Publication

Abstract : This study proposes a deep learning model for the integration of Landsat-8 and Sentinel satellite imagery to enhance the spatial, spectral, and temporal resolution of drought forecasting. Landsat-8 offers a superior spatial resolution but lower temporal frequency, whereas Sentinel imagery provides a higher temporal frequency but a reduced spatial resolution. The integration of these datasets mitigates the limitations inherent in each dataset, thereby improving the quality of Earth observation data and creating a more robust dataset for precise monitoring of droughts. Correlation analysis revealed that the fused dataset achieved stronger relationships between VCI and NDWI (r=0.72) and between VCI and SPEI (r=0.65) than the relationships achieved using the Sentinel-2 (r=0.68, r=0.62) and Landsat-8 (r=0.60, r=0.58) datasets. A bi-directional long short-term memory (Bi-LSTM) model was then applied to predict SPEI-12 from the vegetation and hydrological variables. The model achieved an R2 of 0.7946 and an root mean square error (RMSE) of 0.5709, indicating a strong agreement between the predicted and observed drought values. Compared with conventional LSTM and random forest models, the Bi-LSTM showed approximately 9–12% improvement in R2 and a 15–20% reduction in RMSE, demonstrating superior capability in capturing nonlinear vegetation–climate dependencies.