| dc.description.abstract | The interplay between land use and extreme hydroclimatic events is a crucial aspect of Earth system science, influencing water resources, ecosystem stability, and disaster risk management. Extreme hydrologic events often arise from complex interactions among multiple environmental variables, manifesting as compound extremes where two or more climatic or hydrologic factors coincide, leading to amplified impacts. Examples include active and break spells in monsoonal rainfall, heatwaves, cold-wet spells, and flash floods triggered by snowmelt. The sporadic nature and multivariate dependencies of these events necessitate advanced methodologies for their accurate characterization, prediction, and risk assessment. Traditionally, hydrologic design has relied on univariate annual maxima approaches, but such methods are inadequate for capturing the joint behavior of multiple hydroclimatic variables and their temporal dependencies. This thesis develops novel statistical and deep learning frameworks to enhance the understanding, modeling, and forecasting of hydrologic extremes and land use dynamics.
A key contribution of this research, presented in Chapter 2, is a multivariate framework for analyzing compound extremes using a time-varying interval-censored copula estimation method which explicitly accounts for temporal dependence and tied values in hydrologic data. This approach enables improved estimation of design magnitudes and risk associated with extreme events. The methodology is demonstrated using daily precipitation and temperature data (1977–2020) from the Godavari River Basin, India, focusing on cold-wet compound extremes during the monsoon season. The results emphasize the critical role of ties and temporal dependence in hydrologic design, revealing their profound influence on risk estimates across various spatial scales. However, risk assessment alone is insufficient without predictive capabilities that enable proactive management of hydroclimatic and land parameter assessment.
To further enhance predictive capabilities, deep learning-based multivariate hydrologic time series forecasting are explored in Chapter 3. Traditional statistical approaches, such as Vector Autoregressive (VAR) models, often struggle with long-term dependencies and multi-step-ahead predictions, resulting in high errors. To overcome these limitations, a transformer-based model incorporating the Informer network is developed, leveraging self-attention mechanisms to capture long-range dependencies. The model is validated on the CAMEL dataset for streamflow forecasting, demonstrating superior performance. It is subsequently applied to forecasting temperature, soil moisture, and Normalized Difference Vegetation Index (NDVI) over the Godavari River Basin, achieving Kling-Gupta Efficiency (KGE) values exceeding 0.90 across various LULC types. Additionally, the model's performance is benchmarked against an encoder-decoder sequence-to-sequence LSTM, demonstrating its superior accuracy for multi-time forecasting. These results highlight the robustness of transformer networks in predicting land-atmosphere interactions over diverse landscapes. While these forecasts improve short-to-long-term predictions, the need for a specialized framework for extreme event forecasting and its interaction with land use remains.
This need is addressed in Chapter 4 of the thesis where EXtreFormer, a deep learning framework specifically designed for long-term extreme event forecasting is introduced. By integrating Rotational Position Encoding (RoPE) and Multi-Scale Temporal Attention Networks (mTAN), EXtreFormer is shown to effectively capture the interplay between land use, temperature, and NDVI. Applied to the Godavari River Basin, the model demonstrates strong predictive performance, with KGE scores reaching up to 0.86, while also providing enhanced interpretability through attention maps that highlight key predictive drivers. An ablation study confirms the model’s adaptability across different land cover types, establishing EXtreFormer as a versatile tool for extreme event forecasting.
By integrating multivariate copula-based risk assessment and deep learning-based forecasting, a comprehensive and interpretable framework for tackling hydrologic challenges are presented in this thesis. The findings contribute to better decision-making in water resource planning, disaster mitigation, and sustainable land management, offering practical insights for addressing the growing challenges of climate change and land-use transitions in data-scarce environments. Through this interconnected approach, the research bridges the gap between risk assessment, predictive modeling, and enabling a holistic understanding of land-extreme interactions. | en_US |
