Guan Applies GeoComputational Intelligence and High-Performance Computing to Land Dynamics Modeling

Guan and colleagues used hyperspectral imaging along the North Platte River to detect reeds (Phragmites), shown in red. Top image is from 2006; bottom image is from 2010.
Guan and colleagues used hyperspectral imaging along the North Platte River to detect reeds (Phragmites), shown in red. Top image is from 2006; bottom image is from 2010.

by Qingfeng “Gene” Guan

I joined UNL as an assistant professor in the Center for Advanced Land Management Information Technologies (CALMIT) and School of Natural Resources (SNR) in 2009, specializing in Geographic Information Science (GIScience). My work is based on the premise that the new era of computational geography offers exciting solutions to problems that were previously unsolvable.

My research focuses on complex spatio-temporal land dynamics, which is a generic term referring to the changes that occur on the surface of land. Land dynamics primarily include land-use and land-cover changes (LULCC) such as expansion, retraction, and/or movement of uses or characteristics of land. Some commonly studied land dynamics include urban sprawl, deforestation, and plant species distribution and movement, wildfire regime and spread, cropland change and the spread of plant disease. Land dynamics typically represents one of the most significant challenges in many interrelated sciences, and changes in land use and conditions yield considerable impacts on environmental and socio-economic systems. Understanding land dynamics has become critical in a wide range of scientific research and management practices.

Two major challenges of studying land dynamics have been identified: (1) a specific type of land dynamics is often an extremely complex process that involves multiple nonlinear intra-/inter- relationships/actions among a variety of natural and socio-economic factors, which leads to the fact that our current knowledge and data are often incomplete, and possibly biased; (2) the analysis and modeling of land dynamics often rely on massive amounts of data and complicated processing algorithms, which require extensive, sometimes even intractable, lengths of computing time.

To tackle these challenges, I take a computational thinking approach, and explore combinations of geospatial analytical and modeling techniques and computational science methods, especially methods of computational intelligence (CI) and high-performance computing (HPC). Geospatial analytical and modeling techniques, e.g., cellular automata (CA), agent-based modeling (ABM), and geostatistics are capable of representing complex, collective spatio-temporal dynamics using simple, individual interactions and behaviors, and generating highly visible results. CI methods, such as artificial neural network (ANN) and genetic algorithm (GA), provide adaptive approaches to solving complex problems to which traditional statistical methods may not be applicable, and can be used to identify, measure, and simulate the relationships and interactions between the factors involved in land dynamics. Also, HPC utilizes multiple interconnected computing units (e.g., CPU cores, GPU cores, and PCs) to work on a common task in parallel so as to greatly reduce the computing time and generate results faster.

I am currently working with Dr. Steve Young of the West Central Research and Extension Center of UNL and Dr. Sunil Narumalani of Geography at UNL on the project “Invasive Plant Species Management with Geospatial Information Technologies and Computational Science,” funded by the Strategic Investment Seed Grants for Integrated Projects Program of UNL's Institute of Agriculture and Natural Resources. In this project, an Airborne Imaging Spectroradiometer for Applications (AISA) Eagle hyperspectral imaging system is used to acquire high spatial resolution (1.5 by 1.5 meters), multispectral (62 spectral bands in visible and near-infrared region) images of the North Platte River area of Nebraska in spring and fall. Combined with GPS-based field sample point data, these images are used to produce a time series of maps of the study area delineating and quantifying the spatial distribution of the common reed (i.e., Phragmites) and other land covers. We are now developing a spatially explicit model that combines ANN, CA, and ABM to simulate the spatio-temporal dynamics of the invasive species. Training the ANN using the historical data and other climatic and environmental data will allow it to learn the complex relations and interactions between species and the environment, and to be able to generate forecasts of future scenarios.

I am collaborating with researchers at Georgia Institute of Technology to develop a high-performance CA-based urban land-use change model using Graphics Processing Units (GPUs). Our pilot study implemented a classical CA model, the Game of Life (GOL), on both a GPU-equipped desktop PC and the Keeneland, a hybrid computer system sponsored by NSF that combines multiple CPUs and GPUs. The GPU-GOL was able to reduce the computing time to six minutes, which was 16.7 times faster than a CPU-based GOL, which requires about 100 minutes to complete. The Keeneland-GOL reduced the computing time to 20 seconds, which was 300 times faster, using 20 GPUs. With such promising results, we expect to reduce the calibration time of a CA-based urban land-use model from thousands of hours to just a few hours, and to remove the simplifying assumptions used in previous studies so that the modeling accuracy can be significantly improved.

These techniques are helping solve problems once considered impossible. For example, applying computational intelligence and using high-performance computing will allow the SLEUTH model developed by Dr. Keith Clarke to be calibrated to model urban growth across the entire U.S. Without these techniques, the comprehensive calibration for a medium-sized dataset may take 1,200 CPU hours, which isn't feasible, but with computational intelligence, we can calibrate the model more efficiently by finding the best-fit parameters without examining all possible parameter values, which will significantly reduce computational intensity. With high-powered computing, we can reduce the computing time to a manageable length.

I am also working closely with the National Drought Mitigation Center (NDMC) at UNL. We have developed the next generation of the web-based U.S. Drought Impact Reporter (DIR) mapping service (http://droughtreporter.unl.edu) using dynamic web-GIS technologies. The DIR mapping service provides users a rich set of query and display options, and is capable of querying the back-end database at various spatial, temporal, and categorical scales in real time, visualizing spatial distributions of drought impacts and reports at both state and county levels, and displaying summarized and detailed information about drought impacts and/or reports dynamically in response to users’ interactions with the map. We are now developing new tools that allow users to query and download DIR data for arbitrary shapes of areas other than states and counties, such as watersheds, congressional districts, climate divisions, and Risk Management Agency (RMA) regions. We are also developing methods to statistically compare the U.S. Drought Monitor data (http://droughtmonitor.unl.edu) and DIR data and to study the spatio-temporal patterns of both.

I believe that the advancements in GeoComputational Intelligence and High-performance Geospatial Computing will lead to revolutionary outcomes in geospatial sciences and other related disciplines, including natural resources and ecology. They will provide innovative means to test and verify classical theories as well as stimulate new theory.