Spatial Analysis

The figure below shows the locations of the sensors in the Cub Hill deployment.

 

Cubhill-Location

 

We begin by looking at two months of soil temperature data (Jun 2009- Aug 2009) for the summer at depth 1. The heat map below shows the soil temperature values for different locations in the Cub Hill deployment. Note that not all locations are shown because some had a lot of missing values and faults. A total of 37 locations were used in this analysis. The tags on top indicate the type of location (F=forest, G=grass, E=edge of forest). The vectors are organized using the dendrogram clustering function in Matlab using Euclidean distance and average linkage. The numbers at the bottom indicate the distance between a location and its neighbor (to the right) in the color plot. The dendrogram for the entire site list is shown below too.

heatmap_D1.png

 

Dendro_D1.png

heatmap_D2.png

 

 

The relative heights in meters for the corresponding locations are shown below to see if there is any correlation between cluster locations and heights.

location_height.png

 

Next, we look at the raw time signals for the first 20 days between June 1, 2009 to Aug 1, 2009. Out of 50 locations, around 10 locations have a lot of missing values and noise so they and were removed from this analysis

st_signals.png

 

Correlation Functions for daily means

The heatmap and dendrogram for the mean values is shown below. Note that this plot shows the location mean minus the ensemble mean.

heatmap_mean.png

dendro_mean.png

All Data

Between Forest and Grass

Within Forest and Grass

cov_d1_all.png

cov_d1_between.png

cov_d1_within.png

var_d1_all.png

var_d1_between.png

var_d1_within.png

pairs_d1_all.png

pairs_d1_between.png

pairs_d1_within.png

 

 

Correlation function for the ramp value

Heat map and dendrogram for the daily ramp value

heatmap_ramp.png

dendro_ramp.png

ramp_cov_all.png

ramp_cov_grass.png

ramp_cov_forest.png

 

NOTES

-       The kriging library used is downloaded from http://mgstat.sourceforge.net/. Simple kriging was tried. For each timestep, for 37 locations and 1504 points. Identify the source of why it is so slow.

-       Try the clustering over winter months and the next summer?

-       Leave one-out cross validation

-       Evaluate locations where the relative error is the most.

-       Per cluster interpolation

-       Take more out and see how the error goes to see how many you can do without.

-       Hour of the day what happens.

-       Apply across moisture and see if we can define some sort of coverage concepts to minimize the number of nodes