| 19 Jun 2026
Desert Model Intercomparison Project benchmark framework version 1.0 for assessing land-surface dynamics and surface memory in monthly dust aerosol optical depth over North Africa
Abstract. Deserts are the main global source of atmospheric mineral dust, yet large uncertainties remain in the simulation of dust variability across space and time. Part of this uncertainty may reflect the limited representation of dynamic land-surface states and antecedent surface conditions in current dust-model formulations. Here, we develop an interpretable machine-learning assessment framework, here termed the Desert Model Intercomparison Project (DesertMIP) climate–surface–memory benchmark framework version 1.0, to quantify the added value of land-surface dynamics and surface memory for monthly dust aerosol optical depth (AOD) over North Africa during 2003–2020. Three hierarchical configurations were tested: Climate-only, Climate+Surface, and Climate+Surface+Memory. Among the main explanatory models, Random Forest gave the best overall performance. Test-set skill increased from R² = 0.686 in the Climate-only case to 0.713 after adding surface variables and to 0.736 in the full Climate+Surface+Memory case, while RMSE declined from 0.058 to 0.056 and then to 0.053. The best model also gave MAE = 0.041 and Bias = 0.004. Residuals, defined here as observed minus predicted, were centered close to zero, with a mean of −0.004 and a standard deviation of 0.053, although residual spread increased at higher AOD values. Lag analysis showed a persistent memory-sensitive signal, with the strongest negative associations near a three-month lag for soil moisture (r ≈ −0.75) and vegetation density (LAI; r ≈ −0.82), whereas the highest mean cross-validation skill among the tested memory windows occurred at six months. Despite the gains in overall skill, severe dust outbreaks remained difficult to capture. For events above the 95th percentile, the full model gave POD = 0.423, CSI = 0.269, FAR = 0.574, extreme RMSE = 0.170, and extreme Bias = −0.060. These results support a reproducible pilot benchmark structure toward the DesertMIP.
There is a lot of nice work in this paper, but it not acceptable in a geophysical journal as it stands, as it does not tell us anything about the real world, and only tells us a little about the model world.
Please do not refer to this as a Model intercomparison project: that is usually used for evaluation and comparisons of land or earth system models, which is not what you are doing here.
I think you are using CAMS and ERA5 model output, which also have some assimilation. Please spend some time discussing the value of these model products for your analysis, and what biases they will have. A better approach would be to use real observations. It would be better if you used meteorological parameters that were used consistent with CAMS—maybe you did,I don’t remember (and I shouldn’t have to figure that out) but the model output you are comparing would ideally be consistent with each other. You could also look at model output from the MERRA assimilation system, similarly. But at some point, it would be much more helpful to compare to observations or link to observations, so that your analysis doesn’t just tell us what the model-world is doing, but the relationship between what the model-world is doing and the real world.
The study as it stands is really about evaluating different machine learning approaches to evaluate the CAMS-ERA5 dust/meteorological model output and look for relationships. That would be more useful for a machine learning journal, not an atmospheric journal. I recommend you do the analysis to see what is more robust, etc. and throw that information into the supplement. A paper submitted to GMD should focus more on what this tells us about the models you are evaluating. And why that could be helpful.
A way to make this more interesting to atmospheric scientists, is to contrast the models you chose here (CAMS-ERA5) with observations, and/or other models, for example from the CMIP6 or ACMIP database, and see how different models behave. And then maybe relate that to how these models compare against available observations. So that some models might have longer land surface memories, for example, and simulate the interannual variabilty better--that would provide interesting information to atmospheric scientists.