An Adaptive Landscape Classification Model Utilizing Geoinformatics and Artificial Neural Networks
by André Coleman
Geographic Study Area
The Columbia River watershed boundary can be viewed using Google Earth via the following KML file available from my website:
crws.kml
Proposed Thesis Work
Environmental management and research within large heterogeneous watersheds provides challenges for consistent evaluation and
understanding of system processes. Many large river watersheds throughout the world exhibit highly diverse characteristics
of climate, hydrology, physiography, ecology, and anthropogenic influence within its boundaries. Assessing, mitigating, and
managing large systems is difficult largely due to varying natural characteristics, large geographic areas, domestic and
international political boundaries, and varying degrees of spatial, temporal, and empirical data availability and quality.
|
|
| These characteristics often allow only specific geographic areas and research/monitoring topics to be realized. Through
the development of data relationships and patterning, existing geographically specific studies and data can be used to
infer responses of other areas which have limited available data, but exhibit similar landscape characteristics.
An Adaptive Landscape Classification Model (ALCM) is introduced which utilizes Artificial Neural Networks applied to
observed and derived geospatial data. This research specifically evaluates landscape patterns to estimate characteristic
watershed responses to precipitation and land use for several unique and diverse areas within the Columbia River watershed.
The model provides an adaptive and evolutionary capability in which varying types of data can be incorporated to evaluate
different management needs such as water quality, aquatic habitat, groundwater recharge, and land use. This approach aims
to achieve a consistent, system-wide, and multi-scaled approach for providing data and information to environmental managers
and researchers.
The Columbia River watershed, located in the northwestern United States and southwestern Canada, is the fourth largest river system
in North America and exhibits highly diverse characteristics of climate, hydrology, physiography, and ecology within its boundaries.
The abundant existence of a variety of natural resources within the Columbia River watershed have brought anthropogenic induced changes
to the natural system causing physical and functional damage to terrestrial and aquatic habitats. An important point to consider is
that due to the wide geographic area of the Columbia River watershed, data sets of varying quality, classification, resolution,
and method will intentionally be included into the model to exercise and demonstrate its adaptability in utilizing different data
to provide a consistent determination of the research question. The process of providing data to this adaptive and evolutionary
computing process is conceivably infinite; as more data is provided to the machine learning process, there is a better probability
of providing a more realistic classification.
The overall objective of this study is to research the use of geospatial methodologies combined with Artificial Neural Networks
to develop an adaptive landscape classification model that can be used in a heterogeneous environment of data availability,
standards, quality, resolution, ecology, physiography, and climate. The model results have the potential to provide a consistent,
system-wide, multi-scaled, spatial database for use in adaptive environmental modeling, research, and management as well as
providing predictive capabilities for specific topic areas (i.e. determine probable locations of groundwater recharge zones,
ideal restoration/protection areas, field sampling sites, ideal land use, and various environmental impacts).
Data developed in this process can be used as input parameters for various physical and ecological modeling applications that can be applied
consistently across a watershed and potentially allows management of the basin as a whole instead of through the typical
compartmentalized approach. To apply the ALCM, the model will be trained with existing stream gage data for several watersheds
within the larger Columbia River watershed and will attempt to determine characteristic hydrograph response for all sub-basins in the domain.
The notion of the ALCM looks to utilize available and known spatial information on a fine-scale (first-order watershed) and apply these understood
characteristics and data relationships to other basins with less resolute or available data. The use of geospatial analysis methodologies and
Artificial Neural Network (ANN) algorithms is proposed as a method to discover landscape patterns and similarities between areas in the
Columbia River watershed that are not only spatially disjointed, but dissimilar in their available data. The use of ANN algorithms provides
a key and essential role in this research, as its purpose will be to understand and identify data relationships which are not always obvious
by other means. For example, a very simple data relationship can be established which identifies conditions for which a certain type of
vegetation exists. Utilizing field or remote-sensing observations, the vegetation of interest is delineated and included into our GIS. Using
all of our available spatial and non-spatial data such as elevation, slope, aspect, soil, temperature, and precipitation, an ANN can build
an understanding of what conditions support this vegetation. In the end, these conditions can be mapped back to a larger area which would
spatially represent areas which are likely to have this certain type of vegetation. There are many different types of ANN algorithms, part
of the thesis work will explore and validate these different approaches as it applies to landscape classifications. While the intended research
of using available stream gage data to distribute characteristic hydrograph response to unmeasured watersheds provides only one possible use
of ALCM, future possibilities of assembled data and application of the model will be determined. Some of these possibilities may include the
incorporation of data output from various physical models, "what-if" scenarios, snow melt forecasting, aquatic and terrestrial habitat determinations,
erosion and/or sediment transport probabilities, prediction of mass wasting sites, water quality classifications, groundwater recharge sites,
and measures of environmental quality.
|
|
