Abstract:This study investigates how conditional normalizing flows can be applied to remote sensing data products in climate science for spatio-temporal prediction. The method is chosen due to its desired properties such as exact likelihood computation, predictive uncertainty estimation and efficient inference and sampling which facilitates faster exploration of climate scenarios. Experimental findings reveal that the conditioned spatio-temporal flow surpasses both deterministic and stochastic baselines in prolonged rollout scenarios. It exhibits stable extrapolation beyond the training time horizon for extended rollout durations. These findings contribute valuable insights to the field of spatio-temporal modeling, with potential applications spanning diverse scientific disciplines.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: how to use conditional normalizing flows to perform spatio - temporal prediction on remote sensing data in climate science. Specifically, the research aims to explore the application of conditional normalizing flows in the prediction of climate variables, especially for the prediction of climate variables with different temporal and spatial resolutions. ### Main problems and challenges 1. **Computational complexity**: Global climate models (GCMs) usually describe physical processes based on partial differential equations. These models are computationally expensive due to coarse - grid discretization and solving in a large spatio - temporal domain. 2. **High - resolution requirement**: More accurate earth climate predictions are required. Accurate climate information needs to be provided not only on a global scale but also on a local scale to guide responses to extreme weather events and adaptation strategies (such as water resource management and agricultural practices). 3. **Stability of long - term prediction**: Existing deterministic and stochastic methods have the problem of performance degradation in long - term prediction, especially when it is beyond the training time range. ### Solutions The author proposes a new method - conditional normalizing flow (ST - Flow) to efficiently model the spatio - temporal characteristics of remote sensing data. This method has the following advantages: - **Accurate likelihood calculation**: It can calculate the exact likelihood value. - **Prediction uncertainty estimation**: It can estimate the uncertainty of prediction. - **Fast inference and sampling**: It improves the efficiency of inference and sampling, thus accelerating the exploration of climate scenarios. ### Experimental verification Through experiments, the author shows the performance of ST - Flow on the ERA5 dataset. Especially in long - term prediction, the performance of ST - Flow is better than that of deterministic and stochastic baseline models, and it shows stable extrapolation ability in the extended time range. ### Formula representation The core formulas of conditional normalizing flow are as follows: - The probability density transformation formula of normalizing flow: \[ p_X(x) = p_Z(f_\phi(x)) \left| \frac{\partial f_\phi(x)}{\partial x} \right| \] where \(\left| \frac{\partial f_\phi(x)}{\partial x} \right|\) is the absolute value of the Jacobian determinant of the transformation \(f_\phi\). - The likelihood function of conditional spatio - temporal normalizing flow: \[ p_X(x_t | x_{<t}) = p_\theta(z_t | h_t) \left| \det \frac{\partial z_t}{\partial x_t} \right| \] where \(h_t=\text{GatedConvLSTM}(x_{t - 1},x_{t - 2},\dots)\) represents the hidden state at time \(t\). ### Conclusion By introducing conditional normalizing flow, the author successfully solves the problems of instability and computational complexity in long - term prediction of existing methods, and provides a more stable and efficient solution for climate variable prediction.

Towards Climate Variable Prediction with Conditioned Spatio-Temporal Normalizing Flows

Climate Variable Downscaling with Conditional Normalizing Flows

Conditional Variable Flow Matching: Transforming Conditional Densities with Amortized Conditional Optimal Transport

Multivariate Probabilistic Time Series Forecasting via Conditioned Normalizing Flows

Conformalised Conditional Normalising Flows for Joint Prediction Regions in time series

Flow-based Spatio-Temporal Structured Prediction of Motion Dynamics

Dual-track Spatio-temporal Learning for Urban Flow Prediction with Adaptive Normalization

ClimAlign: Unsupervised statistical downscaling of climate variables via normalizing flows

Probabilistic Forecasting of Irregular Time Series via Conditional Flows

A Tale of Two Latent Flows: Learning Latent Space Normalizing Flow with Short-run Langevin Flow for Approximate Inference

Stochastic Flow Matching for Resolving Small-Scale Physics

Learning Likelihoods with Conditional Normalizing Flows

A probabilistic framework for learning non-intrusive corrections to long-time climate simulations from short-time training data

Conditional Local Convolution for Spatio-Temporal Meteorological Forecasting

Fine-Grained Urban Flow Prediction Via a Spatio-Temporal Super-Resolution Scheme

Continuous and Distribution-Free Probabilistic Wind Power Forecasting: A Conditional Normalizing Flow Approach

Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting

Flood Prediction Using Rainfall-Flow Pattern in Data-Sparse Watersheds

Temporal Interpolation of Geostationary Satellite Imagery with Task Specific Optical Flow

Flow Away your Differences: Conditional Normalizing Flows as an Improvement to Reweighting

Conditional Lagrangian Wasserstein Flow for Time Series Imputation