Research Needs for Water Resource Applications of GIS

Water resources are challenging to model in a geographic information system because these features are often so temporally variable. John P. Wilson, et al. discuss in the article, “Water Resource Applications of Geographic Information Systems” many methods of water resource management and assessment. Often, these topics reiterate the relation of data sharing with the increasing level of technology. These advancements allow a greater number of less professionally trained individuals to contribute what can possibly be high quality data with a varying degree of uncertainty. For this reason, future technology will need to incorporate these individuals’ level of training and geographic and scientific understanding into the framework of tool building. Thus interfaces for using such tools must be created to suit a wide audience efficiently. Otherwise, difficulty in using technology directly translates to uncertainty. Scale is addressed often as well when reviewing other issues ranging from the integration of spatial data acquisition technology to data sharing practices.
Primarily, it is stressed that raster grid resolution selection should attempt to match the size of the smallest features that are being modeled.  Accuracy of these resources is quickly degraded due to excess aggregation of topographic features.  Generalization in this regard minimizes the number possibilities of variation for a given area where a smaller stream feature, for instance, could exist in reality.  Using a raster resolution of 30 meters rather than the 2-10 meter resolutions suggested by several studies mentioned easily smoothes over smaller changes in topography.  Further, a 30 meter cell size yields only 21% - 30% of correct slope gradients.  However, if only larger (major) features are required to be analyzed or communicated, 30 meter resolution can sometimes be quite sufficient.

Many water resource applications are GIS-based.  The National Hydrography Dataset is a prime example of how to digitally store and communicate a hydrologic network by logically displaying geographic information with its organizational and descriptive parameters embedded within in a GIS environment.  An example of temporal water quality monitoring is the Florida DEP's Watershed Monitoring program and EPA & FDEP’s STORET (“STOrage and RETrieval” database) program, where water quality data from across the state of Florida is collected, managed in databases, and can be served internally or publicly via internet mapping service, while being provided to in-house personnel for GIS analysis via ArcSDE & ArcGIS Server. Advancements in technology such as these increase the dissemination of information to a growing number of skilled as well as improperly trained stewards and contributors of information.  Increased availability must be mirrored by an increased accounting of data collection information in the form of metadata to ensure reliability.  A method of controlling uncertainty for every step of data collection is currently available by reporting appropriate metadata parameters prior to distribution of any data.  Although this information is mainly included on official final-drafts of published datasets, more can be done deeper within an organization to include metadata to aide in the quality assurance of final versions of disseminated data.

Overall, it will be important to develop efficient models of water resources that can be used by an audience with contrasting degrees of familiarity with geographic information.  By conducting further research on the most effective way to properly incorporate data with varying degrees of uncertainty, as well as investigating the most efficient scales at which modeling hydrologic networks should take place, the GIS community can continue to accurately model water resources in the future.

Remote Sensing for Landscape Level Dugong Habitat Evaluation

This is a review of an article presented to Dr. Yang's GIS for Environmental Modeling class at Florida State University in the spring of 2008.
Sheppard, James K., Lawler, Ivan R., Marsh, Helene. (2007). Seagrass as Pasture for Seacows: Landscape-Level Dugong Habitat Evaluation. Estuarine Coastal & Shelf Science, 71(1-2):117-132.

View Full Presentation

The dugong is a cousin to the West Indian manatee, which feeds on specific types of seagrasses in area around northern Australia, Indonesia, and along similar latitudes along coastal areas of the Indian Ocean.  This study uses GIS to determine which specific types of seagrasses are preferred by grazing dugongs in waters off of Queensland, Australia.  GIS is used to observe remotely sensed near infrared spectroscopy to quantify the composition and configuration of seagrass communities.

Dugongs are tagged with GPS tracking devices, and their locations are tracked using the Animal Movement Analyst Extendsion in ArcView.  After three months of tracking, a study area is developed from the many line features created by the movement of the animals to estimate the extent of popular grazing locations in Hervey Bay.  Several field methods are used to capture samples of seagrasses at various locations throughout the study area – ranging from intertidal treks in shallow water to sample plant species by hand, to using submarine video equipment to analyze the plant structure from a boat.  Both methods use GPS to derive latitude and longitude data and to record other information.  Further, a steel sediment grab is used to sample benthic composition and sediment nutrients.

The information is loaded into a GIS where bathymetry, seagrass cover, and nutrient profiles are created from the GIS point data that was captured.  Each of these layers are interpolated into a raster surface using the Kriging method, as it yields the most precise output for the data captured.  Nutrient analysis is conducted with Near Infrared Reflectance Spectroscopy (NRIS) due to its ability to identify the composition of organic samples very efficiently.

The analysis sampled five seagrasses, ranging in levels of nitrogen, starch, and fiber.  The dugongs convert nitrogen into protein, and use starch for energy.  Fiber is a tertiary element in the seagrasses that is difficult for the dugongs to digest, therefore it is mainly avoided in large amounts.  The analysis in this study confirms these findings from past research.

Remote sensing in this study is limited in evaluating biomass and species at low densities.  More valid results are easier to obtain when higher concentrations of any particular species exist over a larger area, though patches can be discovered and accounted for when sampling.  This aligns well with the general structure of seagrass meadows.  Further, remote sensing is limited to shallow water, or where seagrass stalks reach to shallow depths of water were the electromagnetic wave can penetrate and return to the capturing device.  Water quickly absorbs much of the beam emitted from sensing devices.  In addition, turbid water is another limiting factor to using remote sensing in open, flowing water.

Overall, this is a very interesting article that touches on several powerful uses of GIS and remote sensing technology.  The biological analysis addressed seems to get moderately complex, however the conclusions are clearly represented and are easily interpreted in context with the spatial component of this research.

An Essay on Spatial Interpolation

 Spatial interpolation is the technique of estimating a complete, continuous raster surface based off of a patchwork of known point values captured across a surface in reality.  The surface that is created attempts to represent the gradation, or smooth progression between varying spatial values found in real-world observations.  Interpolation is a powerful visualization method used to make sense of point data that is otherwise very difficult to envision and interpret.  Applications of such raster surfaces include maps of population densities, precipitation levels, temperature differences, etc.  This technique has nearly limitless uses and can create a surface for any spatially-based phenomenon – one of the most notable and easily interpreted of which is an interpolation of elevation values.  It is possible to use interpolation techniques in a vector environment using a triangulated irregular network (TIN), however much of my experience has remained in the world of raster based interpolation.

Spatial interpolation works by estimating values for the unknown, unsampled area between two or more known points.  A basic method of this is to derive isolines between sample points – this can even be done by hand.  I have performed a manual contour interpolation in a meteorology course in the past, where isopleths – lines of equal barometric pressure – were estimated from a map containing only weather station point data.  These lines were sketched, and another layer – frontal boundaries – could then be inferred from the location and orientation of the isolines.  Although this exercise included only a hand full of points, a surprisingly accurate map could be produced by hand, however the addition of more points can quickly force manual interpolation to become time consuming and increase chances of error.  Another method is to create a continuous surface – usually using advanced computer algorithms to automatically calculate the hundreds, if not millions or more raster cells that make up the interpolated surface.  My experience with these types of raster surfaces begin with a project conducted for the Florida State University Environmental Service Program in 2006, which used raster surfaces to model the efficiency of the University’s recycling program on campus, and to suggest new recycling bin locations.  Point locations were collected for centers of student population around campus – classroom buildings, libraries, parking lots, dorms, etc. – as well as for trash can and recycling bin locations.  Each point was assigned an appropriate weight value, and a raster surface was created for each layer.  The interpolations for each layer of centers of student population were combined using a raster calculation, and contour surfaces were created and overlaid with the raster surface of recycling bin distribution on campus.  It was then very easy to observe how the centers of high student traffic compared with the current distribution of recycling facilities on campus.

One final description of spatial interpolation is enforced by Tobler’s First Law of Geography where locations close to each other have more in common than locations which are farther apart.  Spatial autocorrelation, a formulation of Tobler’s Law, continues by measuring how clustered (positive spatial autocorrelation) or dispersed (negative autocorrelation) a set of spatial features exhibit.  These ideas illustrate the significance of neighboring points (or point sets) upon one another and how neighboring points effect the resulting interpolated surface.  More advanced interpolation techniques using these concepts will be discussed.

As with any area of scientific estimation and modeling, there are many issues that affect the accuracy of an interpolated raster or vector surface.  Basic affects stem from any combination of the number of points used for interpolation (the more, the better), and the distribution (positive or negative spatial autocorrelation) or distance between points (the location and proximity of neighbors significantly effect estimations between points).  More advanced effects begin with directional influence, and continue with the presence of barriers, local neighbor under- or overcorrection, and instrument error.  Finally, possibly one of the most important factors affecting the accuracy of an interpolated surface resides with choosing the most appropriate method to perform the surface interpolation.

Due to this great amount of effect that the method of spatial interpolation holds upon the accuracy of the resulting surface, it is important to carefully choose a method that is most appropriate for the nature of the given data.  The class notes discuss in detail the differences between several methods.  Generally, Inverse Distance Weighted (IDW) methods limit the influence of neighbors as their distance increases (implementing Tobler’s Law).  More sample points promote a more smooth surface, however areas of littler or no data will skew the surface toward the overall mean of the dataset, creating holes: an evenly spaced distribution of input points avoids these holes. Thiessen (Voronoi, or natural neighbor) polygons often have odd-shaped boundaries to transition between polygons.  Continuous variables are not well represented.  The trend surface method uses multiple regression (predicting a dependent elevation, Z variable with independent X & Y, location variables) to approximate values; however this method rarely intersects the original point.  Splines are used to interpolate smooth curves, and are best for surfaces that are already smooth.  Kriging is a more in-depth, random, weighted average technique using more advanced algorithms and spatial autocorrelation, best suited when correlated distances are known or if there is directional bias in the data.  Finally, it is important to understand issues associated with exact, and approximate interpolators, as well as deterministic and geostatistical interpolators.

The goal of interpolation is to minimize error.  It is the responsibility of the GIS user to understand the discussed differences between varying techniques used to interpolate surfaces.  Combining a working knowledge of the issues pertaining to spatial interpolation with the resources of the GIS reference information will assist in the creation of accurate representations.

Geostatistical Modeling of Surface Water Sediment Analytes

Presented April 27th, 2009 for Advanced Quantatitave Geography, Florida State University.

Introduction:

DEP Watershed Monitoring program holds a wealth of sampled water and sediment chemistry data. One particular analyte, Cadmium (Cd), is a toxic metal that is found naturally in soil, however it is also a component used in automobile brake pads. An exploratory investigation was attmempted, with the intentions of correlating lake and stream sediment-laden cadmium levels with population density. After numerous failed attempts of importing this dataset into R and GeoDa, this preliminary review of spatial autocorrelation is abandoned and alternative chosen. The new investigation will attempt to examine the spatial distribution of cadmium levels across the state of Florida.

Examination:

• Plot the raw data
• Plot the emperical variogram
• Develop a reasonable variogram model
• Use model together with the data to create a surface
• Examine model diagnostics

Default kriging interpolation:

[Basic kriging geostatistical interpolation]

[Same surface with station location distribution]

This simple surface does a moderatly decent job of modeling the phosphate belt in west central Florida, extending east of Hillsborough County, as well as similar areas of phosphate mines around the big bend.


Check for dataset normality:

The histogram shows significant skewness. Extreme outliers -- especially the single 17.2 mg/kg -- might need to be removed to observe a more appropriate distribution of analyte values across the dataset, though I'll leave it in for now

[Histogram shows significant skewness]

The QQPlot further suggests normality is highly suspect. Large portions of the dataset deviate from the mean; low values in particular.



A log transformation is applied to restore normality. Box-Cox transformation provides little variation to the original dataset.

Y(s) = log(Z(s)) for Z(s) ≥ 0

[Log transformation successful in minimizing distance and variation about the mean]

Examine spatial distribution of exceedance values:

[Phosphate belt modeled from adjacent large values]

Explore trend:

Very slight upside down U-shaped trend in east-west direction, and a slightly decreasing planar north-south trend. This is attributed to the unique shape of Florida (as opposed to a more positively correlated distribution possibility in Colorado for instance).

[Occurrences of values across Florida]

[Alternative view]

Examine spatial autocorrelation:

[Semivariogram of the cadmium dataset]

Semivariogram cloud:

• x-axis = lad distance (distance separating each pair)
• y-axis = difference² between the values of each pair of locations

- Low lag distance / high variogram value map imply inaccurate data
- Closer locations should be similar
- Distance between location pairs should increase with semivariogram values
- Cloud should flatten out after certain distance; indicates that a relationship between pairs of locations is no longer correlated
- This model experiences significant flattening and inaccuracy; could be alleviated with the removal of the 17.2 outlier

Remove Trend:

• Use a second order log transformation to remove trend in the northwest-to-southeast direction
• Use anisotropy to assist in removing directional components to spatial autocorrelation: the variogram increases more gradually and flattens out
  

[Values are attempted to be kept close to the mean]

Cross validation: Assess the model's performance

• Mean: Close to 0
• RMSE / Average Standard Error: as small as possible
• RMS Standardized Error: Close to 1

[Measuring model's performace]

The 17.2 outlier increases difficulty in judging lower values distribution about the mean, however the lower values are a decent fit.

Results:

Compare cross validation between the original, and fitted model. The map below is then clipped to the state of Florida.

[Cross validation comparison and final interpolated surface]

The fitted model is somewhat more accurate, however this dataset is modeled fairly well with default parameters. This is not the case for other datasets, however. This technique can significantly improve the interpolation of information across space by removing trend, fitting an appropriate model, and finally creating a more reasonable and accurate surface.

Although the levels of cadmium are largely localized, and variable across space, these methods successfully model two important areas of high concentrations of such minerals: the phosphate belt east of Hillsborough county, as well as another region of concentrated mining operations to the north - in the big bend region.