I have a time series of raster images for a study area that have been processed to a binary classification (inundated/not inundated). Each image in the series is time-stamped and can be associated with a specific water level in the study area that's expected to predict what cells are inundated.

I want to generate a model that allows me to interpolate/estimate inundation extent at any given water level. Traditional inundation models (e.g., "flooding" a DEM) are not an option because there is relatively little topographic relief in the study area and DEM data do not have sufficient vertical resolution.

Extensive searching for some kind of applicable modelling approach has proved unfruitful. Conceptually, I think of modeling each cell of the raster as a logistic model against gage height. This is clearly not reasonable, as it treats each cell as independent and ignores the certain spatial correlation in the data. It will also be quite unwieldy to fit hundreds of thousands of models... :)

I know I'm missing an obvious answer here. Can anyone send me towards an applicable method?

  • I am not sure if it could be the right approach but instead of doing a regular logistic regression you could do a Generalized Linear Mixed Model (GLMM) which can incorporate random intercepts and account for the spatial correlation. Furthermore, you could explore Random Forests which apparently are not affected by correlation. Check stats.stackexchange.com/questions/83628/… Commented Jun 5, 2019 at 16:15
  • The answer that you linked is not entirely correct. While it is not effected by parametric assumptions, high degrees of autocorrelation can bias the bootstrap and cause overfit. This is due to non-homogeneity in the bootstrap replicates causing correlation in the ensemble. Commented Jun 5, 2019 at 17:15
  • Thanks. I'm not sure about the RF approach. My initial thought was to try a spatial GAM (e.g., here) that modeled inundation as a function of an x-y smooth and its interaction with gage height, but I don't know if it would bog down under the dense data.
    – adamdsmith
    Commented Jun 5, 2019 at 22:58


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.