|I thank the authors for taking the previous review and subsequent revisions so seriously. The revised manuscript is much improved. I have a few more ideas for improvements.|
Most importantly, section 3.3 must be edited to be much clearer whether the model is extrapolating through time or interpolating across space. In order to extrapolate through time, the training and testing data must be separated through time, meaning the training data must be from a former time point or interval while the testing data must be from a subsequent time point or interval. I think the authors’ model is not designed primarily to extrapolate through time. The authors’ model seems to be designed to interpolate across space, meaning interpolating among the training sites. The manuscript must be clearer exactly which map(s) from which time(s) are used for the training and testing data, and how the samples are selected.
Also the manuscript should be edited to be clearer whether it models either the suitability for a particular category or the suitability for transition to a particular category. For example, figure 4 seems to show high suitability for Built at places that are already Built, which implies that the suitability means suitable for Built, not suitable for transition to Built. If a pixel is already Built, then its suitability to be in a Built state in the future is high, but its suitability to transition to Built is zero. The manuscript must be clearer concerning the distinction between suitability to be in a particular category versus suitability to transition to that category.
Please add to the Acknowledgements “The Massachusetts Geographic Information System and the United States’ National Science Foundation supported the compilation of the Plum Island Ecosystems data through grant OEC-1238212.”
Lines 254-255 of the revised manuscript reports “Parametric models, such as logistic regression, assume the error terms of the input data to be independent and identically distributed (Overmars et al., 2003; Wu et al., 2009).” I am not certain whether that is correct. I think the model assumption is that error terms of the fitted regression surface are independent and identically distributed. I am not certain whether Overmars and Wu got it correct, or whether the authors cite them properly. Even I am not 100% certain whether the authors are wrong, but the authors should double check whether the errors relate to the independent variables or to the fitted relationship.
Lines 368-369 say “either the total number of cells occupied by each category or the number of transitions among the various categories at each time point”. That phrase should be “either the total number of cells occupied by each category at each time point or the number of transitions among the various categories during each time interval”, because occupation occurs at a time point and transitions occur during an interval.
The abstract can cut “In this paper”. The abstract should change from “applications usually only perform the spatial allocation” to “applications frequently perform only the spatial allocation”. The word “only” should be as near as possible to the thing it modifies, which is the spatial allocation, not perform. Also, it is not clear that most models perform only the spatial allocation.
Line 243 should change from “the most common” to “a common”.
The last word of the caption for figure 7 should change from “decreases” to “coarsens”.
I hope my feedback helps. I congratulate the authors on a fine piece of scholarship.