The Physical Limit of Neural Hypoxia Detection in the Black Sea from Satellite Observations

2026

Read The Physical Limit of Neural Hypoxia Detection in the Black Sea from Satellite Observations (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-25608.txt.

Introduction

Mangeleer and three colleagues test how far satellite observations can go in detecting coastal hypoxia in the Black Sea. Summer stratification blocks renewal of bottom oxygen used up by respiration. They frame the task as a Bayesian inverse problem linking surface observations to full three-dimensional physical and biogeochemical states. A deep generative neural network trained on numerical model outputs approximates the true posterior of sea states. Accurate estimation reaches only the mixed layer, and in summer the method detects 38 percent of hypoxic events with a precision of 47 percent.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31293
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.25608

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

The Physical Limit of Neural Hypoxia Detection in the Black Sea from Satellite Observations Thanks: This work was co-supervised by Marilaure Grégoire and Gilles Louppe. Abstract Coastal hypoxia ( O 2 < 63 O_{2}<63 [ m ​ m ​ o ​ l / m 3 mmol/m^{3} ]) threatens ocean health worldwide. Bottom oxygen consumed by respiration cannot be renewed, making monitoring essential to protect vulnerable marine ecosystems and reduce biodiversity loss. Despite the growing availability of Black Sea satellite observations, no operational system currently exploits them to directly infer the oxygen state in real time. This can be framed as a Bayesian inverse problem relating surface observations to the complete Black Sea states. Here, we solve it using a deep generative neural network trained on numerical model outputs, providing a tractable approximation of the true posterior distribution of sea states. We find that accurate state estimation is limited to the mixing layer, because its homogeneity makes surface conditions representative of subsurface states. During summer, we detect 38% of all hypoxic events shelf-wide with a precision of 47%. Improving results will likely require longer assimilation windows or additional observations. † † journal: Geophysical Research Letters † † authors: Victor Mangeleer Luc Vandenbulcke Marilaure Grégoire Gilles Louppe † † corresponding: Victor Mangeleer, [email protected] \justifying Department of Modelling for Aquatic Systems (MAST), University of Liège Montefiore Institute of Electrical Engineering and Computer Science, University of Liège keypoints The prediction of bottom coastal hypoxia from surface satellite-type observations is investigated with a deep generative neural network. The neural network emulator detects 38% of summer hypoxic events shelf-wide with a precision of 47%. The complete three-dimensional state is emulated with robust accuracy within the mixing layer throughout the year. Plain Language Summary In coastal waters, excessive fertilizer use causes eutrophication, producing organic matter whose decomposition consumes oxygen. When ventilation cannot replenish this loss, oxygen drops below a critical threshold, creating hypoxia that threatens marine ecosystems. Despite the growing availability of satellite observations, no operational system currently exploits them to directly infer oxygen levels in real time. We address this gap by solving the Bayesian inverse problem that estimates the complete Black Sea state, including oxygen, from satellite surface observations alone. This is challenging since multiple states may produce the same observations. Traditional methods estimate only the most likely state, overlooking rare hypoxic events. Here, we use a deep generative model trained on numerical model outputs to recover the full range of plausible states, from which oxygen is inferred alongside the other observed variables. We find that surface observations enable accurate state estimation within the mixing layer year-round, because homogeneity within this layer makes surface conditions representative of subsurface states. During summer, we detect 38% of all hypoxic events shelf-wide, with 47% of our detections correctly identifying hypoxia. Improving accuracy across depths will likely require longer assimilation windows or additional observations (e.g., float measurements). 1 Introduction Hypoxia ( O_{2}<63 [ mmol/m^{3} ]) is a worldwide phenomenon affecting the bottom of many coastal ocean areas 26 ; 5 ; 21 . It has detrimental effects on benthic life, particularly crustaceans and bivalves, and more generally on low-mobility organisms that cannot escape. Hence, accurate and timely estimates of oxygen levels are key to supporting informed management decisions 6 . The standard approach for estimating oxygen levels and detecting hypoxia relies on numerical ocean models, which simulate the full dynamics of the ocean. In the framework of the Copernicus Marine Environment Monitoring Service (CMEMS) 30 , 10-day forecasts and multidecadal reanalysis are routinely produced for the ocean physics and biogeochemistry, and, in particular, for oxygen. Yet these physics-based models are complex to run and not easily amendable by non-experts. In operational conditions, satellites continuously observe the sea surface and measure temperature, salinity, chlorophyll, and sea surface height at ever-increasing temporal, spatial, and spectral resolutions 30 . However, these observations remain confined to the sea surface and do not directly measure oxygen levels. No operational system currently exploits them to infer oxygen without first running a numerical simulation and then using the observations to correct it. The real-time availability of these observations thus raises the question of whether surface data alone could enable reliable state estimation or hypoxia detection without requiring a numerical simulation. The possibility of estimating the current sea state, and particularly inferring oxygen levels, from the present-day surface observations is supported by physical and biogeochemical processes 6 . Bottom hypoxia typically occurs when oxygen consumed by benthic respiration exceeds physical replenishment. This happens under poor ventilation and intense organic matter degradation, most commonly in summer when: ( 1 ) the water column is stratified, ( 2 ) bottom temperatures peak, and ( 3 ) sediments receive organic matter exported after the main surface bloom. This summer decoupling between surface and bottom waters challenges bottom condition retrieval from surface observations. Yet, winter mixing couples surface (temperature, chlorophyll) to the bottom oxygen levels, which condition summer hypoxia. Overall, we argue that the inverse problem linking current surface observations to the subsurface state is indeed physically supported. Motivated by the remarkable effectiveness of deep generative neural networks in solving inverse problems across various tasks 26 ; 14 ; 32 ; 3 ; 33 ; 20 , we address ours through a two-stage approach. First, we train a diffusion model on a numerical simulation of sea physics and biogeochemistry evolution to learn daily dynamics. Then, we use the diffusion model to estimate the actual sea state by constraining the generation to surface observations. Due to its stochastic nature, the diffusion model enables generation of multiple state estimates. Combined, these plausible solutions quantify estimation uncertainty and define the posterior distribution of the inverse problem. Concurrently, 29 used the same approach to reconstruct subsurface velocity and buoyancy fields from sea surface height in an idealized double-gyre configuration. Our work extends theirs to a realistic sea and to biogeochemical variables. The paper is organized as follows. We first validate our diffusion prior through experiments that include analyzing seasonal means, power spectra, and reproducibility of spatial structures. We then solve the inverse problem and assess state estimation accuracy, identify its depth limitations, and establish baseline hypoxia detection results from surface observations alone. 2 Background An inverse problem can be posed as a Bayesian inference problem 25 ; 2 , where the solution is the posterior distribution p(x\mid y)\propto p(y\mid x)\,\,p(x) . Here, p(x) is the prior distribution over plausible sea states x , and p(y\mid x) is the likelihood describing how observations y relate to a given state x . To solve the inverse problem, we start by learning a diffusion model of the prior p(x) . Following 28 , samples x\sim p(x) are progressively perturbed through a diffusion process expressed as a stochastic differential equation (SDE) \operatorname{d}\!{x(t)}=f(t)\,x(t)\operatorname{d}\!{t}+g(t)\operatorname{d}\!{w(t)}\,, (1) where f(t)\in\mathbb{R} is the drift coefficient, g(t)\in\mathbb{R}^{+} is the diffusion coefficient, w(t) denotes a standard Wiener process, and x(t) is the perturbed sample at time t\in[0,1] . Since the SDE is linear in x(t) , the perturbation kernel from x to x(t) is Gaussian and takes the form p(x(t)\mid x)=\mathcal{N}(x(t)\mid\alpha(t)\,x,\Sigma(t)) , where \alpha(t) and \Sigma(t)=\sigma(t)^{2}\,\mathbb{I} are derived analytically from f(t) and g(t) . The forward SDE ( 1 ) admits a family of reverse SDEs 1

Read the complete text