.png){kind=icon}
.png){kind=icon}
Return
GeoAI-Based multi-model framework for spatiotemporal drought modeling in arid environments
.png){kind=icon}
PRE
AI
A
DOI：10.1080/01431161.2026.2689555.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate drought prediction in arid regions is complicated by strong climate variability and frequent regime shifts. This study develops and evaluates two complementary modeling frameworks, XGBoost and ConvLSTM, to model the Standardized Precipitation Evapotranspiration Index (SPEI-3) across the Al-Kharj-Riyadh Hydrological Basin (KRHB) using 14 hydro-climatic variables from remote sensing and reanalysis datasets. A strict chronological train–validation–test split (2000–2024) was implemented to avoid temporal leakage and ensure a robust assessment. XGBoost exhibited strong temporal predictive skills, achieving training R2=0.99, validation R2=0.72 and test R2=0.65 with an RMSE of 0.23. SHAP analysis confirmed that SPEI3 lag 1 was the dominant predictor, with an importance of 0.52, followed by precipitation (0.21) and temperature-related variables. However, spatial transferability varied substantially, with performance ranging from R2=0.73 at well-behaved locations to negative R2 in highly heterogeneous grid cells. This sensitivity highlights XGBoost’s limited ability to capture micro-scale spatial variability. Moreover, ConvLSTM demonstrated stronger spatial coherence and higher KGE-based hydrological fidelity, whereas XGBoost achieved higher test-set R2 and lower RMSE. On the independent test period, ConvLSTM achieved a Kling-Gupta Efficiency (KGE) of 0.71, R2=0.58 and RMSE = 0.42, indicating its ability to reproduce the main temporal and spatial structure of drought and wetting events, although some amplitude smoothing remained during rapid transitions. Crucially, the ConvLSTM remained stable during the 2023 pluvial regime shift, successfully tracking anomalies exceeding SPEI = +1.2, whereas the XGBoost exhibited mild amplitude damping. The comparative analysis shows that XGBoost provides strong, interpretable, autoregressive performance, whereas the ConvLSTM delivers higher robustness to non-stationarity and higher reconstruction of spatial drought dynamics. These results suggest that deep spatiotemporal models can be a useful modeling option for drought monitoring and decision-support applications in arid environments such as the Arabian Peninsula.
Keywords:
Drought
ConvLSTM
XGBoost
GeoAI
climate
SPEI
Arid Regions
Remote Sensing
Journal
IF:
2.6
Papers:
1.2W
Citations:
2.7W
Organization
Cited Papers