A preprint posted on arXiv on 9 October by researchers at Brazil’s Federal University of Ceará examined how far desertification indicators could be forecast in the semi-arid state of Ceará. The team combined ten annual observation variables for 2000–2023 on a grid with cells of roughly 2.2 kilometres and trained a convolutional LSTM (ConvLSTM). Its target was an index derived from MapBiomas land-use data that reflects the presence and intensity of degradation-prone cover. The historical mean index rose from 0.238 in 2000 to 0.350 in 2023, while annual changes in individual cells were mostly volatile jumps.

In a held-out southern validation band, the model reduced error relative to a persistence forecast by 1.6% one year ahead and 4.4% four years ahead; the paper reports that it did not beat simple methods for a one-year forecast at the state-wide scale. The researchers conclude that satellite indicators can capture a regional average trend, but cannot reliably identify which individual cells will degrade next. A scenario from 2023 to 2030 assumes average climatic forcing and is not an observed future outcome. The findings are limited to this dataset and model experiments, and the work is a preprint that has not yet undergone peer review.