Mathieu Gravey

The content below is no longer maintained. Mathieu Gravey’s official webpage is here: https://www.mgravey.com.

KEYWORDS: multiple-point statistics, MPS simulation, remote sensing, satellite imagery, colorization, spectral enhancement, GPU-based topography point clouds completion methods, Machine Learning, Neural Network, GAN, VAE/AE.

RESEARCH STATEMENT

Recent satellite imagery is a vast improvement in terms of spatial, spectral and temporal resolution over archive imagery. My Ph.D. focused on a statistical approach that reconstructs the spectrum of old satellite imagery to match quality of recent imagery. This enables the use of a single processing pipeline on a variety of different datasets acquired with diverse sensors.

Methodologically, I improve and develop new and efficient algorithms especially for pixel-based multiple-point statistics (MPS) simulations. My research develops MPS algorithms that are easier to configure and that benefits from new computing architecture, for CPUs with parallelization, multi-socket, vectorization and advance use of Xeon Phi, and for coprocessors such as GPU or FPGA. Currently, I am working on a autocalibration for MPS to enable non-expert and user-friendly utilization of these methods.

My methods are applicable, for example, for spectral enhancement, also known as colorization, of old or incomplete satellite imagery. This process can be divided into two main topics: the disaggregation, that can be seen as a deconvolution, like reconstruction of color from gray; and extrapolation, like reconstructing near infrared from color.

Aside from my research on methodology I am working through collaboration on a number of applications. First, collaborations where the method I developed is used, but more generally on any methodologies involving or requiring statistics, Machine Learning, algorithmic or intensive computation (cloud/cluster computation). From a dataset and application point of view, most of my collaborations are currently focusing on the use of remote sensed data (satellite, drone,…) in the context of quantifying changes (vegetation, land cover, snow cover, glaciers… ) in particular using Google Earth Engine to target small to large-scale studies in time (from a single moment to a complete archive) and space (from a single glacier up to worldwide studies).

BIOGRAPHY

After acquiring a strong foundation in mathematics and physics in “Classes Préparatoires aux Grandes Ecoles”, I completed 3 years of engineering at the Mining Engineering School of Alès, with specialization in computing (equivalent to a M.Sc.). My strong knowledge in computing and in particular in software optimization and parallelization were recognized through the Intel® Modern Code Developer Challenge 2015, where I ended Grand Prize Winner. Because of my interdisciplinary scientific fascinations and my quest for new challenges, I did my Ph.D. in earth science at the University of Lausanne in January 2016, that I finished in January 2020. Since I have the opportunities to collaborate with the IGD ( Institute of Geography ) at UNIL , on a project to quantify land cover change in Vietnam over the last 50 years.

Recently, award of SNSF early postdoc mobility, my next stop would be in Stanford University to work on Machine Learning applied for stochastic geospatial simulations.

CODE

The QuickSampling code is freely and openly available for Linux / macOS / Windows 10 and is usable form Matlab / Python3 and R on the dedicated GitHub repository. Please check the documentation for installation and basic usage.

PUBLICATIONS

PRESENTATIONS

Poster presentation

Invited presentation / lectures