Machine Learning methods for the Atmosphere, the Ocean, and the Seabed

Abstract

Understanding, observing, and forecasting the environment are all essential step to support a more sustainable interactions between human activities and the environment. Several areas of environmental modelling and classical analysis can be beneficial from the application of novel approach such as Machine Learning techniques. In particular, we are currently working on multiple areas for the development of machine learning techniques to be applied for (1) the modeling of the convection permitting dynamical model for precipitation forecasting, (2) data interpolation for ocean observing systems, in particular using data collected with ARGO floats, and (3) the automatic classification of seabeds for the assessment of geological hazards. Here we detail the current state of the projects in those area and directions for future research.

Keywords

• eol>Deep Learning
• Digital Twins of the Ocean
• Oceanography
• Seabed Classification
• Geohazards
• Disaster Risk Forecasting

-

1. Introduction diferent aspects of the environment, adding ML as a support for existing simulation models or to manual labeling and analysis of data.

Artificial Intelligence (AI) and machine learning (ML) methods are starting to change the way the environment is modeled, forecasting is performed, and, in general, the way scientific computing will support research in the 2. Deep Learning for Precipitation future. In fact, as proposed in [ 1 ], there is a new merging and Associated Risk of scientific computing, scientific simulation, and AI that results in what the authors call “simulation intelligence”. Every year across the world natural catastrophes due

The works presented in this abstract are all part of to extreme weather and climate events cause casualties this general movement towards the integration of ML and significant damage to properties and assets. Disaster techniques in the scientific field. In particular, here we risk forecasting highly depends on the ability to correctly present three diferent areas in which there are currently quantify the phenomena related hazards, specifically at active project: high spatial and temporal resolution. As for the precipitation phenomenon, the classical method of deriving • Deep learning techniques for the forecasting of precipitation distribution from simulations of dynamical the precipitation distribution; models is computationally too expensive when it comes • Deep Learning for data interpolation in ocean to high resolution, limiting its application. ML models observing systems; can help in this regard, by levering the huge amount of • Automatic classification of the seabed. data available from historical records and models simuAs it is possible to observe, the three projects covers lations. For the precipitation phenomenon few studies have been carried out to date, among them [ 2 ] and [ 3 ], Ital-IA 2023: 3rd National Conference on Artificial Intelligence, orga- both based on deep learning techniques. nized by CINI, May 29–31, 2023, Pisa, Italy In this direction, we are developing a novel deep learn*$[e-mail] (V. Blasone); [e-mail] ing framework which represents a first attempt in emu(U. Di Laudo); [e-mail] (G. Pietropolli); lating the convection permitting dynamical models. The [e-mail] (L. Bortolussi); [e-mail] main objective is to improve the projection of climatic (S. Ceramicola); [e-mail] (G. Cossarini); impact-drivers relevant for risk assessment, in a much [e-mail] (L. Manzoni) more eficient way. In its first version, the framework 0000-0001-5814-8532 (V. Blasone); 0000-0001-7623-8419 includes convolutional, recurrent and graph neural net(0G00.0P-i0e0tr0o2p-1o3ll1i)8;-01020702-(0S0.0C1-e8r8a7m4i-c4o0l0a1);(0L0.0B0o-0rt0o0l1u-s7s8i)0;3-8568 works, to deal with the intrinsic characteristics of the data (G. Cossarini); 0000-0001-6312-7728 (L. Manzoni) and the associated challenges. The input data is derived © 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License from the ERA5 reanalysis dataset [ 4 ], with hourly values CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) for temperature, specific humidity, eastward/northward of Deep Learning model and data assimilation to improve wind components and geopotential atmospheric parame- the forecast skill of a marine model forecast system of ters, on a low-resolution grid of approximately 25 km. In the Mediterranean Sea [ 6, 7 ]. Specifically, the deep learna supervised perspective, the target is represented by the ing model is set to generate relationships between highGRIPHO hourly precipitation observations dataset on a frequency sampled variables and low-frequency ones high-resolution grid of 3 km [ 5 ]. The framework was ap- with the aims to generate reconstructed observations for plied to the northern Italy and successfully trained using the data assimilation. We consider as a dataset to train a time span of 15 years, from 2001 to 2015. Projections the deep learning model the collection of measurements in terms of yearly and seasonal precipitation distribu- collected by the so-called ARGO profiling float. Existing tion maps for the year 2016 were derived and compared applications based on a feed-forward model (e.g., mulwith observed values, showing a good capacity in resem- tilayer perceptron) are unaware of the typical shape of bling the precipitation distribution. Finally the model the profiles of biogeochemical variables that they try to performance was tested in describing an extreme event. infer. To overcome this issue we tested an innovative

Future research will focus on improving the frame- approach based on convolutional deep learning architecwork, extending its applicability in spatial-temporal ture to reconstruct nutrient profiles. The underpinning terms. The medium-term goal is to replace the input idea is that the typical shape of the vertical profiles of a reanalysis data with data from simulations of dynamical variable is a constraint that has to be learned during the models for the same atmospheric parameters. The long- training. Preliminary experimental results confirm that term goal is to produce risk maps (e.g., floods) for the the curves produced by the convolutional architecture coming decades, integrating hydrological and vulnerabil- guarantee the generation of smoother – and thus, more ity information. similar to those of the real world – profiles. A second application consists of the direct integration though a deep learning approach of observations and the determin3. Deep Learning for Ocean istic model output to predict 3D fields of biogeochemical Observing Systems variables in the Mediterranean Sea by integrating observations and the output of an existing deterministic model, that is MedBFM. The deep learning architecture that we exploited for the aforementioned task is based on the inpainting [ 8 ], a computer vision technique developed to ifll missing pixels of a considered image. Here, the resulting ML model can instead be used to fill the gaps between the observations in a way that “corrects” the output of the determinstic model. We successfully developed and trained this ML model on a portion of the Mediterranean Sea, obtaining promising results, and now we are working on the extension of this to the whole Mediterranean area.

Improving the capability of monitoring and predicting the status of the marine ecosystem has important implications, also considering the changes caused by human activities. In fact, marine ecosystem health is impacted by human activity: during the last decades, the Ocean has been increasingly afected by global changes (e.g. acidification) caused by the exponential augmentation of human assets. Among the many applications of AI in the oceanographic field, we deal with the automated realtime production of short-term forecasts of the state of the sea, with the aim of increasing the reliability of modeling predictions by correcting model results based on real-time observations of physical, chemical, geological and biological processes in the seas and oceans. Investi- 4. Automated classification of the gating marine ecosystem evolution and variability can be seabed based on observations and modelling. In general, observations are accurate but limited and sparse both in time Geological hazards or geohazards are the result of natural, and space, and, most importantly, unevenly available active geological processes. They include volcanic erupamong diferent variables. On the other hand, models tions, earthquakes, tsunamis, landslides and several types reproduce ecosystem dynamics and cover diferent spa- of mass wasting phenomena. Geohazards can endanger tial and temporal scales of the processes but they can be and cause damage mainly in coastal areas where people inaccurate due to several source of uncertainties. live and important economic infrastructures (harbors,

Integrating models and observations is widely used to highways, airports, and so on) are located. provide optimal (e.g., in statistical sense) estimates of the The assessment of geohazards is the basis for carrying state of the oceans. Classically, it is done through data out susceptibility and risk assessment and to apply a susassimilation approaches (e.g. Kalman filter, variational tainable management of the seafloor and coastal areas approaches). Here, we propose two novel approaches us- as indicated in several European and international direcing AI techniques to advance the model/observation inte- tives. Assessing geohazards in marine environments is a gration. The first application consists on the integration very time consuming and costly exercise as it implies using research vessels, geological and geophysical expertise useful to the coastal community dealing with hazards and expensive tools. In addition geological interpretation and disasters. and seabed mapping can be very subjective to experience and time available.

Having the help of human-supervised AI could be 5. Final Remarks strategic when dealing with huge databases and to reduce the inevitable human subjectivity when interpreting As it is possible to observe, ML model are being develseabed data, thus providing a uniform interpretation. The oped for a large number of applications for climate, sea, goal of this project is to construct a model that performs and seabed. In all cases it is necessary to have a close an automated classification for the seabed and that in par- collaboration with the domain experts in order to underticular, is able to find and recognize in automated way (or stand the requirements for a model and to evaluate it. better automated human supervised) features indicative Some advantages of machine learning in those areas are or prone of to be geohazards. the ability to automate expensive human activities, e.g.,

First, we want to construct a model that automati- manual labeling of geohazards, to integrate information cally map the seabed, using a bathymetric map as input. from multiple sources even if the relation between them For this purpose, we are using a model [ 9 ] that consist has not already be formalized (i.e., there is no “classical” in two Region-based Convolutional Neural Network (R- model explaining the relation between two variables), CNN) [ 10 ]. GIS data (that specify the depth, the slope and to capture spatial and temporal relations only from and the curvature of the seabed) are used as input and the data. One obstacle is, however, that a ground truth treat them as an RGB image. The image is cut in small is not always present, but usually only sparse data and windows and for each window they apply the Selective deterministic models (which are themselves approximaSearch algorithm in order to localize features and then tions of the reality). Hence, ML models should be able train the networks [ 11 ]. The first CNN is basically a bi- to deal with uncertainty in the input data – and possibly nary classification and tell us if there is or not a feature provide a corresponding estimation of uncertainty in the indicative of hazard in some region; the second one classi- output. ifes features into diferent classes. Transfer learning from VGG19 is used to train the networks. The overall objec- References tive is to improve the existing model described above in order to have a best eficiency and less time-consuming predictions.

Once we have a model that, given unlabelled data, returns us the overall structure of the seabed, it comes the complex step of assessing geohazards. Indeed, the fact that a seabed feature (i.e., a submarine canyon) is prone or not to hazard, depends from several factors: its depth, its geological activity, its proximity to the coast etc. It is easy to understand that it is not trivial to develop such a model but would be a significant support to practitioner to enable the reduction of vulnerability and to enhance community resilience to disasters. For these reasons it is important for the resulting ML-driven method to be able to estimate the likelihood that seabed features may represent potential hazards. Given that this step is currently not automate, such a model would provide a signficant step forward in geohazards assessment for huge regions such as for the Mediterranean sea.

The main goal of the project is to construct an AI tool that is able to label a seabed map with all the possible features indicative of geohazards. An further step would be to provide an explainable model, providing a justification for the choice the model makes about the possibility of having or not an hazard. In other words developing an AI model with human supervision able to provide information and, possibily, explanation, in the assessment of geohazards in marine environment would be extremely

References

[1] A. Lavin , D. Krakauer , H. Zenil , J. Gottschlich , T. Mattson , J. Brehmer , A. Anandkumar , S. Choudry , K. Rocki , A. G. Baydin , C. Prunkl , B. Paige , O. Isayev , E. Peterson, P. L. McMahon , J. Macke , K. Cranmer , J. Zhang , H. Wainwright , A. Hanuka , M. Veloso , S. Assefa , S. Zheng , A. Pfefer , Simulation intelligence: Towards a new generation of scientific methods , 2022 . arXiv: 2112 . 03235 .

[2] J. Wang , Z. Liu , I. Foster, W. Chang , R. Kettimuthu , V. R. Kotamarthi , Fast and accurate learned multiresolution dynamical downscaling for precipitation , Geoscientific Model Development 14 ( 2021 ) 6355 - 6372 . doi: 10 .5194/gmd-14- 6355 - 2021 .

[3] C. Wang , P. Wang , P. Wang , B. Xue , D. Wang , A spatiotemporal attention model for severe precipitation estimation , IEEE Geoscience and Remote Sensing Letters 19 ( 2021 ) 1 - 5 . doi: 10 .1109/LGRS. 2021 . 3084293 .

[4] H. Hersbach , W. Bell , P. Berrisford , A. Horányi , M.- S. J., J. Nicolas , R. Radu , D. Schepers , A. Simmons , C. Soci , D. Dee , Global reanalysis: goodbye erainterim, hello era5 , 2019 . URL: https://www.ecmwf. int/node/19027. doi: 10 .21957/vf291hehd7.

[5] A. Fantini , et al., Climate change impact on flood hazard over italy , 2019 .

[6] S. Salon , G. Cossarini, G. Bolzon , L. Feudale , P. Lazzari , A. Teruzzi , C. Solidoro , A. Crise , Marine ecosystem forecasts: skill performance of the cmems mediterranean sea model system (????).

[7] G. Cossarini , L. Feudale , A. Teruzzi , G. Bolzon, G. Coidessa, C. Solidoro , V. Di Biagio , C. Amadio , P. Lazzari , A. Brosich , et al., High-resolution reanalysis of the mediterranean sea biogeochemistry ( 1999 -2019), Frontiers in Marine Science 8 ( 2021 ) 1537 .

[8] G. Pietropolli , G. Cossarini , L. Manzoni, Gans for integration of deterministic model and observations in marine ecosystem , in: Progress in Artificial Intelligence: 21st EPIA Conference on Artificial Intelligence, EPIA 2022 , Lisbon, Portugal, August 31- September 2 , 2022 , Proceedings, Springer, 2022 , pp. 452 - 463 .

[9] R. Soemantoro , M. Gaimann , Machine Learning How to Sea: Recognition and classification of subsea structures using artificial intelligence and highperformance computing , PRACE: Partnership for Advanced Computing in Europe , 2022 , pp. 35 - 37 .

[10] R. Girshick , J. Donahue , T. Darrell , J. Malik, Rich feature hierarchies for accurate object detection and semantic segmentation , in: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) , 2014 , p. 580 - 587 .

[11] J. R. R. Uijlings , K. E. A. van de Sande, T. Gevers , A. W. M. Smeulders , Selective search for object recognition , in: International Journal of Computer Vision , Springer, 2013 , pp. 154 - 171 .