geomid2.py

from sklearn.neighbors import KDTree

def nearest_neighbor(df1, df2):
    kdt = KDTree(df1.as_matrix())
    distances, indexes = kdt.query(df2.as_matrix(), k=1, return_distance=True)
    return pd.Series(distances.flatten(), index=df1.index.values[indexes.flatten()])


import sys
import pandas as pd
import numpy as np
args = sys.argv[1:]

df1 = pd.read_csv(args[0], index_col="GEOM_ID")
df1["area"] = df1.area.apply(np.sqrt)

df2 = pd.read_csv(args[1], index_col="GEOM_ID")
df2["area"] = df2.area.apply(np.sqrt)

s = nearest_neighbor(df1, df2).order()
print s.describe()
print s.tail()

matches (distant).csv

This is another try at creating geom ids, but with a different strategy. Here I'm just trying to do "matches" of a list of geometries to another list of geometries. I do this by creating a dataframe which contains centroidx, centroidy, length, area, and the four corners of the bounding box and put them into a dataframe. One dataframe for the "haystack" and one for the "needles."

In my test, I'm pretty sure all the needles are present in the haystack, but several operations have been done that manage to apply some "noise." For my example (of the 2M parcels in the Bay Area), the describe of the distances to the nearest neighbor looks like this:

count    1951744.000000
mean           0.000098
std            0.000056
min            0.000004
25%            0.000065
50%            0.000085
75%            0.000117
max            0.024922

I added a csv of the 2 rows which are the two geometries which were furthest apart (that .0249222 number) - they still look pretty close.