are the pixel values and the output is the real-valued eccentricity of the crater (Fig. 1.1).

Figure 1.1 Example input and output pair for supervised classification (top) and regression (bottom).

In unsupervised learning, the input data is similar but no labels are available for training. In the following subsections, we will discuss methods for supervised and unsupervised learning in more detail. In semisupervised learning, models make use of both labeled and unlabeled data samples during training—in many applications there are far more unlabeled examples available than labeled examples. This is especially true in planetary science applications, partly because labels are more difficult to acquire than traditional machine learning applications because labeling may require domain knowledge or may not be clearly defined. For many planetary science applications, there may be very little data–labeled or unlabeled–requiring specialized methods. Section 1.4 will cover methods for dealing with small labeled datasets, including semisupervised learning.

1.2 Supervised learning

Supervised learning refers to the set of machine learning techniques in which the model learns from input examples that contain an associated label for every example. There are two main categories of supervised learning techniques: classification, in which the model prediction is a discrete class or label, and regression, in which the model prediction is continuous (i.e., a real number).

1.2.1 Classification

Though supervised classification problems have inspired the majority of machine learning solutions, many real-world applications require significant effort to compile a training dataset since the data samples are usually labeled by humans. For example, in order to train a model to classify whether an image contains a crater or not (Fig. 1.1), one must first assemble a dataset containing hundreds or thousands of examples of images and an associated label of crater or noncrater. Techniques for labeling planetary science datasets are discussed in more detail in Chapter 5 while techniques for making use of a small number of labeled examples are discussed in Section 1.4. Classification involves predicting a class that corresponds to a given input. The learning model is essentially a function that maps the input feature values to the output class value, as illustrated in Fig. 1.2.

Figure 1.2 Classifier as function.

One of the simplest classifiers is a linear classifier, in which the function that maps input to output is a linear function. For example, a simple linear function of the input that distinguishes two classes is:

(1.1)

where x is the input feature vector, W is called the weight matrix, and is called the bias. For this function, a decision rule can be used to distinguish the classes:

(1.2)

The decision boundary is therefore defined by the linear relation , which is a hyperplane of dimension where N is the dimensionality of the input features. One way to illustrate a linear classifier is shown in Fig. 1.3a for the case of a 2-dimensional feature vector. The blue (dark gray in print version) circles represent data samples from class while the red (mid gray in print version) squares are data samples from class . The magenta (light gray in print version) line represents the decision boundary at that separates the classes.

Figure 1.3 Linear decision boundary separating data samples from two classes.

Although the boundary shown in Fig. 1.3a is a simple line, the data samples that correspond to each class may not be linearly separable (e.g., Fig. 1.3b). In such cases, models that are capable of finding nonlinear decision boundaries (e.g., support vector machines and other kernel-based methods, or neural networks) or different feature extraction methods that enable the classes to be linearly separable are required.

A common obstacle in developing supervised classifiers is overfitting, which occurs when the classification model parameters are designed to perform particularly well for the training data, but the model performs relatively poorly when tested with data unseen during training (e.g., testing or validation data). Fig. 1.4 shows an example of a decision boundary for which overfitting has occurred. This problem can occur when the model parameters are specifically tailored to the training samples used rather than the underlying distribution of the samples. In this example, a higher-order polynomial was chosen to separate the classes and (magenta (light gray in print version) decision boundary), but the underlying distribution likely could be approximated using a simpler (e.g., linear) decision boundary and fewer model