Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-Modal Learning-based Reconstruction of High-Resolution Spatial Wind Speed Fields

Matteo Zambra, Nicolas Farrugia, Dorian Cazau, Alexandre Gensse, Ronan Fablet

TL;DR

This work proposes a scheme based on both data assimilation and deep learning concepts to process spatiotemporally heterogeneous input sources to reconstruct high-resolution time series of spatial wind speed fields and shows that the proposed framework outperforms a deep learning–based inversion scheme and can successfully exploit the spatiotemporal complementary information of the different input sources.

Abstract

Wind speed at sea surface is a key quantity for a variety of scientific applications and human activities. Due to the non-linearity of the phenomenon, a complete description of such variable is made infeasible on both the small scale and large spatial extents. Methods relying on Data Assimilation techniques, despite being the state-of-the-art for Numerical Weather Prediction, can not provide the reconstructions with a spatial resolution that can compete with satellite imagery. In this work we propose a framework based on Variational Data Assimilation and Deep Learning concepts. This framework is applied to recover rich-in-time, high-resolution information on sea surface wind speed. We design our experiments using synthetic wind data and different sampling schemes for high-resolution and low-resolution versions of original data to emulate the real-world scenario of spatio-temporally heterogeneous observations. Extensive numerical experiments are performed to assess systematically the impact of low and high-resolution wind fields and in-situ observations on the model reconstruction performance. We show that in-situ observations with richer temporal resolution represent an added value in terms of the model reconstruction performance. We show how a multi-modal approach, that explicitly informs the model about the heterogeneity of the available observations, can improve the reconstruction task by exploiting the complementary information in spatial and local point-wise data. To conclude, we propose an analysis to test the robustness of the chosen framework against phase delay and amplitude biases in low-resolution data and against interruptions of in-situ observations supply at evaluation time

Multi-Modal Learning-based Reconstruction of High-Resolution Spatial Wind Speed Fields

TL;DR

This work proposes a scheme based on both data assimilation and deep learning concepts to process spatiotemporally heterogeneous input sources to reconstruct high-resolution time series of spatial wind speed fields and shows that the proposed framework outperforms a deep learning–based inversion scheme and can successfully exploit the spatiotemporal complementary information of the different input sources.

Abstract

Wind speed at sea surface is a key quantity for a variety of scientific applications and human activities. Due to the non-linearity of the phenomenon, a complete description of such variable is made infeasible on both the small scale and large spatial extents. Methods relying on Data Assimilation techniques, despite being the state-of-the-art for Numerical Weather Prediction, can not provide the reconstructions with a spatial resolution that can compete with satellite imagery. In this work we propose a framework based on Variational Data Assimilation and Deep Learning concepts. This framework is applied to recover rich-in-time, high-resolution information on sea surface wind speed. We design our experiments using synthetic wind data and different sampling schemes for high-resolution and low-resolution versions of original data to emulate the real-world scenario of spatio-temporally heterogeneous observations. Extensive numerical experiments are performed to assess systematically the impact of low and high-resolution wind fields and in-situ observations on the model reconstruction performance. We show that in-situ observations with richer temporal resolution represent an added value in terms of the model reconstruction performance. We show how a multi-modal approach, that explicitly informs the model about the heterogeneity of the available observations, can improve the reconstruction task by exploiting the complementary information in spatial and local point-wise data. To conclude, we propose an analysis to test the robustness of the chosen framework against phase delay and amplitude biases in low-resolution data and against interruptions of in-situ observations supply at evaluation time
Paper Structure (23 sections, 15 equations, 13 figures, 6 tables)

This paper contains 23 sections, 15 equations, 13 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Problem statement
  3. Data
  4. Low-resolution data
  5. High-resolution data
  6. In-situ time series
  7. Preprocessing scheme
  8. Methods
  9. Trainable data assimilation scheme
  10. Direct learning-based inversion
  11. Learning scheme
  12. Numerical implementation
  13. Results
  14. Evaluation framework
  15. Benchmark analysis results
  16. ...and 8 more sections

Figures (13)

  • Figure 1: Dataset qualitative characteristics. Panel (a): Geographical region considered. The red markers represent the buoys positions. Panel (b): Sample in-situ time series pseudo-observations obtained from wind speed values at buoys positions. Panel (c): The downsampling-reinterpolation step to obtain low-resolution pseudo-observations from ground-truths. The spatial resolution for this case is 30 kilometers.
  • Figure 2: Panel (a): Temporal sampling patterns of the data used. The items "HR (12 h)" and "HR (24 h)" refer to the datasets in which the HR observations are simulated to have temporal frequency of 12 or 24 hours. These items refer to different experimental configurations and are depicted on the same plot for graphical convenience. Panel (b): An example of the dataset items. The temporal sampling frequencies of HR and LR fields are fictitious and aim to illustrate the dataset.
  • Figure 3: First row: (left) Original data, (middle) reconstruction of the $B_1$-SR baseline, (right) reconstruction of the $M_m$-C3 4DVarNet. Second row: (left) Map of average MSE related to the $B_1$-SR baseline, (middle) $M_m$-C3 4DVarNet and (right) map of the average relative gain of the 4DVarNet w.r.t. the baseline. The temporal sampling frequency for high-resolution fields is 12 hours. The two rows are displaced in order for the baseline and model reconstructions and error maps to match vertically.
  • Figure 4: Maps of average relative gains. Left panel: Non-trainable observation operator. Average gain of model $M_s$-C3 w.r.t. $M_s$-C1. Right panel: Trainable observation operator. Average gain of model $M_m$-C3 w.r.t. $M_m$-C1. The temporal sampling frequency for high-resolution fields is 12 hours.
  • Figure 5: Top row: test case of LR simulated delay. Bottom row: test case of LR simulated re-modulation. The suffixes "-rd" and "-ri" identify the models trained in case of random delay and re-modulation, respectively. The experimental configuration used for this experiment is the case C3 for the $M_m$ model and the SR, C3 cases for the $B_1$ baseline. The temporal sampling frequency of high-resolution fields is 12 hours.
  • ...and 8 more figures