Hybrid Machine Learning Methods for the Reconstruction and Disaggregation of Gridded Climatic and Hydrological Variables

Ujjwal Singh

Doctoral thesis

Abstract:

Ongoing climate change examining scenarios, accurate precipitation estimation and understanding runoff dynamics are vital in managing hydrology, meteorology, and water resources. While different in focus, these two domains contribute to an intricate understanding of water cycle components, particularly in regions experiencing water resource degradation. In both studies presented within this doctoral thesis, sophisticated machine learning techniques deliver in-depth insights. Specifically, this research delves into two distinct facets of hybrid modelling: the disaggregation of satellite-based precipitation data and the reconstruction of historical runoff reconstruction. The first study focuses on the methodological hybrid machine learning modelling framework for achieving accurate precipitation estimation at high spatial and temporal resolution. Downscaling or disaggregation is a key method for enhancing the spatial-temporal resolution of satellite data. The authors propose a non-parametric method involving a hybrid Extreme Gradient Boosting (XGBoost) approach combined with multivariate spatial-temporal Fuzzy clustering. This clustering relies on Integrated Multi-satellite Retrievals for GPM (IMERG) precipitation and Shuttle Radar Topography Mission (SRTM) Digital Elevation Data. Experimentation was conducted to downscale 255 months of IMERG satellite data over the Czech Republic. The utilization of eight stations for training and validation, and 19 additional stations for validation, showed a strong agreement with ground-observed precipitation. The proposed methodology substantially enhances the accuracy of IMERG precipitation data and holds promise for applications in other regions with remotely sensed data, particularly where ground-measured station data is sparse. Apart from precipitation, runoff plays a significant role in the water cycle. The second study emphasizes the reconstruction of historical runoff provides vital information for climate change adaptation. However, the conventional hydrological models can exhibit bias, leading to the necessity of machine learning-based computing models. A novel Hybrid Ensemble Multi-Model Framework (HEMMF) was proposed to reconstruct the gridded runoff of Europe over 500 years. The HEMMF combines non-parametric extended data pattern recognition and data-driven methods, employing nine different machine learning algorithms and four ensembles of ML. Using different datasets, including Moran's spatial autocorrelation (SPA) index and various climate-reconstructed data, the study found that the HEMMF offers the best reconstruction performance with a merged dataset. The reconstructed runoff, made possible by this advanced methodology, helps explain runoff trends, drought propagation, and relationships with climate variables. The proposed methodology has wide-ranging applications, including potential usage for past hydroclimatic data and related analyses across diverse temporal periods, climate scenarios, and geographical scales. Both methodologies play a crucial role in aligning and refining hydrological and meteorological observational records. Utilizing hybrid machine learning techniques for disaggregating precipitation and reconstructing historical runoff becomes particularly important when considering climate extremes. Notably, discerning trends on land becomes increasingly challenging in areas with limited data coverage. Addressing and bridging these data gaps is paramount. This ensures that such inconsistencies don't influence future projections, guiding our mitigation and adaptation strategies more effectively.

Keywords

Hybrid Machine Learning, Reconstruction, Disaggregation

Language used: English

Date on which the thesis was submitted / produced: 21. 2. 2024

Thesis defence

• Supervisor: doc. Ing. Petr Máca, Ph.D.
• Reader: Adam Vizina, externi, Martina Zeleňáková, externi, Golan Bel, externi

Czech University of Life Sciences Prague

Faculty of Environmental Sciences

Doctoral programme / field:
Landscape Engineering / Environmental Modelling