Spatio-temporal drought dynamics in Niamey District, Niger (2005-2022): multi-index deep learning approach leveraging...

In this research, we developed and applied a multi-sensor Long short-Term Memory (LSTM) pipeline for drought monitoring and analysis in the Niamey District of Niger from 2005 to 2022. Data inputs included Landsat 5 TM and Landsat 7 ETM + satellite scenes (resolution of 30 m, acquired between 2005 and 2013), Sentinel-1 C-band SAR and Sentinel-2 MSI imagery (resolution of 10 m, fused after 2014-2015), MODIS MOD13Q1 NDVI and MOD11A2 LST datasets (available throughout the study period), Climate Hazards Group InfraRed Precipitation with Stations satellite-derived precipitation data ("CHIRPS"; 0.05° resolution), and Climatic Research Unit ("CRU TS"; 0.5° monthly) surface climate data. Drought was quantified using SPI, SMI, PDSI, and VHI. Periodicity analysis detected a significant dry period from 2005 to 2012 and a slow but incomplete meteorological recovery trend from 2013 to 2022 during which SPI and PDSI remained relatively constant. Unexpectedly, given nearly two decades of hydrological drought conditions in the region after 2012, VHI exhibited a slight but significant positive trend (τ =  + 0.083, p = 0.0497), suggesting some overall improvement in vegetation condition throughout the region since 2005 which could be driven by factors such as CO₂ fertilization, land use/management change, or changes in vegetation community composition towards more drought-tolerant species, among other possibilities. Care should be taken when considering these results. The Long Short-Term Memory (LSTM) drought modelling and prediction pipeline had acceptable performance across metrics (R² = 0.93, NSE = 0.93, Bias =  - 0.714) with SPI having the highest modelling efficiency among drought indices (R² = 0.836). SMI and VHI proved to be less predictable likely due to the inherent coupling between soil and vegetation properties and non-linear relationship between root zone soil moisture and vegetation spectral indicators.