| 16 Sep 2026
Projecting global changes in land use and ecosystem services using SEALS (Spatial Economic Allocation Landscape Simulator) version 2.0
Abstract. Land-use change alters ecosystems and the services they provide, but calculating these effects requires high-resolution land-use maps. Ecosystem service models rely on satellite-derived maps for present and historical conditions, using spatial resolutions of 10 to 300 meters. Scenarios of future land-use dynamics, however, are typically provided at resolutions of 0.25 arc-degrees (~30 km at the equator) or coarser, far too coarse for ecosystem-service models. Here, we describe and evaluate the Spatial Economic Allocation Landscape Simulator (SEALS) model, a global, computationally tractable model that downscales coarse land-use projections into high-resolution maps suitable for ecosystem service analysis. SEALS estimates 3.12 million parameters applied to spatial layers for land constraints, transition suitability and multi-scale adjacency convolutions, identified separately for each 1-degree tile globally. The parameters are trained on observed land-use change from 2000 to 2015 and validated on withheld data from 2020. SEALS can generate global 10 arc-second land-use maps in minutes on a standard laptop. We show that SEALS places change 1.9 times more accurately than a random allocation of the same change and that the advantage persists when near-misses are credited. Passing the downscaled maps to the InVEST habitat-quality model reproduces habitat quality calculated from observed land cover more closely than coarse accounting in nine of ten deforestation regions, showing that resolving where change falls improves a downstream ecosystem-service estimate. As a demonstration we downscale eight Shared Socioeconomic Pathway scenarios to 2100 and release the resulting global 10 arc-second land-use maps.
The manuscript presents SEALS 2.0 a model for downscaling of land use predictions. While the underlying model looks promising and appears to be sound there are some issues with the presented manuscript. I will first state my general observations and then present my observations of specific parts in the manuscript.
General observations
- The flow of the paper seems to be off at certain points. The main issues are:
- In Section 2 “Model description” I am missing a general overview of the flow which ties the different sections together and provides me with a reference of how each part contributes. Sections 2.2 to 2.4 present different components of the calibration suitability without it ever being properly introduced. Figure 1 tries of providing such information, but the text does not properly reflect it.
- Section 3.4 and 3.5 provide experimental setups, which are then used (with the theory reiterated) in Section 4, the rest of section 3 talks about methodology and implementation details. Also, what is the reason for Section 3.1 “Loss function” being part of Section 3 instead of Section 2.
- In the evaluation section, why is no comparison made to other downscaling models? In both Section 1 and Section 5.3 a lot of other models are presented, and I would like to see how SEALS compares against them and not only a random placement. I think that this would draw a much better picture on where SEALS is located in the field of available methods. There are some other issues with the evaluation section which I discuss below.
- The authors present quite a lot of supplementary material, which in itself is no problem, but one might want to think about if all of it is needed and if not some of it should be moved to the manuscript instead.
Major observations
- Line 177, P_{ijk} was never properly introduced. This is also related to line 171 where it is mentioned that sections 2.2 to 2.4 present the three components of the transition suitability without properly explaining what these components are. Line 149 mentions P_{ijk} but does not explain it. The first proper mention of it is on line 249.
- Both Sections 2.2 and 2.3 could benefit of some examples of what constrains and spatial variables can be.
- Section 3.2 “Calibration” is quite short and does not go into much detail on how the actual calibration is done. In point (d) the authors write “iteratively updates the parameters based on the value of the loss function, […]” how does this iterative update look like? What is the governing formula?
- Section 4.1 aims to provide prove that the held out validation skill is the same as for non held out data. But the authors never make any comparison between the prediction of the year 2020 and the one of the year 2015, instead they compare the model’s performance of the year 2020 to the performance of random placement in the year 2020. Here a proper comparison between the models prediction in held out data and non held out data is needed. Also a justification on why comparing a 15 year prediction to a 5 year prediction is valid is needed.
- Section 4.3, the authors state that “We do not analyse the scenarios further […]”, but I think that this is required here. Not from an environmental or sociological stand point of view, but from a model performance one. The thing that immediately caught my eye is the strong artifacting that occurs, what impact does this have on subsequent research? How do other models perform here? I think this section should be expanded to discuss these things.
- Section 4.4 “Computational performance”, there you state that “[...] scales to high-performance computing with thousands of CPUs [...]”, additionally line 83 states that SEALS has “[...] highly optimized, [and] parallel computation [...]”. I do not think that the content of Section 4.4 is enough to justify these claims. Firstly, does SEALS support any distributed computing? Because if not then it is going to be difficult to scale it more then one node which means that it will never be able to reach the few thousand CPUs. Secondly, from the current analysis one cannot really see any scaling. Comparing two different systems against one another and then trying to make statements about scaling is basically impossible. What should be done is to take System B and then run the program multiple times with different numbers of CPUs and then based on this make the scaling study.
- Section 6 “Conclusion” gives a quick summary of the different parts shown the manuscript but does not draw a conclusion from this.
Minor observations
- Line 99 to 102, the three sentences are confusing, the authors first state that Section 5.2 is describing the “larger modelling pipeline”, then they state that the manuscript itself evaluates the fine-downscaling step, and in the next sentence they state that the manuscript “[…] describes and evaluates the SEALS model.”
- Line 114, the paragraph seems out of place. It talks about why high-resolution land use change data is important, but at this point the model is already presented. This paragraph would fit better in the beginning of the introduction.
- Section 2.6 could be a bit more detailed for the assignment rule.
- Line 339, What is meant by pixel? Is it the same as a fine cell?
- Line 346 to 349, Is there some prior research that supports the claim that random placement is an appropriate null?
- Line 389, the paragraph talks about how polygonal data instead of gridded one can also be used, this is never explored any further. The paragraph seems somewhat out of place.
- Line 429, the given statement is never proven.
- Line 484, how are the deforestation regions chosen?
- Figure 5, why is there no plot for the random fine prediction? For the bar plot it might be useful to use a log scale for the y-axis as it is currently hard to read the values for the SEALS prediction.
- Line 506, a justification is need on why having shown the correctness of the models 5 year prediction capabilities, it also creates correct predictions for 85 years.
- Line 513, what meaning do the percent changes carry? Isn’t this already given by the coarse changes?
- Line 534, the statement of higher clock speed and smaller core count is too general. There are a lot more things than just clock speed and core count that play into the performance of a system.
- Table 1, Why is there no data for the “peak resident memory” in System B?
- Section 5.2, this section seems somewhat out of place.
- Section 5.3, a lot of talk about different models is done and how they compare against SEALS without showing the actual performance differences.
- Section 5.4, what are some of the limitations that we observe in the model output? E.g., artifacting.
Typos
- Line 374, \sigma_m not in math mode.
I believe that if these points, especially the ones concerning the model evaluation, are adequately addressed the manuscript would be ready for publication.