COS 63-1 - Weatherclim: Towards a Method for Expert-Informed Dimensionality Reduction of Daily Weather Time Series

Abstract: One of the core successes of modern machine learning is that of feature discovery, often outperforming both expert and competing data-driven methods. In ecology, realizations of complex climate systems are often reduced into a set of 19 Bioclim variables, and are used in numerous applications, including as predictors in the majority of species distribution modeling studies. These expert-selected and interpretable variables, many of them correlated, are often further reduced by principal component analysis (PCA), selection of particular features according to knowledge of the system of interest, or the statistically-driven selection of a subset of uncorrelated features. Each step of dimensionality reduction discards a substantial amount of ecologically relevant data, in the process embedding conscious and unconscious assumptions on the dimensions of a species’ fundamental niche. Going to the source, we demonstrate the ability of common autoencoders to directly learn meaningful representations of three decades of daily weather time series for the state of Texas. However, the powerful encoders and decoders that allow for nonlinear compression can make it difficult to build interpretable scientific models on this lower-dimensional, or latent space. Such off-the-shelf approaches also provide no knobs for expert control of the process. In this proof-of-concept, we demonstrate the use of Bioclim as an expert in proposing the optimal structure for an interpretable set of latent variables that represent the environment. Specifically, the use of methods from continuous deep metric learning help to create an overarching structure in which coarse linear movements in latent space correspond to interpretable linear movements in the space of Bioclim features, while finer-scale variation in the latent space captures the fluctuations of weather around these climatic means. We contextualize this early, exploratory work as the start of a longer-term effort towards the derivation of datasets by expert-informed nonlinear dimensionality reduction for ecological modeling. In the context of species distribution modeling and ecological forecasting, the joint learning of climate and weather features, nicknamed Weatherclim, has the potential to serve as a practical methodological bridge for the consideration not just of climatic means but of yearly weather patterns and fluctuations as a definitional part of a species’ fundamental niche. In addition, these techniques can facilitate more responsive forecasting of species distributions under future climate/weather regimes.