|
|
|
Automated Orbital Mapping: Statistical Data Mining of Orbital Imagery to Analyze Terrain, Summarize its Characteristics and Draft Geologic Map
|
|
|
| Planetary exploration has entered a new era in which our data-gathering capability has outpaced our capacity for timely analysis. A
combination of high-resolution instruments, spacecraft autonomy and mobility, and extended lifetimes is adding to a vast library of
image and spectral data. The Mars Reconnaissance Orbiter is the latest following this trend, with HiRISE and MARCI images
providing unprecedented resolution of surface features. The catalog of images will continue to grow. The amount of resources devoted
to analysis remains mostly static, so we require automated methods to improve image analysts' productivity. We address the analysis
bottleneck in data analysis by creating methods to automatically generate maps from image and spectral data. Our method exploits
recent advances in machine learning and object recognition. The system will learn to infer mapping rules statistically from "training"
data in the form of existing geologic maps and registered orbital images. This is a natural application of existing learning-based
methods of image segmentation and boundary detection. We will extend these methods to the orbital mapping domain with
geologically-relevant descriptors like texture, spectral signatures, and surface features like boulders, craters, and dunes. The result is a
system that uses features to predict the locations of geologic units and their boundaries. Automating geologic mapping will yield
immediate benefits for the science community. It will dramatically improve the speed of image analysis which is limited by manual
feature labeling. Maps offer fast summaries of gigapixel images that may include hundreds of thousands of features. These "draft"
maps can be refined by human experts in less time than it would take to construct an entire map from scratch. Our mapping system
will also permit new classes of spatial data mining. Exhaustive statistical analysis of image libraries can quantify trends and discover
hidden correlations and anomalies. The system will yield summary statistics that improve geologists' ability to understand large
datasets rapidly. Automated mapping can also improve the quality of data analysis: it can quickly examine thousands of images,
perfectly, consistently and without fatigue. This reduces the need for redundant analysis for quality assurance. Automated mapmaking
is not intended to replace human expertise; both human and automated image analysis play to different strengths. The human can
offer expert geologic insight and authoritative segmentations of geologic units. The information system can examine many images
quickly, quantify features with accuracy and consistency, and perform exhaustive statistical analysis to search for unexpected patterns
and anomalies. In this sense the roles of the geologist and the automated mapping system are strongly complementary. The mapping
system automates some of geologists' more arduous tasks, freeing them to focus on interpretation. |
|
|
