Desertification forecasts hold for the regional trend, not for individual 2.2 km cells: a 24-year Ceará record puts a number on the ceiling

Desertification forecasting is split into two scales: the regional mean trend can be projected, while next-year change at an individual 2.2 km cell largely cannot. Earth Observation work has gone mostly into mapping and classification, leaving forecasting of the degradation process itself comparatively underexplored and without a model-independent reference for what is predictable. State-wide mean degradation index rose from 0.238 in 2000 to 0.350 in 2023, about 0.005 per year, but the standard deviation of annual cell change is 0.053, more than ten times the mean drift, and the median change is zero in every degradation class. A gradient-boosting model given all ten channels explained only 4.7% of next-year change variance over 2018–2023, an RMSE 2.9% below persistence.

The model no longer redraws the whole degradation field each year; it learns only the change relative to the last observed state. An earlier version with a sigmoid output had to reconstruct the entire field every year, and its predicted changes were uncorrelated with observed ones, producing a spurious decline when run recursively. The residual ConvLSTM starts from persistence and bounds its output to about 0.05 in normalized change; on the held-out southern band it lowers one-year RMSE by 1.6% (95% interval 1.1–2.0%) and four-year RMSE by 4.4% against persistence, reproducing 53–69% of the regional drift, with a predicted-to-observed change correlation of about 0.13 against 0.22 for the gradient-boosting reference.

Under climatological forcing, the state-wide mean degradation index rises from 0.350 in 2023 to 0.370 in 2030, about half the recent rate. The observed rise over 2018–2023 was faster than the 2000–2023 average, and the model reproduced only 53–69% of the regional drift in hindcast. Monte Carlo dropout gives a 2030 95% band of 0.366–0.374, entirely above the 2023 value; stochastic scenario maps keep 54% of cells unchanged each year in hindcast, close to the observed 48–55%, against 11–15% for the expected-value forecast.

The analysis covers Ceará, Brazil, using ten Earth Observation variables from 2000 to 2023 at roughly 2.2 km spatial and annual temporal resolution. The predictability ceiling is conditional on these ten variables, this resolution and this period, and the authors note that richer inputs could raise it. The regional trend and its uncertainty can be projected from the record, while per-cell maps mainly carry the 2023 pattern forward and should not be read as predictions of where degradation will appear next. The authors propose repeating the analysis in other drylands to separate region-specific dynamics from limits that come from partial observability in general.

The record spans only 24 years, limiting training samples and the statistical power of every comparison; the validation set is 18 patch-windows from a single band in the far south, making metrics sensitive to that region's 2018–2023 trajectory. The zone-stratified evaluation does not retrain with zones withheld, and the channel knockout does not retrain without the removed channels, so both describe one trained network. The 2024–2030 projection uses climatological forcing, excludes future droughts, El Niño events or land-use change, and its Monte Carlo dropout bands reflect model spread rather than the cell-level variability with a standard deviation of about 0.05 per year. The scenarios assume the jump statistics of 2000–2023 continue unchanged.