Presentation given to computer science undergraduate students at the University of Houston July 2007

Transcription

1 Presentation given to computer science undergraduate students at the University of Houston July 2007

2 Machine Learning and Data Mining in Mars Tomasz F. Stepinski Lunar and Planetary Institute

3 MARS/EARTH COMPARISON MARS EARTH Diameter 4,217 miles 7,922 miles Distance from Sun Atmosphere million miles million miles 7-9 millibars; 95% CO 2, 3% N Magnetic field Extremely weak Strong Axial tilt Revolution period (year) Rotation period (day) Orbital eccentricity 687 days 365 days hrs hrs Surface gravity 3.69 m/s m/s 2 Escape velocity 11,185 mph 25,055 mph Surface temperature 1014 millibars; 77% N; 21% O to +72 F -27 to +136 F

4 Mars Orbiters Mars Exploration Rovers NASA Mars Reconnaissance Orbiter NASA Mars Odyssey Data: images, spectra taken locally ESA Mars Express Data: images, spectra, topography, Radar, etc. taken globally

5 Data from orbiters

6 More data from orbiters

7 Data from rovers

8 The role of computer science 1. Spacecraft design 2. Spacecraft navigation and operation 3. Onboard science support 4. Off-line data analysis Machine Learning and Instrument Autonomy Group Lunar and Planetary Institute

9 The Onboard Autonomous Science Investigation System OASIS system has been developed to evaluate, and autonomously act upon, science data gathered by in situ spacecraft such as planetary landers and rovers. Project OASIS

10 Identifying rocks in rover images

11 Where in the world is the Lunar and Planetary Institute? Houston Galveston Bay Gulf of Mexico NASA JSC Clear Lake

12 What is the Lunar and Planetary Institute? The LPI was established during the Apollo missions to foster international collaboration and to serve as a repository for information gathered during the early years of the space program. Today, the LPI serves as a scientific forum attracting visiting scientists, postdoctoral fellows, students, and resident experts supports and serves the research community by organizing meetings collects and disseminates planetary data educates the public about space science

13 Manual versus machine making of categorical maps Traditional manual mapping: Machine automated mapping: 1) Accurate 1) Fast and cheap 2) Slow and expensive 2) Objective 3) Subjective 3) Could be a part of database 4) detached 4) Lacks human touch

14 Two types of automatic mapping on Mars Survey of specific features Valley networks Thematic (categorical) mapping Impact craters

15 Mapping streams and valley networks

16 Terrain morphology-based valley mapping algorithm

17 Example of automatic survey: valley networks

18 Mapping craters

19 Mars has variety of different kind of craters The Catalog of Large Martian Impact Craters > 42,000 craters

20 Identification of craters from topographic data ADVANTAGES: DEM provides a direct, threedimensional representation of Martian surface. No problems associated with visibility. Possibility of calculating more parameters, such as crater depth. DISADVANTAGES: Low resolution of available data. Image, 100 meters/pixel Topography, ~500 meters/pixel

21 Identification of craters from topographic data Design criteria: Robust, works well on all types of Martian surfaces. Fast, permitting generating a catalog of craters over the entire Martian surface. Scale-independent, can be applied to other topographic datasets. Simple, can be offered as a download.

22 Final results: Tisia Valles 31 craters identified No false positives Some false negatives Catalog contains: 1) Coordinates 2) Radius 3) Shape descriptors 4) Depth

23 Automatic survey of craters on Mars Terra Cimmeria Hesperia Planum Stepinski et al., LPSC XXXVIII poster

24 Application: mapping depth-diameter ratio Terra Cimmeria region 7845 craters

25 Maps of depth-diameter ratio Implications for ground ice

26 Automatic categorical mapping on Earth Mapping of geological units from multi-spectral image Aster image, bands 9, 6, and 4 Automatic geologic map Manual geologic map Southern Mongolia Area: 1415 km 2

27 Our approach to automated mapping Method of choice: Segmentation-based classification

28 Automatic generation of geomorphic maps: Segmentation 6593 segments For each pixel we assign a vector of topographic features: slope curvature flood x an y coordinate The site is segmented into a given number of segments so each segment is approximately uniform in all features.

29 Automatic generation of geomorphic maps: Training set 829 classified segments

30 Automatic generation of geomorphic maps: Classification Naïve Bayes Trees C4.5 algorithm Support vector machines

31 Classification: other sites Vichada Valles Al-Qahira

32 Classification: other sites

33 Near term plans and prospects: 1. Further development of automatic mapping method 2. Use for a recognition of crater parts (floor, walls, rim, etc.). 3. Application to other planets: Earth, Moon. 4. Ultimate application: automated comparative analysis of landscapes.