Computational Intelligence 

Abstract: Downscaling is a common task in climate science and meteorology in which the goal is to use coarse scale, spatio-temporal data to infer values at finer scales. Statistical downscaling aims to approximate this task using statistical patterns gleaned from an existing dataset of downscaled values, often obtained from observations or physical models. In this work, we investigate the application of domain alignment to the task of statistical downscaling. We present ClimAlign, a novel method for unsupervised, generative downscaling using adaptations of recent work in normalizing flows for variational inference. We evaluate the viability of our method using several different metrics on two datasets consisting of daily temperature and precipitation values gridded at low (1° latitude/longitude) and high ( and ) resolutions. We show that our method achieves comparable predictive performance to existing supervised statistical downscaling methods while simultaneously allowing for both conditional and unconditional sampling from the joint distribution over high and low resolution spatial fields. To the best of our knowledge, this is the first proposed method for unsupervised statistical downscaling, and one of very few proposed methods that allows for efficient sampling of synthetic data.

Abstract: Deep learning (DL) models are progressively being applied to climate applications due to their ability to learn complex nonlinear spatiotemporal patterns, typically present in the atmosphere. In particular, deep learning has landed on the downscaling field, providing high-resolution climate change projections crucial for sectorial applications. Despite their merits, they are still seen as “black boxes” generating distrust among the climate community and thus limiting their use in real applications. Therefore, there is a need to develop techniques that unravel the knowledge hidden in the neural models to 1) gain understanding about their decisions and 2) make these models more reliable to the community. In this study, we adopt a technique used in computer vision to visualize the decisions, to convolutional-based downscaling models. The results show comprehensive links learned by the network connecting the large-scale to the local-scale and prove the implicit feature selection that occurs within the hidden layers. To our knowledge, this is the first study that properly assesses a methodology to unravel the “black box”, in particular information concerning the predictor-predictand link, in a downscaling application.

Abstract: This work demonstrates that deep unsupervised learning holds much promise for rare event detection, even when labeled data is limited. Rare event detection is challenging for traditional supervised learning approaches due to high class-imbalance. We demonstrate the efficacy of deep unsupervised learning, in particular a variational autoencoder (VAE), when used for anomaly detection, in an application to avalanche detection in the French Alps, from satellite SAR imagery and a limited on-the-ground survey. Remarkably, our results demonstrate that supervision (i.e., access to labeled data) is only needed in the validation phase of our training pipeline, to tune a hyper-parameter: the threshold on reconstruction error of the VAE (trained in an unsupervised fashion) that will be used to designate an observation as an anomaly, i.e., an avalanche in this context. Our method outperforms previous benchmarks in avalanche detection, including supervised learning a convolutional neural network on an artificially balanced version of the same data. To the best of our knowledge, this is the first step in exploring the potential of applying deep unsupervised learning methods to detect avalanche deposits. Our proposed semi-supervised learning pipeline is also promising for real-world settings in which the data is very unbalanced and labels are limited.