University of Twente Student Theses
Comparative analysis of rainfall surface generation using deterministic, stochastic and deep learning approaches
Saha, Swaraj (2024) Comparative analysis of rainfall surface generation using deterministic, stochastic and deep learning approaches.
|
PDF
3MB |
| Abstract: | Rainfall surface generation plays a critical role in climate science, providing essential insights into the spatial distribution of precipitation. Accurate rainfall data are pivotal for disaster management, water resource planning, and understanding regional climate dynamics, particularly in regions with complex topography like the Himalayas. Despite advances in spatial interpolation techniques, generating high-resolution rainfall surfaces that accurately reflect regional variability remains a challenge. Traditional deterministic and stochastic methods often struggle with capturing fine-scale variations, especially in areas with sparse observational networks. The advent of deep learning, however, offers new opportunities to enhance the precision and efficiency of rainfall surface generation. This thesis addresses the problem of generating high-resolution, daily rainfall surfaces for the state of Uttarakhand, India, a region characterized by challenging terrain and significant risk of natural disasters. The research identifies a gap in existing methodologies, which often fail to integrate the computational power of deep learning with traditional spatial interpolation techniques. The objective of this study is to develop a novel spatial interpolation framework that leverages deep learning to generate precise, daily gridded rainfall data and to compare its performance against traditional methods. The methodology involves a comparative analysis of deterministic, stochastic, and deep learning approaches to rainfall surface generation. The study uses daily rainfall observations from automatic weather stations (AWS) in Uttarakhand, employing geostatistical interpolation methods such as Kriging and Regression Kriging. The deep learning model, based on the Graph Neural Network (GNN) framework and Gaussian Mixture Model (GMM) convolutions, is trained to predict rainfall surfaces from these observations. The model's performance is evaluated using Root Mean Square Error (RMSE) and compared with traditional methods and existing gridded data provided by the Indian Meteorological Department (IMD). Results indicate that while traditional Kriging methods capture general trends, they often fail in areas with sparse data, leading to inaccurate predictions in unobserved regions. The deep learning model, on the other hand, shows promising results, particularly in capturing complex spatial patterns and providing moderate to high accuracy acrossthe study area. The study also revealsthat elevation, while statistically significant, has a limited impact on rainfall predictions in this context, suggesting the need to incorporate additional climatic variables for improved accuracy. The implications of this research are significant for regional planning and disaster management in Uttarakhand. The enhanced rainfall surfaces can improve early warning systems for floods and landslides, contributing to better risk management and spatial planning. The deep learning model, once further refined and trained on larger datasets, could be deployed in real-time applications, providing continuous updates on rainfall patterns. Future work will focus on expanding the study area, increasing the number of training samples, and incorporating additional environmental variables to improve model accuracy. Moreover, integrating satellite-based rainfall products with ground observations could offer a more comprehensive approach to rainfall prediction on regional and global scales. This research demonstrates the potential of deep learning as a powerful tool in climate science, capable of overcoming the limitations of traditional interpolation methods and providing accurate, high-resolution rainfall data critical for environmental management. |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Subject: | 38 earth sciences, 50 technical science in general, 54 computer science |
| Programme: | Geoinformation Science and Earth Observation MSc (75014) |
| Link to this item: | https://purl.utwente.nl/essays/103602 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page