Meteorological drought assessment in northern Bangladesh: A machine learning-based approach considering remote sensing indices

Meteorological drought, driven by inadequate precipitation, has significant repercussions for water resources, agriculture, and human well-being. This study conducted an extensive assessment of meteorological drought in northern Bangladesh, employing remote sensing indices and machine learning techniques. The main aim was to evaluate meteorological drought occurrences in northern Bangladesh from 2010 to 2019, utilizing seven drought parameters and a machine learning model. Utilizing a Random Forest (RF) model, this study employed the Standardized Precipitation Index (SPI) as the dependent variable and seven remote sensing indices as independent variables. Through this methodology, the study assessed the significance of these indices generated by the model and integrated them, culminating in the creation of a meteorological drought distribution map spanning 2010 to 2019. This approach offers novel insights by probing the interplay and collective impacts of these indices, shedding light on previously unexplored aspects of regional drought patterns of northern Bangladesh. The major findings showed that precipitation strongly influenced both short-term and long-term meteorological drought episodes. Moreover, land surface-related indices, such as Evapotranspiration (ET) and Normalized Difference Water Index (NDWI), exhibited a more pronounced impact on short-term drought occurrences, while vegetation-related indices like Normalized Multi-band Drought Index (NMDI) and Normalized Difference Vegetation Index (NDVI) demonstrated greater influence over long-term drought events. During this timeframe, the Rajshahi division experienced frequent extreme and severe drought events. Moderate droughts and abnormally dry conditions were widespread. The Barind tract area consistently faced moderate to extreme droughts, with exceptions in 2011, 2014, and 2019. On average, over 5% of the region had extreme droughts, while more than 12% experienced severe droughts during this decade. Long-term drought indicators (SPI 6 and SPI 9) consistently showed higher frequencies of extreme and severe droughts compared to short-term indicators (SPI 1 and SPI 3), emphasizing the influence of prolonged rainfall deficits on extreme droughts and the relevance of longer time frames for severe drought dynamics. The RF model demonstrated strong performance with accuracy ranging from 81% to 95%. Low prediction errors (RMSE 6% to 31%) and high out-of-bag (OOB) accuracy ranging from 76% to 98% highlighted its accuracy. The F1 score consistently exceeded 76%, indicating high precision and recall. Cross-validation values ranged from 78% to 94%, affirming reliable generalization to new data. Incorporating the main findings, this study contributes valuable insights for the formulation of targeted drought mitigation strategies in northern Bangladesh. It is imperative to note that the scope of this study is confined to the northern region of Bangladesh, and generalizing these findings to other regions should be exercised with caution. Nevertheless, the research methodology and approach can serve as a model for future studies in related fields, advancing knowledge of how to assess droughts using remote sensing and machine learning methods.