Application of Machine Learning Algorithms to Develop new Frameworks for Efficient Drought Assessment, Monitoring and Forecasting

Abstract

Drought is a complex and least understood natural hazard that occurs across the globe. It is also considered a creeping phenomenon with a much wider spatial, temporal, and areal extent than any other natural hazard. Drought is caused by natural climatic variability, leading to prolonged precipitation deficits in a region (months to years). It has widespread, devastating impacts on the global economy, ecosystem, agriculture, natural hebetate, and human life. Most nations around the globe are susceptible to drought. Climate change and global warming accentuate the importance of accurate drought assessment and management systems. Moreover, accurate assessment, monitoring, forecasting, and analysis of spatial patterns of drought using precise and innovative frameworks plays a pivotal role in developing adequate early warning and drought mitigation policies. From this perspective, this thesis presents five contributions that allow for reliable and faster seasonal and spatial assessment, monitoring, and forecasting of drought hazards. In the first contribution of the thesis, we have proposed the appropriateness of the Extreme Learning Machine (ELM) algorithm for drought forecasting to improve drought prediction accuracy. For quantitative assessment, time-series data of the Standardized Precipitation Temperature Index (SPTI) is used for nine meteorological stations located in diverse climatological zones of Pakistan. The prediction accuracy of ELM for SPTI with various months timescales was compared with Multilayer Perceptron (MLP) algorithm and Autoregressive Integrated Moving Average (ARIMA) models and analyzed using six performance evaluation metrics. Results indicate the superior performance of the ELM algorithm. The ELM algorithm can provide insights to freshwater resource managers and stakeholders to facilitate decision-making and the development of drought mitigation policies. We propose a Seasonally Blended Regionally Integrated Drought Index (SBRIDI) for regional drought characterization in the second contribution. We have developed a probabilistic framework based on the two-stage Bayesian Networks (BNs) theory. At first stage, BNs blends meteorological characteristics of various seasonally segregated Standardized Drought Indices (SDIs) at each of the individual meteorological stations in the region and produce seasonally relevant SDIs. In the second stage, BNs integrate seasonally prominent SDIs to produce spatially relevant SDIs for each season. The outcome of the framework named SBRIDI can efficiently assess and monitor drought hazards at the regional scale. The SBRIDI can highlight vi flashpoints of drought-prone zones, which could help stakeholders and relevant agencies manage at the regional level. In the third contribution, we propose a new Multiscalar Weighted Amalgamated Drought Index (MWADI) by developing a probabilistic framework based on BNs. We introduced a new generalized weighting scheme that integrates climatic information from multiple seasonal SDIs at individual observatories. The MWADI reduced the uncertainties associated with individual SDIs for accurate reporting of drought hazards. In the fourth contribution, we have examined seasonal and spatial drought frequencies, spatiotemporal patterns of drought, and inter seasonal drought persistence at various locations in northeastern Pakistan. Bayesian Logistic Regression Model (BLRM) describes the spatial patterns of drought occurrence and its persistence from the current season to the subsequent season. Results revealed that drought occurrence frequency and inter-seasonal drought persistence have certain spatial and seasonal patterns. The outcome associated with BLRM can provide a massive outline for the agricultural sector to manage drought-resilient crops in drought-prone regions. In the fifth contribution, we have examined the prevalence and persistence of different drought categories at selected meteorological stations. The Proportional Odds Model (POM) is used to calculate the odds ratios and probability of drought persistence for various climatic divisions (seasons). The associated outcome provides a basis for identifying spatial inter-seasonal propagation of meteorological drought.