Trained on 24 years of satellite data, a ConvLSTM model outperforms simple persistence by up to four years and predicts that soil degradation will continue between now and 2030.
Desertification is well mapped, but rarely anticipated. Researchers have trained a ConvLSTM network on ten annual Earth observation variables (2000-2023) covering the semi-arid state of Ceará, Brazil, on a grid of about 2.2 km of resolution. Objective: To measure how much of the evolution of a MapBiomas derived degradation index can be predicted by satellite.
A predictable trend, unpredictable cells
The average state index has increased from 0.238 to 0.350 in 23 years, but at the scale of each cell the variations are dominated by year-to-year jumps: a gradient boosting model accounts for only 4.7%. In an area away from training, the ConvLSTM reduces the error from 1.6% to one year and 4.4% to four years, and reproduces 53-69% of the regional drift. A first version, which redesigned the entire map each year, produced changes unrelated to reality.
Under climatological forcing, the average index would reach 0.370 by 2030. The authors conclude that satellite forecasting is reliable for the regional trend, not individual plots, and accompany their uncertainty estimates maps.
Read the full article on the original source
Go to the source →