A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Remote Sensing and Geographic Information Systems

Thesis (M.Eng.) - Asian Institute of Technology, 2025

Drought is a recurring phenomenon that significantly impacts the living conditions of local populations and destabilizes the surrounding environment in Northeastern region, Thailand. This situation poses a considerable challenge to the region's socioeconomic stability, necessitating comprehensive and proactive solutions to address the complexities arising from dry conditions. This research aims to overcome and prepare solutions for the difficulties associated with the dry condition by employing the machine learning technique of Long Short-Term Memory (LSTM) with varied fixed window lengths for predicting drought risk in the area under study. The primary meteorological (PCI and TCI), agricultural (VCI), and hydrological (WCI and ETCI) drought indices, alongside socioeconomic vulnerability factors, collected over a 22 year period from 2002 to 2024, serve as crucial data for developing predictive drought risk models. The research findings indicate a high probability of drought localization in the central to lower sub-region of northeastern Thailand, encompassing Nakhon Ratchasima, Buri Ram, Surin, Si Sa Ket Khon Kaen, and Maha Sarakham. Notably, different drought levels, ranging from mild, moderate, severe, and extreme, have occurred over the years. For prediction purposes, the study utilized two types of LSTM models: To-One and To-Many, to forecast various environmental factors and detect drought in the subsequent one and several months, respectively. To assess the models' performance, statistical measures such as the correlation coefficient (r), coefficient of determination (R2), and root mean squared error (RMSE) were calculated. The To-One LSTM model demonstrated high accuracy in providing 1-month drought forecasts, achieving strong correlations with actual data across multiple scenarios, establishing it as a reliable model for drought prediction. Specifically, when comparing suitable window lengths for drought indices generation, the highest correlation in R2 of 0.926 with suitable window lengths of nine months was identified for the drought index of TCI, whereas the lowest coefficient of determination in the ETCI of 0.140 was found during the one-month fixed window length. For other indices, the suitable range of window length expand from 1-month to 6-month period with R2 of 0.434 to 0.688. For the To-Many LSTM model, the performance exhibited less accuracy, showing more noise and visual artifacts. In examining drought risk performance by combining different drought type indices, the research offers a comprehensive image of dryness conditions by observing and analyzing based on various categories. Overall, by utilizing machine learning for constructing drought risk assessment with satellite drought indices, the results present high potential in drought monitoring and forecasting, particularly when integrated with in-situ information. In future study, this approach can be a notable prediction method for better understanding drought assessment.