Data-based local rainfall modeling through global climate information

Abstract

Climate is a global system whose subsystems interact complexly. Deterministic models are capable of describing the climate phenomena with physical detail around the globe. Nonetheless, the several concurrent global climate patterns make the numerical modeling challenging for tropical regions. This is because of inadequate parameterizations and systematic errors, typical of physics-based models. Additionally, the morphology of the mountains in the tropical Andes generates complex spatial patterns for the fluxes. A strategy to circumvent the climate modeling complexity in tropical mountain systems is based on the following considerations. 1) Although complex, the climate in the tropical Andes is strongly seasonal. 2) The climate is a network system in which global patterns greatly influence that seasonality. Both considerations seem to be rational criteria to devise a simplified but meaningful modeling process. This research thesis is about the modeling of local monthly rainfall signals using global climate patterns. It is assumed that global climate signals are crucial drivers for the local seasonal features. A signals’ decomposition using the well-known Dynamic-HarmonicRegression (DHR) helps determine which global climate signals influence the local climate. The DHR technique allows the rainfall to be separated into non-stationary trends and quasi-periodical signals. On the one hand, trends are used to find out inter-annual connections with global patterns. On the other hand, the non-stationary amplitudes of periodical components allow finding intra-annual connections. In a second stage, the identified global signals are included as exogenous variables in a harmonic model for simulating the monthly local rainfall. Global patterns determine the non-stationary properties of trends and periodical components through non-linear functions. The non-linearity is attained by the State-Dependent (SDP) technique, which infers non-parametrical functions between the harmonic’ parameters and global climate states. A preliminary evaluation reveals a model with abilities to accurately predict monthly rainfall signals, which points to potential fundamental climate mechanisms and conceptual links between the local seasonal behavior and global climate states. Finally, the model is data-driven in principle, synthesizing local seasonal features driven by global climate patterns, providing the model with a process-driven flavor. Because of this, the model’s evaluation requires a more comprehensive perspective, responding to both the data-driven and process-driven nature. In that sense, a predictability evaluation of the proposed model in contrast to other empirical alternatives is carried out. This predictive-based evaluation is typical for data-driven techniques. In addition, this work considers a process-oriented assessment based on the model’s capacity to mimic intrinsic seasonal and temporal characteristics. This reveals a model with better predictive accuracy than its alternatives in statistical terms and other attributes. It is argued that the predictability of the proposed model is attributed to its capacity for mimicking local rainfall features driven by global climate patterns.