Advances in remote sensing techniques have made available large datasets of topographic measurements pertaining to terrestrial and planetary land surfaces. However, the scientific utilization of these datasets is hampered by a lack of tools for effective automated analysis. This project seeks to develop a system for fast, objective and transparent conversion of topographic data into knowledge about land surfaces. The project has two complementary goals: 1) to develop a tool that autonomously produces geomorphic maps mimicking traditional, manually derived maps in their appearance and content, and 2) to develop a tool that classifies entire topographic scenes into characteristic landscape categories. The mapping tool is based on the object-oriented supervised classification principle. A number of novel solutions, including semi-supervised learning, meta-learning, and a wrapping technique coupling classification and segmentation, are proposed to address challenges posed by the specificity of topographic data. The scene classification tool is based on information-theoretic metrics and incorporates novel solutions to problems posed by the raster character of topographic datasets.
Intellectual Merit
The project employs a novel fusion of machine learning and computer vision techniques to open new possibilities. In the process of constructing the mapping and classifying tools, novel machine learning methodologies will be developed and tested. The products of this research will enable a qualitatively new type of analysis of land surface topography: the large scale statistical comparison of spatial distribution of landforms.
Broad Impact
Successful mapping and classifying tools will have impact beyond the analysis of natural landscapes; they can be also be applied to the study of surface metrology (the numerical characterization of industrial surfaces). The nature of this project will attract interest and collaboration with specialists from diverse disciplines, such as computer science, remote sensing, geomorphology and hydrology. Such links will broaden the base of expertise for each discipline, as well as enrich participants from contributing domains.
". The goal of the project is to develop a robust system for the automatic geomorphic mapping of land surfaces using a fusion of pattern recognition tools, including machine learning, computer vision, and data compression. The proposed algorithms work on remotely sensed, spatially extended data. On the technical side, we covered three major phases: 1. We built a computer algorithm that divides the different classes and subclasses of landforms on Mars using a hierarchical representation. For example, one interesting landform on Mars is that of a crater; they abound in certain regions on Mars but not so much in other regions. Craters can be further decomposed into subparts: crater walls, crater rims. crater floors, etc. Our algorithm is able to first recognize a crater as a whole, and then it can its identify different subparts (e.g., outer crater walls). 2. We performed a theoretical study to decide the best strategy to learn the hierarchical model corresponding to landforms on Mars. Specifically, we worked on the following problem. In identifying landforms on Mars, one can use a "global" model that captures all different classes at once. An alternative is to use many "local" models that capture different classes individually; local models then need to be combined into a single model. Our results show that a combination of many local models tends to outperform a global model even if the number of local models is very large (in the hundreds). 3. We developed a learning technique that captures different patterns over the surface of Mars. The attached figures show our approach. The figures show a color-labeling of Tisia Site on Mars. The first figure shows the correct labeling for different landforms on Mars. The second figure shows the predictions made by a traditional technique. The last figure shows the labeling provided by our approach. Overall our approach tends to make less mistakes and closely resembles the true class labeling. In terms of outreach and educational activities, support from NSF helped the continuation of the High School Science Club Summer Camp. The program brought together a total of approximately 60 students --from major school in the Houston area-- to work on a real scientific projects during the summers. During the span of the project we worked on fostering an appreciation for science among high school students. Students demonstrating a level of high proficiency in school were assigned to a research project under the supervision of an active scientist. Researchers and local teachers acted as mentors as the student explored, discovered, and proposed various research directions. The goal of the camp is to open a communication line between talented young students and the scientific community, by working hand-on-hand in real state-of-the-art scientific projects. Camps lasted for 6 weeks. During this time, students worked to solve the intellectual challenge posed by the researcher assigned to their team. Each week there were sessions where all the participants on the project gathered together. At the end of the six weeks we had a poster session where each team will showed the results obtained from their work. We awarded the three best projects with an honorific diploma. Additionally, in terms of graduate student involvement. A total of three graduate students participated in this project. Two MS and one Ph.D. students obtained a degree in Computer Science with a thesis related to the different phases of this project.
