FORECASTING PRECIPITATION IN CUBA USING Graph based deep learning

Author

• Taufiqul IslamFollow

Graduation Semester and Year

Spring 2026

Language

English

Document Type

Thesis

Degree Name

Master of Science in Earth and Environmental Science

Department

Earth and Environmental Sciences

First Advisor

Ghanbarian, Behzad

Abstract

Daily precipitation forecasting remained a challenging problem in regions characterized by strong spatial heterogeneity, nonlinear atmospheric dynamics, and intermittent rainfall behavior. Cuba represented a particularly complex case due to the combined influence of tropical cyclones, easterly waves, mesoscale convective systems, and orographic effects, which produced highly variable rainfall patterns in both space and time. Conventional statistical and machine-learning models typically treated stations independently and therefore overlooked spatial dependencies that strongly influenced rainfall variability across the island.

This study developed a spatiotemporal deep-learning framework for daily precipitation forecasting across 40 spatial nodes in Cuba using data from 1979 to 2023. The proposed approach was based on a graph convolutional recurrent neural network (GCRNN) that explicitly modeled spatial relationships through a physically informed graph constructed using training-period data only. A strict forward-chaining temporal split was adopted (training: 1979–2015, validation: 2016–2019, testing: 2020–2023), along with spatial generalization in which 20% of nodes were withheld during training to evaluate performance at unseen locations.

To ensure a leakage-free comparison, three model configurations were evaluated: a masked long short-term memory (LSTM) serving as a temporal baseline without spatial information, a partially masked LSTM incorporating limited spatial features, and the graph-based GCRNN that performed spatial message passing across nodes while restricting target-node information during training. This framework allowed systematic assessment of the role of spatial dependency in precipitation forecasting.

Results showed that the GCRNN consistently outperformed both LSTM-based models across all evaluation stages. During the validation period (2016–2019) at the 32 training nodes, the GCRNN achieved RMSE values ranging from approximately 0.12 to 1.83 mm day⁻¹ and MAE from 0.03 to 0.31 mm day⁻¹, with R² (coefficient of determination) values between 0.97 and 0.99. During the testing period (2020–2023) at the eight held-out nodes, the GCRNN achieved RMSE values ranging from approximately 0.39 to 1.69 mm day⁻¹ and MAE from 0.15 to 0.42 mm day⁻¹, with R² values between 0.97 and 0.99. In contrast, during the same testing stage, the partially masked LSTM showed substantially larger errors, with RMSE ranging from approximately 1.19 to 5.08 mm day⁻¹, MAE from 0.27 to 1.28 mm day⁻¹, and R² ranging from approximately 0.83 to 0.98. The masked LSTM baseline performed substantially worse than the partially masked LSTM and failed to generalize to held-out test nodes, producing negative R² values and significantly larger errors compared to the GCRNN.

Overall, the results demonstrated that incorporating spatial structure through graph-based learning significantly improved precipitation forecasting, especially under the challenging node-generalization setting. The proposed GCRNN framework provided a robust and transferable approach for modeling complex spatiotemporal rainfall dynamics and offered practical value for hydrological forecasting, water resource management, and climate-resilience applications in tropical regions.

Keywords

Precipitation forecasting, Graph neural network, GCRNN, LSTM, Deep learning, Spatiotemporal modeling, Cuba precipitation, Climate resilience, Hydrological forecasting, Graph-based learning

Disciplines

Climate | Environmental Monitoring | Hydrology | Meteorology | Other Environmental Sciences | Water Resource Management

License

This work is licensed under a Creative Commons Attribution 4.0 International License.

Recommended Citation

Islam, Taufiqul, "FORECASTING PRECIPITATION IN CUBA USING Graph based deep learning" (2026). Earth & Environmental Sciences Theses. 3.
https://mavmatrix.uta.edu/ees_theses2/3