1,721,016 research outputs found
Apprentissage profond pour reconstruire la hauteur de la surface océanique à partir d’observations satellites multivariées
No full textCette thèse de doctorat porte sur la reconstruction d'images satellites de la surface de l'océan à partir de mesures éparses et bruitées. Son objectif est l'estimation de la hauteur de la mer (SSH), une variable importante pour approximer les courants de surface. Elle est actuellement mesurée par des altimètres pointant au nadir, laissant de nombreuses zones non observées. Les cartes complètes de SSH sont produites en utilisant des interpolations optimales linéaires présentant une faible résolution effective.D'autre part, la température de surface de la mer (SST) est observée sur des zones plus étendues et est physiquement liée aux courants géostrophiques à travers l'advection.Cette thèse explore les algorithmes d'apprentissage profond pour estimer les champs de SSH. En s'appuyant sur des années de données de simulation et d'observations, les réseaux neuronaux profonds sont capables d'apprendre des relations complexes entre les variables SSH et SST. Nous utilisons ces algorithmes ainsi que les observations de température, pour reconstruire la SSH d'abord dans une perspective de réduction d'échelle sur une simulation physique. Ensuite, nous considèrerons le problème de son interpolation sur des données de simulation et d'observation, en nous concentrant particulièrement sur la manière de transférer l'apprentissage dans des contextes opérationnels. Enfin, nous adaptons notre méthode pour produire des estimations en temps réel et des prévisions.This Ph.D. thesis focuses on reconstructing satellite images of the ocean surface from sparse and noisy measurements. Our objective is the Sea Surface Height (SSH), an important variable to estimate surface currents. It is retrieved through nadir-pointing altimeters, leaving important observation gaps due to their remote sensing technology. Complete SSH maps are produced using linear Optimal Interpolations with low effective resolution.On the other hand, Sea Surface Temperature (SST) products have much higher data coverage, and SST is physically linked to geostrophic currents through advection.This thesis explores deep learning algorithms to estimate SSH fields. Relying on years of data from simulation and observations, deep neural networks are able to learn complex relationships between SSH and SST variables. Using these algorithms and SST observations, we first enhance SSH mapping from a downscaling perspective on a physical simulation. Then, we tackle the SSH interpolation problem on simulation and observation data, with a particular focus on how to transfer the learning in operational settings. Finally, we adapt our method to produce near real-time and forecast estimations
Can a neural network learn a numerical model ?
No full textInternational audienceNumerical models are used to simulate the evolution of atmosphere or ocean dynamics. They are implemented through a computer code, that contains predefined rules specifying how to compute the evolution of some outputs (e.g sea surface height) from inputs (e.g. previous states of the model, satellite or in situ observations of other parameters). A machine learning approach, in contrast, infers its internal set of rules from a large amount of data. In many fields (image recognition, automatic translation, speech recognition, ...), the more traditional methods, which rely on predefined rules, have been outperformed by machine learning algorithms. This performance was made possible by advances in Convolutional and Recurrent Neural Networks. This work addresses the question of the application and the usefulness of machine learning for numerical modeling in Geophysics. Results are presented using a demonstration model : A shallow-water model including a forcing by the wind, a diffusive term and a dissipation term. We evaluate the ability of a neural network to reproduce the numerical model "rules" given only the output fields of the model. We also investigate the ability of this neural network to simulate only some specific parts of the numerical model (e.g. diffusion or dissipation) and discuss the potential combination of approaches
Leveraging Deterministic Weather Forecasts for In Situ Probabilistic Temperature Predictions via Deep Learning
No full textInternational audienceWe propose a neural network approach to produce probabilistic weather forecasts from a deterministic numerical weather prediction. Our approach is applied to operational surface temperature outputs from the Global Deterministic Prediction System up to 10-day lead times, targeting METAR observations in Canada and the United States. We show how postprocessing performance is improved by training a single model for multiple lead times. Multiple strategies to condition the network for the lead time are studied, including a supplementary predictor and an embedding. The proposed model is evaluated for accuracy, spread, distribution calibration, and its behavior under extremes. The neural network approach decreases the continuous ranked probability score (CRPS) by 15% and has improved distribution calibration compared to a naive probabilistic model based on past forecast errors. Our approach increases the value of a deterministic forecast by adding information about the uncertainty, without incurring the cost of simulating multiple trajectories. It applies to any gridded forecast including the recent machine learning-based weather prediction models. It requires no information regarding forecast spread and can be trained to generate probabilistic predictions from any deterministic forecast
Infering the Chlorophyll-A Vertical Distribution in the Ocean From Satellite Data by Using Hidden Markov Models and Self Organizing Maps
No full textRemote sensing the sea surface CO<sub>2</sub> of the Baltic Sea using the SOMLO methodology
No full textInternational audienceStudies of coastal seas in Europe have noted the high variability of the CO2 system. This high variability, generated by the complex mechanisms driving the CO2 fluxes, complicates the accurate estimation of these mechanisms. This is particularly pronounced in the Baltic Sea, where the mechanisms driving the fluxes have not been characterized in as much detail as in the open oceans. In addition, the joint availability of in situ measurements of CO2 and of sea-surface satellite data is limited in the area. In this paper, we used the SOMLO (self-organizing multiple linear output; Sasse et al., 2013) methodology, which combines two existing methods (i.e. self-organizing maps and multiple linear regression) to estimate the ocean surface partial pressure of CO2 (pCO2) in the Baltic Sea from the remotely sensed sea surface temperature, chlorophyll, coloured dissolved organic matter, net primary production, and mixed-layer depth. The outputs of this research have a horizontal resolution of 4 km and cover the 1998–2011 period. These outputs give a monthly map of the Baltic Sea at a very fine spatial resolution. The reconstructed pCO2 values over the validation data set have a correlation of 0.93 with the in situ measurements and a root mean square error of 36 μatm. Removing any of the satellite parameters degraded this reconstructed CO2 flux, so we chose to supply any missing data using statistical imputation. The pCO2 maps produced using this method also provide a confidence level of the reconstruction at each grid point. The results obtained are encouraging given the sparsity of available data, and we expect to be able to produce even more accurate reconstructions in coming years, given the predicted acquisition of new data
Retrieving the evolution of vertical profiles of Chlorophyll-a from satellite observations using Hidden Markov Models and Self-Organizing Topological Maps
No full textInternational audienc
Predicting Ocean Dynamics through Machine Learning: Application on Sea-Surface Suspended Particulate Mater
No full textInternational audienceIn the satellite age, geoscientist have acquired an unprecedented aboundance of data describing the earth (ocean and land) surface. This accumulation of observations with high spatio-temporal sampling has generated a demand in ways to optimally extract from these data the useful features which have the ability to forecast the evolution of some key parameter. In this work we explore the high potential of using advanced machine learning techniques for the prediction of the temporal evolution of 2D oceanographic parameters.We chose to present an experiment on the prediction of sea-surface fields of the total suspended particulate mater in the english chanell. This choice was motivated by the complexity of the phenomenons impacting this oceanic variable: it is driven both by the neap-tide cycle, storms and general circulation oceanic currents.The predicting system used is constructed using three successive blocks. The first is consisting in a convolutional neural network to extract useful feature and reduce the dimension of the input. The second is a recurrent neural network which is used as a feature predictor. The last block is a convolutional neural network used to reconstruct the image from the predicted feature of the last block.An additional motivation was the frequent missing values caused by the cloud cover over the area. A number of neuronal methods are able to produce good predictions despite missing values.The methodology we selected to implement is a combination of convolutionary neuronal networks and long short-term memory networks.Preliminary results indicate a predictive power for the mean situation and also for extreme events (e.g. storms) than is comparable or better than traditional approaches
Deep prior in variational assimilation to estimate ocean circulation without explicit regularization
No full textInternational audienceMany applications in geosciences require solving inverse problems to estimate the state of a physical system. Data assimilation provides a strong framework to do so when the system is partially observed and its underlying dynamics is known to some extent. In the variational flavor, it can be seen as an optimal control problem where initial conditions are the control parameters. Such problems being often ill-posed, regularization may be needed using explicit prior knowledge to enforce satisfying solution. In this work we propose to use a deep prior, a neural architecture that generates potential solution and acts as implicit regularization. The architecture is trained in an fully-unsupervised manner using the variational data assimilation cost so that gradients are backpropagated through the dynamical model and then through the neural network. To demonstrate its use, we set a twin experiment using a shallow-water toy model, where we test various variational assimilation algorithms on a ocean-like circulation estimation
Generating ensembles of spatially coherent in situ forecasts using flow matching
No full textInternational audienceWe propose a machine‐learning‐based methodology for in situ weather forecast postprocessing that is both spatially coherent and multivariate. Compared with previous work, our Flow MAtching Postprocessing (FMAP) represents the correlation structures of the observation distribution better, while also improving marginal performance at stations. FMAP generates forecasts that are not bound to what is already modeled by the underlying gridded prediction and can infer new correlation structures from data. The resulting model can generate an arbitrary number of forecasts from a limited number of numerical simulations, allowing for low‐cost forecasting systems. A single training is sufficient to perform postprocessing at multiple lead times, in contrast with other methods, which use multiple trained networks at generation time. This work details our methodology, including a spatial attention transformer backbone trained within a flow‐matching generative modeling framework. FMAP shows promising performance in experiments on the EUMETNET Postprocessing Benchmark (EUPPBench ) dataset, forecasting surface temperature and wind‐gust values at station locations in western Europe up to five‐day lead times
- …
