Lab 13 – Effects of Scale

Quantitatively, I compared the DEM elevation values directly, as well as two derivatives: Aspect and slope.  I first made both an Aspect and Slope surface for these DEMs.  I then used Batch processing for the Get Raster Properties tool selecting MINIMUM, MAXIMUM, and MEAN for each of the now six total rasters.

The elevation data comparison shows that the LiDAR DEM had a greater range of values compared to the SRTM DEM; the minimum value was smaller and maximum value was larger for the LiDAR DEM.  This may be an indication of greater elevation resolution.  The mean elevation value was lower for the LiDAR DEM.  Though these differences are apparent here, the magnitude of the difference is small and not likely significant. 

The slope and aspect summary statistics are highly similar between the two datasets as well.  The mean slope of the SRTM DEM is smaller, though this can be inferred based on the smaller range in elevations of this DEM.  It may be assumed that the LiDAR data is more accurate; however, the overall difference between the two datasets is very minimal.  I suspect the LiDAR dataset is more accurate for two reasons: First, it is the product of a resampling technique whereby the underlying accuracy of the high resolution 1-m DEM is certainly higher than the derive 90-m DEM.  Second, the SRTM DEM was created via orbital spacecraft, which, inherently introduces a higher degree of vertical measurement error.

Lab 12 – Geographically Weighted Regression

Spatial regression can be used in GIS to model a phenomenon of interest.  In non-spatial regression analysis, spatial auto-correlation is generally undesirable.  Spatial regression attempts to quantify auto-correlation and use it as an explanatory variable.  Geographic Weighted Regression (GWR) is a specific spatial regression used to account for multicolinearity.  In the lab this week we compared GWR to OLS regression.  The model output from GWR regression was better (i.e. had a lower AICc) than the OLS model.  Further the z-score was lower, indicating that there was spatial dependence in the phenomenon of interest.

Lab 11 – Multivariate Regression, Diagnostics and Regression in ArcGIS

Regression analysis can be used in ArcGIS to model a phenomenon which may vary spatially.  There are three main reasons for regression analysis: 1) to predict values in unknown or un-sampled areas, 2) to measure the influence of variables to a particular phenomenon, or 3) to test hypotheses about the influence of variables on a phenomenon.  It is very important to test the usability, performance, or predictive power of a model due to its potential in policy decisions.  There are several diagnostics to test the performance of a regression model that go past simply determining the R-squared value (which may be very misleading).  ArcMap contains an Ordinary Least Square regression tool, among others, which produces some of these diagnostics, however it is up to the user to evaluate these statistics in the context of the model.  The six step process outlined by ESRI to interpret these statistics provides a foundation for determining the "best" model.  One important thing to note is the frequency distribution of residuals in a histogram.  The residuals should be distributed standard normal.  A positive or negative skew may be the result of spatial auto-correlation.  This is perhaps the biggest use of regression analysis is GIS, as spatial analysis is central in ArcMap processing.

Module 10 Lab: Supervised Classification

Above is a supervised classification of Germantown, Maryland generated using
ERDAS Imagine supervised classification.  The map was ultimate composed in
ArcMap.

1.      I used the seed polygon method to generate the signature polygons.  A distance threshold value of approximately 25-50 was generally used.  In cases when only one class was created and the class was relatively obvious, for example water, then I used a higher distance threshold value.  In cases when there were many classes for the same classification, for example Agriculture 1-4, I used a lower threshold values.  The lower values and many classes allowed for high coverage with minimal (or none) mis-classification.

Module 9 Lab: Unsupervised Classification

A map depicting the unsupervised classification of University
of West Florida campus.  There are four different Information
Classes.