Giorgos Mountrakis

Associate Professor in GIS and Remote Sensing, Dept. of Environmental Resources Engineering, SUNY ESF

Posts in the active research category

Collaborators: Dr. Giorgos Mountrakis, Dr. Bill Porter, Dr. Colin Beier, Dr. Ben Zuckerberg, Dr. Lianjun Zhang, Dr. Bryan Blair

Bird Distribution with NY State (credit B. Zuckerberg)

Bird Distribution within NY State (credit B. Zuckerberg)

Funders: NASA Biodiversity Program

Motivation: Recent work has demonstrated a northern shift in bird habitats over the past 20 years in NY State. Our motivating question is to assess whether this northern shift is attributed to land use land cover changes and/or to climatic changes.

Methodology: Our research team is composed of engineers, ecologists and biologists. We will establish a biodiversity model linking land use land cover (LULC) alterations and climate changes to bird spatial distribution. Changes in LULC will be assessed directly or indirectly using remotely sensed information. In July 2009 we flew our own airborne mission using the Laser Vegetation Imaging Sensor, a waveform LiDAR sensor expected to be on board the 2017 DESDynI mission (budget permitting).

Findings: We have received the LiDAR data in April, 2011, so stay tuned.

Relevant papers: A few under development, more to come in late 2011/early 2012.

Collaborators: Dr. Giorgos Mountrakis, Jida Wang, Dr. Dimitrios Triantakonstantis

Urban growth model improvements

Urban growth model improvements

Funders: NSF, NASA, SUNY ESF Graduate Assistantship

Motivation: Urbanization is an important issue concerning scientists from different disciplines such as urban planners, biologists, engineers and is also extensively considered by policy-makers. A computational model quantifying locations and quantities of urban growth would provide environmental and socioeconomic benefits. Traditional urban growth models are based on a single-algorithm fitting procedure and thus restricted on their ability to capture spatial heterogeneity.

Methodology: A GIS-based modeling framework titled Multi-Network Urbanization (MuNU) model is developed that integrates multiple neural networks. The MuNU model enables a filtering approach where input data patterns are automatically reallocated into appropriate neural networks with targeted accuracies. We hypothesize that observations classified by each neural network share greater homogeneity and can be modeled more accurately with a collection of targeted algorithms.

Findings: Land use and land cover datasets of two time snapshots (1977 and 1997) covering the Denver Metropolitan Area are used for model training and validation. Compared to a single-step algorithm – either a stepwise logistic regression or a single neural network – several major improvements are pronounced in the visual output of the MuNU model. Statistical validations further quantify the superiority of the MuNU model given its effectiveness to incorporate spatial heterogeneity. Improvements range between 3% and 11% at the 1km scale when compared with a single neural network and logistic regression respectively.

Relevant papers: One published in IJGIS, another to be submitted in 2-3 weeks (as of April, 2011).

Collaborators: Dr. Giorgos Mountrakis

MSRBF Evaluation

MSRBF Evaluation

Funders: University of Maine covered patent application fees.

Motivation: One of the constraints of radial basis function (RBF) neural networks revolves around the ability to combine local functions of variable spreads. Development of a multi-scale RBF would allow RBFs to exhibit their full potential in regression and classification tasks from multiple disciplines.

Methodology: The underlying idea is to develop a training mechanism that considers both local and global errors when evaluating candidate node functions.

Findings: The initial experiments are very encouraging in terms of accuracy, consistency and repeatability.

Relevant papers: Patent 7,577,626 awarded by the USPTO – Full patent. A manuscript publication is under development.

Collaborators: Dr. Giorgos Mountrakis, Lori Luo, Dr. R. Watts

Integrating multiple classifiers

Integrating multiple classifiers

Funders: Multiple sources.

Motivation: Remote sensing as a field of study has reached its adulthood; computer-assisted classifiers have been in development for more than two decades. The complexity of remote sensing classification has led to a variety of methods, some of them based on artificial intelligence (AI). Recently, we have also observed a significant increase in parallel processing capabilities. Computer workstations with multiple processors are becoming the mainstream in research laboratories. Physical limitations in processor design indicate that future computational power improvements will result from parallel processing rather than single processor advances. Parallel computing presents a unique opportunity and challenge for image classifiers. The question arises: how can we harvest this new power to improve classification results?

Methodology: One solution is to implement hybrid classifiers, i.e. methods that merge multiple approaches together. When multiple classifiers are integrated correctly, the ability to harvest the power of multiple algorithms with a critical eye on the application at hand improves classification accuracy.

Findings: We have shown that hybrid classifiers perform similarly or better than single classifiers. We are currently focusing on intelligent selection from multiple classifiers and adjustments in the multi-step process.

Relevant papers:
L. Luo, G. Mountrakis (to appear). Integrating intermediate inputs from partially classified images within a hybrid classification framework: An impervious surface estimation example. Remote Sensing of Environment.

G. Mountrakis, R. Watts, L. Luo, J. Wang (2009). Developing Collaborative Classifiers using an Expert-based Model. Photogrammetric Engineering and Remote Sensing, 75(7):831-844.[pdf]

G. Mountrakis (2008). Next generation classifiers: Focusing on integration frameworks. Highlight article for Photogrammetric Engineering and Remote Sensing, 74(10):1178-1180.[pdf]

More papers in preparation.

Collaborators: Dr. Giorgos Mountrakis, Kari Gunson

Spatiotemporal MVCs

Moose-Vehicle Collisions in Space-Time

Funders: Humane Society, SUNY ESF Graduate Assistantship

Motivation: Moose-vehicle collisions (MVCs) pose a serious safety and environmental concern in many regions of Europe and North America. For example, in the state of Vermont one-third of all reported MVCs resulted in motorist injury or fatality while collisions have increased from two in 1982 to 164 in 2002. We studied varying spatiotemporal patterns of wildlife collisions along roads to assist transportation planners in identifying optimal mitigation strategies along specific roads, such as deciding on location and spatial length for permanent and expensive measures (e.g. crossing structures and associated fencing) versus less permanent and inexpensive structures (e.g. wildlife signage and reduced speed limits).

Methodology: Our work used a MVC dataset from 1983 to 1999 in the Northeastern Highlands of Vermont (four major roads) to perform space, time and spatiotemporal analyses and guide future mitigation strategies. An adapted kernel density estimator was implemented for exploratory analyses to detect high density collision hotspots on roads. We also developed an adapted Ripley’s K-function to test the hypothesis that MVCs clustering occurred at multiple scales in space, in time and in space-time combined.

Findings: The kernel in space showed seven major density peaks which varied in magnitude and spread between roads. The kernel estimator in time for all roads showed an exponentially increasing trend with annual periodicity and a seasonal cyclic component, where the majority of collisions occurred from May to October. Spatiotemporal kernel estimation exhibited discontinuous density hotspots in time and space suggesting changing animal movement patterns across roads. Statistically significant spatial clustering was evident on all roads at spatial scales from 2 to 10 kilometers. A more consistent clustering in time occurred on all roads at a scale distance of 5 years. Similar to the kernel estimation, annual periodicity was also evident. Positive space-time clustering was present at small spatial (5 km) and temporal scales (2 years) indicating that where MVCs occur is also influenced by when they occur. In retrospect, using multiple road lengths, and the combined kernel estimation and Ripley’s K-function in time and space, provided a powerful methodology to study varying spatiotemporal patterns of wildlife collisions along roads.

Relevant papers: G. Mountrakis, K. Gunson (2009). Multi-scale spatiotemporal analyses of moose-vehicle collisions: A case study in northern Vermont. International Journal of Geographical Information Science, 23(11):1389-1412.

K. Gunson, G. Mountrakis, L. Quackenbush (2011). Spatial wildlife-vehicle collision models: A review of current work and their application to transportation mitigation projects. Journal of Environmental Management, 92(4):1074-1082.