The successful candidates will be based in the Universitat de València, Spain. Prof. Camps-Valls is the coordinator of the Image and Signal Processing (ISP) group. The group is devoted to the development of machine learning and signal processing techniques for remote sensing image processing, Earth observation data analysis and the Geosciences. Several topics are treated in our research group and projects: regression, causality & information theory, Earth observation data analysis, physics-aware machine learning, generative modeling, eXplainable AI and feature ranking, and anomaly detection. Applications in Earth and Climate sciences fields.
We are looking for several Phd students and postdocs to work in the intersection of Machine Learning and Earth/Climate Sciences. We are searching for outstanding, highly motivated students with a Master/PhD in computer science, statistics, machine learning, electrical engineering, physics, or mathematics. Good programming skills, a critical and organized sense for data analysis, as well as maturity and commitment, strong communication, presentation and writing skills are a big plus.
Positions are open in the context of the following research projects:
Marie Curie ETN iMIRACLI (imiracli.eu) — a European excellence network for PhD training [3 PhD students]
This is an excellence training network in europe around ML for cloud-aerosol interactions. The network aims to bring together leading climate and machine learning experts across Europe to train a new next generation of climate data scientists. The project to be developed at the ISP is described in detail in ESR14, but the ISP group will also co-supervise ESR1 and ESR2. Note: The official application should be done through the iMIRACLI site
XAI4EO — eXplainable AI for Earth observation data understanding [1 postdoc, 1 PhD]
Blink, a new deep network is applied in Earth sciences. But, what has the networked learned? Why? and for maximizing what? This project aims to develop new data analysis techniques in mathematics and computer sciences to explore the solution space of deep architectures, with special focus on Earth observation problems. We will explore invariances in latent spaces, deep causal methods and basis visualization. Experience in deep learning, signal and image processing, color vision, wavelets and proficiency in maths are required.
COSMO — Causality for Soil Moisture Observations [1 postdoc, 1 PhD]
The impact of global warming on the water cycle still remains uncertain. One of the relevant and tractable tracers of water is soil moisture. Soil moisture, acts as the memory of precipitation; it governs infiltration, runoff and evapotranspiration and drives the surface atmosphere exchanges. Thanks to advances in instrumentation, space technology and inversion techniques in the microwave L-band frequencies, the Earth's surface soil moisture is now mapped globally for the first time. These measurements can also be used to infer other key variables such as vegetation water content, sea surface salinity, and thin sea ice thickness. The project has two main goals: 1) improve the spatio-temporal resolution of SM and VOD by spatio-temporal deep networks (convolutional to extract spatial patterns, recurrent to account for memory effects) by multisensor fusion. The new products will allow us to study memory effects of soils and impacts on droughts; and 2) infer causal relationships from this new observational data set to get a further understanding on the Earth's water cycle and its recent changes. Causal inference from empirical data allows going beyond simple association links and tries to determine relations of causes and effects between variables and observational data. Experience in deep learning and remote sensing data processing are required.
CAUSEME — A web platform for causal discovery method evaluation [1 PhD, 1 software engineer]
EMULA2 — Physics-aware Gaussian processes for advanced emulation [1 postdoc, 1 PhD]
Machine learning models are widely used to learn both the forward and inverse functions, and now routinely replace complex models and sub-components to improve scalability and mathematical tractability. These models are commonly known as emulators and report excellent accuracy-speedup trade-offs compared to simulators, besides elegant ways to do uncertainty quantification, error propagation, and sensitivity analysis. In this project we will focus on probabilistic machine learning models, and in particular on Gaussian processes (GPs), which have excelled in both prediction and emulation. As any other data-driven technique, however, they do not necessarily respect physical or causal relations. Furthermore, GP models are still computationally costly and often yield predictions that are inconsistent with physics principles. Combining model simulations and observational data is an interesting way, and has been recently approached with joint GPs and multifidelity GPs. Nevertheless, advances are still needed in terms of efficiency, physical plausibility, and interpretability. Experience in Gaussian processes, Bayesian inference, and remote sensing are required.
AWILD — Anomalies in the wild [1 postdoc, 1 PhD]
The Earth is a complex dynamic network system and in the last few hundred years human activities have precipitated enormous changes in the Planet. It goes without saying that the most important challenge for today's Science is to detect and attribute the causes of such changes. In this scenario, Earth observation data allows us to automatically detect anomalous changes, as well as extreme events, on the land-cover at both spatial and time domains. This is now possible by exploiting high resolution satellite images and long time series of images and Earth observation products, along with powerful statistical techniques to process them. However, in recent years, the big and heterogeneous data streams acquired by satellite constellations hamper the adoption of advanced machine learning statistical techniques for anomaly change detection and extreme events identification. This project proposes to develop, characterize, and apply novel anomaly detectors, under the framework of Gaussianization networks, Variational Autoencoders and Normalizing flows, to exploit the wealth of information contained in these data. The goal is to develop online detection algorithms, to accommodate multi-source data, model complex distributions, to cope with high-dimensional data, and deal with unevenly sampled time series and missing data. Detection of rare, unexpected changes and events under the developed statistical framework will constitute the stepping stone before the more ambitious far-end goal of machine attribution of anthropogenic climate change causes. Experience in deep neural networks, Earth and Climate sciences are required.
HighRes - High spatio-temporal resolution products with AI in the cloud [1 postdoc, 1 PhD]
The project will develop new deep learning techniques exploiting large spatio-temporal data in the Google Earth Engine (GEE) platform to provide high-resolution cloud-free composiste. These new reflectance gap free data are combined with machine learning models to create robust remote sensing based land information products at a finer continental/global spatial detail. The new land products (e.g. LAI, FAPAR, Land cover, biomass, carbon/water fluxes, etc.) are of high interest for the scientific community, governmental organizations, and the private sector due to the broad spectrum of potential applications possible: from high precision agriculture and crop and/or forest monitoring, to crop yield prediction, and land use change assessment. They endorse an enhanced understanding of biodiversity and nature of the Earth system with a high level of spatial detail.
ERC Consolidator grant SEDAL — a European Research Grant to advance AI in Earth observation [1 postdoc, 1PhD]
We are looking for outstanding postdoc candidates with a strong interest in machine learning and geosciences to cover a post-doc position in the Image and Signal Processing (ISP) group in the Universitat de Valencia, Spain, http://isp.uv.es. This particular position is focused on learning causal models to explain complex interactions in essential climate variables and remote sensing observations, and discover hidden essential drivers and confounding factors in Climate/Geo Sciences applications.