When analyzing data it is important to have a basic familiarity with the data structure. With tabular data this often means creating histograms and scatter plots to visualize the structure and relationship between point values. Also useful are knowing descriptive statistics such as minimum, maximum, mean, and standard deviation values. Familiarity with spatial data should include measures of their geographic dispersion, autocorrelation, and value aggregation. Within ArcGIS these characteristics can be measured using “Average Nearest Neighbor”, “Spatial Autocorrelation (Global Moran’s I)”, and “Hot Spot Analysis Getis-Ord Gi*)” tools, respectively. In this example I look at the spatial structure of a sample of satellite image-mapped forest disturbances in Oregon’s west Cascades. The data are polygons representing unique disturbance events, with attributes including: year of disturbance detection, magnitude of disturbance, and duration.
1. Average nearest neighbor.
Magnitude of disturbance was divided into three classes (low, medium, and high). Each class was run through the average nearest neighbor tool to determine if the spatial pattern is clustered, random, or dispersed. The pattern for low magnitude disturbance is random, whereas medium and high are clustered. This pattern of disturbance severity and its distribution is possibly a function of the disturbance agent. Low magnitude disturbances are typically natural, which may be more random than anthropogenic disturbances, like clearcuts, which dominate the medium and high magnitude classes. Note that nearest neighbor analysis is highly sensitive to the data extent. A larger of smaller extent, would likely change the result, therefore the stated results are only meaningful for the area and extent used, not an indication of universal pattern.
2. Spatial autocorrelation (Global Moran’s I)
Global Moran’s I was applied to disturbance magnitude (without classification based on severity). Global Moran’s I indicated that the disturbances are clustered by magnitude. This means that there is autocorrelation within data, where disturbances close to one another have similar magnitudes. The results are the same as nearest neighbor evaluated by severity classes, except that magnitude was explicit in the analysis with Global Moran’s I (no classification needed). The interpretation is the same as that for nearest neighbor.
3. Hot spot analysis tool (Getis-Ord Gi*)
Getis- Ord Gi* calculates a z-score that relates to the clustering of either high or low valued features. The results, based on the entire range of magnitudes, shows significant clustering of high values, but not of low values, which is consistent with nearest neighbor analysis. The areas showing greatest significance of high magnitude clustering have relatively large gaps between neighbors, which could be a consequence of the “look-to-distance” of the analysis.
It might be interesting to describe the spatial distribution using standard deviational ellipses… sometimes the shape of the study area is so strong that you don’t see any kind of variations, but it might be worth a try anyway. Try creating standard deviational ellipses for low, medium, and high magnitude events and see if the ellipses are different in terms of size, orientation, or center. (If each polygon is labeled “low”, “medium” or “high”, use this field as the Case field to create all the ellipses at once). I would select 1 standard deviation for size in order to see core areas. Create an ellipse of the events (no case; no weight) to serve as a baseline. Then weight by time (create an integer date like Julian dates), weight by duration, etc. If, for example the ellipse weighted by duration shifts north from the baseline it would indicate the longer duration events are in the north.
Hope this analysis provides interesting results!
Best wishes,
Lauren