Desertification Forecasting in Semi-Arid Regions: Using Deep Learning to Probe the Observability Limits of Land Degradation Dynamics

Filipe Moura Da Silva; Cláubio Landney Lima Bandeira

eess.IVOct 9, 2026

Desertification Forecasting in Semi-Arid Regions: Using Deep Learning to Probe the Observability Limits of Land Degradation Dynamics

Authors: Filipe Moura Da Silva, Cláubio Landney Lima Bandeira

Organizations: EEEP Avelino Magalhães Tabuleiro do Norte, Cear´a, Brazil · Universidade Federal do Cear´a Russas, Cear´a, Brazil

Abstract

Forecasting desertification has received far less attention than mapping it. We use a convolutional LSTM (ConvLSTM) trained on ten annual Earth Observation (EO) variables (2000-2023) over the semi-arid state of Ceará, Brazil, on a 256x206 grid at roughly 2.2 km resolution, to measure how much of the year-to-year evolution of a MapBiomas-derived degradation index (IDV) can be predicted from satellite data. The network predicts a bounded change on top of the last observed state, is trained in two stages (single-step, then recursive rollout) and is evaluated on a contiguous spatial and temporal holdout. State-wide mean IDV rose from 0.238 to 0.350 between 2000 and 2023, about 0.005 per year, but cell-level change is dominated by year-to-year jumps, and a gradient-boosting model given all channels explains only 4.7% of its variance. Against this ceiling, the ConvLSTM lowers RMSE relative to persistence by 1.6% one year ahead and 4.4% four years ahead in a held-out southern band (p < 0.001), beats persistence with a constant drift and reproduces 53-69% of the regional drift. Drift-aware baselines only help from the third year onward, by up to 4.4% in RMSE at six years. An earlier formulation that redrew the whole field each year predicted changes unrelated to the observed ones (correlation -0.06) and produced a spurious decline when run recursively. Channel knockouts suggest that elevation and EVI degrade the forecast, while slope and aspect have no measurable effect. Monte Carlo dropout and stochastic scenario maps convey the uncertainty that the expected-value forecast hides. Under climatological forcing, mean IDV rises from 0.350 in 2023 to 0.370 in 2030. EO-based desertification forecasting is therefore reliable for the regional trend, not for individual cells.

Figures & tables

Fig. 1: Conceptual illustration of the decline in predictive performance as forecast lead time increases. Spatial prediction maps (top row) and temporal performance curves (bottom row) are shown for short-range ( $t+1$ ), medium-range ( $t+5$ ) and long-range ( $t+10$ ) forecasts. Short-horizon predictions reproduce the target patterns closely, medium-range forecasts drift as errors accumulate and observational constraints weaken, and at long lead times uncertainty dominates.

Variable	Source	Native res.	Aggregation
IDV (target)	MapBiomas	30 m	annual
Precipitation	CHIRPS	0.05°	annual sum
LST	MOD11A2	1 km	annual mean
NDVI	MOD13Q1	250 m	annual mean
EVI	MOD13Q1	250 m	annual mean
Elevation	SRTM	30 m	static

TABLE I: Variables in the Earth Observation Dataset. All Layers Are Resampled to the Analysis Grid of About 2.2 km. The IDV Is Both the Target and an Input Channel.

Fig. 2: The ConvLSTM cell. The input $X_{t}$ is processed through convolutional gates that update the memory $C_{t}$ and the hidden state $H_{t}$ . The forget gate $f_{t}$ , input gate $i_{t}$ and output gate $o_{t}$ regulate memory retention, the incorporation of new information and the exposure of memory to the output. In this study the memory spans the three-year input window.

$Y_{t}$ class	Share	Mean $\Delta Y$	Huber loc.	Mean $\Delta Y$	Unchanged
	(train)	(train)	(train)	(val)	(train)
0–0.01	0.265	$+0.0056$	$+0.0046$	$+0.0059$	0.90
0.01–0.2	0.260	$+0.0072$	$+0.0055$	$+0.0084$	0.52
0.2–0.5	0.228	$+0.0066$	$+0.0055$	$+0.0081$	0.33
0.5–0.8	0.146	$+0.0016$	$+0.0028$	$+0.0033$	0.28
0.8–0.99	0.078	$-0.0040$	$-0.0018$	$-0.0031$	0.39

TABLE II: Annual IDV Change by Current IDV Level, Valid Cells. Training Transitions 2000–2017 and Validation Transitions 2017–2023.

All valid cells
Baseline	$k=1$	$k=2$	$k=3$	$k=4$	$k=5$	$k=6$
Persistence	0.0513	0.0759	0.0954	0.1104	0.1282	0.1453
+ global drift	0.0514	0.0760	0.0952	0.1094	0.1259	0.1413
+ level-dependent drift	0.0514	0.0757	0.0946	0.1083	0.1242	0.1389
Per-cell linear trend	0.0536	0.0813	0.1042	0.1222	0.1412	0.1589
Southern validation band

TABLE III: RMSE of Simple Baselines Forecasting 2018–2023 from the Observed 2017 Map, by Horizon

k

in Years.

Model	RMSE	MAE	$R^{2}$	Pearson	SSIM
Persistence	0.0508	0.0250	0.975	0.988	0.974
CNN	0.0546	0.0375	0.971	0.986	0.958
LSTM	0.0626	0.0437	0.962	0.981	0.948
Residual ConvLSTM	0.0502	0.0274	0.975	0.988	0.971

TABLE IV: Single-Step (

k=1

) Performance on the Validation Band, Averaged over Validation Patches

ConvLSTM relative to
Model	$k=1$	$k=2$	$k=3$	$k=4$
Persistence	0.0514	0.0769	0.0999	0.1228
Persistence + drift	0.0512	0.0764	0.0988	0.1206
CNN	0.0546	0.0802	0.1037	0.1259
LSTM	0.0651	0.0838	0.1054	0.1273
Residual ConvLSTM	0.0509	0.0755	0.0969	0.1173

TABLE V: Multi-Step RMSE on the Southern Validation Band by Forecast Horizon

k

(Full-Scene Rollouts Started in 2018, 2019 and 2020). Bottom Rows: Relative RMSE Change of the ConvLSTM, with 95% Spatial Block-Bootstrap Intervals

Fig. 3: RMSE and $R^{2}$ by forecast horizon for persistence, CNN, LSTM and the residual ConvLSTM on the validation band, averaged over validation patches. Values differ slightly from the full-scene rollouts of Table V .

Configuration	$\Delta$ RMSE [95% CI]
Without SPI-12	$+0.00010$ [ $-0.00001$ , $+0.00026$ ]
Without NDVI	$-0.00007$ [ $-0.00017$ , $+0.00001$ ]
Without EVI	$-0.00045$ [ $-0.00085$ , $-0.00013$ ]
Without LST	$+0.00052$ [ $+0.00021$ , $+0.00091$ ]
Without precip.	$+0.00007$ [ $0.00000$ , $+0.00015$ ]
Without vegetation	$-0.00045$ [ $-0.00096$ , $-0.00004$ ]

TABLE VI: Channel Knockout on the Validation Rollout (

k=1

–

4

), No Retraining. Change in RMSE Relative to the Full Model (0.0855), with Paired 95% Bootstrap Intervals

Fig. 4: Change in validation RMSE when channels are replaced by a neutral constant, with paired 95% bootstrap intervals. Positive values mean the network relies on the channel. Bars in blue indicate channels whose removal lowers the RMSE.

Fig. 5: RMSE of the residual ConvLSTM on patches inside and outside each climate zone, target years 2018–2023. Error bars are 95% bootstrap intervals. The model is not retrained per zone.

Panel of Figure 12 — Fig. 6: Top: recursive projection of the IDV for 2024–2030 under climatological forcing, cells outside Ceará held fixed. Bottom: mean IDV over valid cells, observed (2000–2023) and projected (2024–2030).

Fig. 7: Hindcast 2018–2023. Share of valid cells with IDV $\leq 0.01$ (left), with IDV $\geq 0.99$ (centre) and unchanged from the previous year (right), for the observed maps, the expected ConvLSTM forecast and the mean of 40 scenarios.

Fig. 8: Observed IDV in 2023 (left) and expected change from 2023 to 2030 in the ConvLSTM projection (right).

Fig. 9: From left: observed IDV in 2023, expected ConvLSTM projection for 2030, two stochastic scenarios for 2030, and the fraction of 40 scenarios in which a cell’s IDV rises by more than 0.1 between 2023 and 2030.

Fig. 10: State-wide mean of the MC Dropout projection, 2024–2030. The band is the average over valid cells of the cell-level 95% intervals.

Estimator	2024	2025	2026	2027	2028	2029	2030
MC Dropout ( $D=50$ )	.0008	.0012	.0014	.0017	.0018	.0019	.0021

TABLE VII: Mean MC Dropout Standard Deviation over Valid Cells, by Forecast Year

Spatiotemporal Forecasting

Spatiotemporal forecasting predicts how quantities change across locations and time, from traffic and weather to disease and environmental systems.

See all 429 papers

A Sequence-to-Sequence ConvLSTM Approach for Leaf Area Index Forecasting over the South-Central United States
Aug 1, 2026Zhixing Ruan, Lixin LuTime Series Forecasting Spatiotemporal Forecasting
Cooperative Institute for Research in the Atmosphere, Colorado State University Fort Collins, Colorado, USA
EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting
Jun 25, 2026Junwei Luo, Shuai Yuan, Zhenya Yang +3World Model Learning Forecasting Benchmarks
1The University of Hong Kong · 2Wuhan University
State-Space Unlearning for Non-Stationary Bias in Land Surface Forecasting
Sep 30, 2026Anidipta PalMachine Unlearning State Space Models
Heritage Institute of Technology, Kolkata, India

Aug 1, 2026J/K move · Enter open · S save

A Sequence-to-Sequence ConvLSTM Approach for Leaf Area Index Forecasting over the South-Central United States

Zhixing Ruan, Lixin Lu

Cooperative Institute for Research in the Atmosphere, Colorado State University Fort Collins, Colorado, USA

Time Series Forecasting Spatiotemporal Forecasting

Open paper AskPDF arXiv

Leaf Area Index (LAI) is a fundamental biophysical variable governing land-atmosphere interactions; however, LAI forecasting at high spatial resolution remains an unsolved challenge. While recent machine learning approaches have demonstrated LAI estimation at point or regional scales, none provides a gridded, meteorology-driven prognostic forecast suitable for subseasonal land surface and climate modeling applications. Here we present a sequence-to-sequence Convolutional LSTM (ConvLSTM) framework that generates daily 1-km LAI forecasts up to 30 days ahead, driven by historical LAI sequences and daily meteorological forcing including temperature and precipitation. Trained and evaluated over the South-Central United States -- a region of strong climate gradients and diverse vegetation -- the model achieves a domain-averaged RMSE of 0.36 at a 30-day lead time, more than a third lower than the persistence baseline. Forecast skill remains robust across seasons, geographic distributions, and plant functional types, including forests, grasslands, shrublands, and croplands. To our knowledge, this is the first demonstration of skillful LAI forecasting at a 30-day horizon at 1-km resolution.

Abstract

Spatiotemporal Forecasting

Explore similar work

A Sequence-to-Sequence ConvLSTM Approach for Leaf Area Index Forecasting over the South-Central United States

EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting

State-Space Unlearning for Non-Stationary Bias in Land Surface Forecasting