I'm working on a comparison of two different algorithms that detect weather features. The input data produced by each of the two detection methods is 3D arrays of binary (0/1) data with axes (lon,lat,time). For now, I'm just looking at comparing individual time slices with dimensions (lon,lat). The goal is to quantify the amount of similarity between the two methods in terms of which features they detect.
It was suggested that I use the chi-squared test to quantify the amount of similarity between the two methods-- so I'd be starting out with a 2x2 contingency table for methods 1 and 2 that's essentially
M1 M2 Pres 100 150 Abs 1000 950
(I'm just making those numbers up)
When I run
chisq.test, the p-value would tell me to either accept or reject the null hypothesis that these two variables are independent.
The other suggestion is the Jaccard index,
J = (A&B)/(A+B-(A&B))
Where the numerator is the intersection of the data and the denominator is the union of the data.
HOWEVER. The presence points are clustered together (because they're detecting one or two distinct features on a map), which means that I need to factor spatial autocorrelation into the presence counts, right?
I've found resources for testing for spatial autocorrelation, but not for how to adjust! Any guidance would be greatly appreciated.