I have 218 benthic sediment grabs from the continental shelf ranging from 20 to 130 meters deep. These samples were taken from eight sites spread from Northern California to Southern Washington. Within each site, samples were randomized along depth gradients.
Each sample consists of species counts plus depth, latitude, and sediment characteristics such as grain size (i.e., sand versus silt), organic carbon, and nitrogen concentrations. Using Bayesian Belief Networks, species– habitat associations were calculated and established relationships were used to make regional predictive maps. Final map products depict the spatial distribution of suitable habitat where a high probability indicates a high likelihood of finding a species given a location and its combination of environmental factors. While sampling points were not taken across a consistent gridded scale, suitability maps were scaled to 250 meter resolution.
As in any habitat modeling process, the “best” model was chosen by looking at model performance and the amount of error, or misclassification between what was observed and what was predicted. Error of commission occurred when probability scores were high for a location where the species was actually observed to be absent. Errors of omission occurred when probability scores were low for a location where the species was actually observed to be present.
I am interested in two questions. The first is whether there is a spatial pattern to the error observed and the second question is at what scale is this error significant? Error may be caused by variation in the environment that occurs at a finer scale than what my modeling structure captures.
To explore these two questions, I intend to conduct a spatial autocorrelation analysis on the error for each local site to determine if there is any potential spatial pattern, and if so, if there is an associated environmental pattern to the error (i.e., does most of the errors occur in shallower or deeper water?). I am also interested in creating high resolution local maps of sediment characteristics (grain size, organic carbon, and nitrogen) through spatial interpolation technique using sediment grab data. For these local sites, I will then recreate predictive maps and compare to the 250 meter predictive maps.
First I have a couple of questions.
1. What is keeping you from creating regional maps at a finer scale that 250m? I’m guessing it is the scale of one of your inputs?
2. Are you incorporating other habitat characteristics such as slope of shelf in your model? ( I can’t see the inputs into your net from the picture). It looks like most of your error points are on the deeper end of the shelf but there may be some local variation in slope that could be contributing to the error.
I agree that creating fine scale local maps is going to do a lot to reduce the error that you are seeing. Alternately using the MGET (marine geospatial ecology tools) package to create another map (including more fine scale habitat variables) and compare your outputs might be another way to figure out the source of error.
Looks like an interesting project. My first questions is: what do you mean, exactly, by “at what scale is this error significant?”?
Secondly, I would suggest conducting a univariate analysis, using your predictor variables, to see the likelihood of a false positive (i.e. error of omission) associated with that variable. You could conduct the same analysis for false negative (i.e. error of commission). This analysis may point to a variable in the analysis that could be giving rise to these observed error in prediction. If you have questions regarding how this may be done, please feel free to talk to me in class about it. 🙂