Project

Back to overview

Bridging the gap between machine learning and geostatistical simulations

English title Bridging the gap between machine learning and geostatistical simulations
Applicant Gravey Mathieu
Number 195057
Funding scheme Early Postdoc.Mobility
Research institution Institut des dynamiques de la surface terrestre Université de Lausanne
Institution of higher education Institution abroad - IACH
Main discipline Other disciplines of Environmental Sciences
Start/End 01.02.2021 - 31.10.2021
Show all

All Disciplines (3)

Discipline
Other disciplines of Environmental Sciences
Other disciplines of Earth Sciences
Mathematics

Keywords (5)

Multiple-point statistics; Machine learning; Statistical simulation; Geostatistics; Interpolation

Lay Summary (French)

Lead
Dans le monde, les données sont collectées de manière sporadique en quelques points précis uniquement. De ce fait, nous avons besoin de méthode pour interpoler sur les zones inconnues, et ce de manière réaliste et efficace.
Lay summary
Ce projet vise à utiliser l'apprentissage automatique pour produire des interpolations réalistes. La solution proposée surmonte les limitations actuellement présentes dans les méthodes existantes. Les méthodes d'interpolation sont actuellement des outils essentiels dans divers domaines, de la géologie et l'hydrogéologie en passent par les sciences de la surface de la Terre (par exemple, la télédétection, la modélisation du climat, la météo, l'écologie). Tous ces domaines d'étude nécessitent les méthodes d'interpolation afin de comprendre comment une variable évolue dans l'espace. Chaque méthode d'interpolation convient le mieux à un domaine d'études spécifique. Dans ce projet on se concentre sur les interpolations qui produisent des structures spatialement cohérentes appliquées à la télédétection et particulier la colorisation d’ancienne image satellite. La méthode proposée propose plusieurs scénarios d’interpolation, ce qui permet d’étudier l’incertitude qui y est associée. La méthode proposée utilise une image d’entrainement pour inférer la structure spatiale à utiliser pour l’interpolation. Enfin, une approche basée sur l’apprentissage automatique permet à l’utilisateur d’avoir une solution clé en main, et n’a donc pas besoin de perdre du temps dans la fastidieuse tâche qui consiste à trouver les bons paramètres de calibration.
Direct link to Lay Summary Last update: 20.05.2020

Responsible applicant and co-applicants

Abstract

This project aims at using machine learning (ML) to improve state of the art methods to produce realistic spatial interpolations. The proposed solution overcomes the limitations that are currently present in the existing methods. The current interpolation methods are vital tools in various fields, from geology and hydrogeology to earth surface sciences (e.g., remote sensing, climate modeling, weather, ecology). All these fields of study require the interpolation methods in order to understand how a variable evolves in space. Each interpolation method suits a specific field of study best; this project focuses on interpolations that produce spatially coherent structures applied to remote sensing (in particular on the enhancement of poor - and usually cheap - satellite images). To handle the variability in the spatial structures and also to take uncertainty into consideration, many different interpolations are required (each called a simulation). Each interpolation represents a different scenario, respecting underlying spatial structure. Nowadays methods can enhance poor satellite images, however, they only use a small fraction of all the information available, and therefore we are still relatively far from the optimal outcome.This problem of generating structurally coherent interpolations was historically addressed by two distinct approaches: 1- Gaussian random fields which uses a mathematical framework and 2- multiple point statistics (MPS) which uses images of analogues to obtain the information needed for the interpolation. While the first approach is mathematically robust, the major issue lies in its inability to simulate complex structures. In contrast, structural complexity is embedded in the training images used by MPS algorithms to do simulations. However, these solutions come with some limitations of their own, for instance they need to be placed over a grid and they cannot well reproduce the rare events (usually extremes). But probably the most critical limitation is that extrapolating to new (unencountered) scenarios is impossible. Finally, both approaches require extra algorithmic parameters or mathematical models to be tuned. These parameters are sometimes hard to determine and need extensive trial-and-error.Current solutions are either limited to simple models, or by the impossibility of generalizing structures to new situations. In all cases, current approaches struggle with huge datasets. This project aims to develop a 3rd approach that challenges these established facts, by proposing a state-of-the-art Machine Learning approach that will allow extrapolation, reproduction of complex structures, self-parametrization using the significant dataset available in remote sensing field. The proposed approach in this project capitalizes on sequential simulations that are well known as operational strategies. Furthermore, ML will allow quick adaptation to new situations and datasets by learning from the previous case studies.
-