Lunar Biomes
In a previous post, I used climate simulation software to model what the moon's climate would be if it had an Earth-like atmosphere (TL;DR: hot & wet).
I had hoped to repeat the process I used to assign biomes to Antarctica, training a random forest model on Earth's biomes and then applying it to the map of modeled temperature & precipitation data for the moon. (Essentially, this approach goes pixel by pixel through a map and asks: 'what part of Earth does this pixel most resemble?')
Unfortunately, that approach works best when the range of values is similar to Earth's, and that does not hold true for the simulated Lunar climate. The moon is only moderately warm, but it's extremely rainy for its temperature range:
So, we can't meaningfully map the Lunar climate to Terran biomes, but we can still try to slice the lunar surface into meaningful climate zones using a clustering algorithm. (I opted for k-means because it's easy).
Because the variables in question (annual temperature mean & sd, annual precipitation mean & sd, simulated vegetation cover, elevation, distance to water body, & log flow accumulation) have wildly different scales and are probably somewhat correlated, I scaled & centered them and then ran a principal components analysis. Five orthogonal components is enough to capture 93% of the variance in the 8 variables.
Importance of components:
Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Comp.6 Comp.7 Comp.8
Standard deviation 1.9659689 1.1590938 1.0256901 0.89120019 0.66697123 0.64757722 0.2843308 1.172976e-02
Proportion of Variance 0.4831293 0.1679373 0.1315051 0.09927974 0.05560634 0.05241954 0.0101055 1.719842e-05
Cumulative Proportion 0.4831293 0.6510666 0.7825717 0.88185142 0.93745776 0.98987730 0.9999828 1.000000e+00
We can take a look at how each component is calculated:
Loadings:
Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Comp.6 Comp.7 Comp.8
tas_mean 0.474 0.185 0.301 0.797
tas_sd -0.400 -0.142 0.806 -0.404
pr_mean 0.422 -0.458 0.125 -0.267 -0.712
pr_sd 0.417 -0.471 0.133 -0.273 0.702
vegplantc_mean -0.294 -0.554 -0.424 -0.395 -0.226 0.467
elevation -0.360 -0.287 0.300 0.197 0.811
wdist -0.216 -0.388 0.865 -0.209
flowacc -0.145 -0.938 0.149 0.274
Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Comp.6 Comp.7 Comp.8
SS loadings 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000
Proportion Var 0.125 0.125 0.125 0.125 0.125 0.125 0.125 0.125
Cumulative Var 0.125 0.250 0.375 0.500 0.625 0.750 0.875 1.000
Component 1, accounting for 48% of the variability, is higher when temp and precipitation are high, temp consistently so and precipitation variably so, with less (simulated) plant cover, low elevations, and proximity to the sea. Component 2, accounting for around 1/6th of the variability, is skewed toward dry lowlands. And component 3 seems to favor highlands.
Mapping the results, we get this:
Anyways, we can feed these into a clustering algorithm to try to identify groups of pixels with similar climate. I tried a few groupings and opted for four, as when I ran five clusters I had trouble distinguishing clusters 1 and 5. Here they are (light blue is ocean):
To characterize the clusters, I made some plots:
The clusters break neatly by elevation band which is a little disappointing but makes sense.
- Cluster 1: High elevations, dry, low but highly variable temperatures
- Cluster 2: Second lowest elevations, somewhat inland, moderate temperatures, relatively dry
- Cluster 3: Second highest elevations, dry, low but variable temperatures
- Cluster 4: Hot, wet coastal lowlands
So not that interesting, but scanning over the map there are enough deviations from purely elevation-based bands to still be somewhat compelling.
I'm imagining that the coastal regions in cluster 1 are basically rain-drenched mangrove swamps, with a gradient of forests in clusters 2 and 3 giving way to more arid montane forests (maybe 800-foot-tall lunar sequoias?) in cluster 4.
Here's the cluster map with shades of green assigned roughly corresponding to moisture level:
