Movatterモバイル変換

Modified Box and Whisker Plots

Whiskers extend from the 5th to 95thpercentiles, thebody represents the 25th through 75th percentiles, and the dot is the50th percentile.

Suggested usage:

• Gauge overlap between map units in terms of boxes (25th-75thpercentiles) and whiskers (5th-95th percentiles).
• Non-overlapping boxes are a strong indication that the centraltendencies (of select raster data) differ.
• Distribution shape is difficult to infer from box and whisker plots,remember to cross-reference with density plots below.

Density Plots

These plots are a smooth alternative (denistyestimation) to the classic “binned” (histogram) approachto visualizing distributions. Peaks correspond to values that are mostfrequent within a data set. Each data set (ID / variable) are rescaledto {0,1} so that the y-axis can be interpreted as the “relativeproportion of samples”. Note that density estimates are constrained tothe range defined by the 1–99 percentiles.

Suggested usage:

• Density plots depict a more detailed summary of distributionshape.
• When making comparisons, be sure to look for:
  • multiple peaks
  • narrow peaks vs. wide “mounds”
  • short vs. long “tails”

Monthly Summaries

Median PPT vs. PET, bounded by 25th and 75th percentile.

Modified box-whisker comparisons by month.

Monthly inter-quartile range.

Tabular Summaries

Table of selectpercentiles, byvariable. In these tables, headings like “Q5” can be interpreted as thethe “5th percentile”; 5% of the data are less than this value. The 50thpercentile (“Q50”) is the median.

Geomorphon Landform Classification

Proportion of samples within each map unit that correspond to 1 of 10possible landform positions, as generated viageomorphonalgorithm. Landform classification bythis methodis scale-invariant and is therefore not affected by computational windowsize selection.

Suggested usage:

• Use the graphical summary to identify patterns, then consult thetabular representation for specifics.
• “Flat” is based on a 3% slope threshold.
• Map units are organized (in the figure) according to the similarity,computed from proportions of each landform position.
• Thedendrogramon the right side of the figure describes relative similarity. “Lowerbranch height” (e.g. closer to the right-hand side of the figure)denotes more similar landform positions.
• Landform class labels and colors are aligned with an idealizedshedding →accumulating hydrologic gradient.

Landcover Summary

These values are from the2011 NLCD (30m)database.

Multivariate Summary

The following “ordination” summarizes environmental variables byMLRA. The flattening of multivariate data (16 dimensions) onto anoptimal 2D projection is performed usingprincipalcoordinates. Ellipses represent 50% probability contours viamultivariate homogeneity of group dispersions. MLRAdelineations with more than 1,000 samples are (sub-sampled viacLHS).MLRA with very low variation in environmental variables can result intightly clustered points in the ordination. Seethischapter, from the newStatistics for Soil Scientists NEDScourse, for an soils-specific introduction to these concepts.

Suggested usage:

• Colors match those used in the density plots above. Be sure tocross-reference this figure with density plots.
• The relative position of points and ellipses are meaningful;absolute position will vary each time the figure is generated.
• Look for “diffuse” vs. “concentrated” clusters: these suggestrelatively broadly vs. narrowly defined map unit concepts.
• Nesting of clusters (e.g. smaller cluster contained by largercluster) suggests superset/subset relationships.
• Overlap is proportional to similarity.

Pair-wise comparisons at the 90% level of confidence.

Raster Data Correlation

The following figure highlights shared information among raster datasources based onSpearman’sRanked Correlation coefficient. Branch height is associated with thedegree of shared information between raster data.

Suggested usage:

• Look for clustered sets of raster data: typically PRISM-derived andelevation data are closely correlated.
• Highly correlated raster data sources reduce the reliability of the“raster data importance” figure.

Raster Data Importance

The following figure ranks raster data sources in terms of howaccurately each can be used to discriminate between map unitconcepts.

Suggested usage:

• Map unit concepts are more consistently predicted (by supervisedclassification) using those raster data sources with relatively larger“Mean Decrease in Accuracy” values.
• Highly correlated raster data sources will “compete” for positionsin this figure. For example, ifelevation andmean annualair temperature are highly correlated, then their respective“importance” values are interchangeable.

Reportconfiguration and sourcecode are hosted on GitHub.