Multi-modAl Earth obServaTion Image Analysis – MAESTRIA

The MAESTRIA project (Multi-modAl Earth obServaTion Image Analysis) aims to solve the methodological challenges related to the fully automatic analysis of the massive amount of images acquired by Earth Observation platforms. MAESTRIA targets to generate land-cover and land-use
descriptions at country scale at various spatial resolutions and sets of classes. The ultimate goal is to provide a continuum of spatially and semantically consistent products, that are relevant for many end-users and applications. Both public policies at local or national levels and scientific models would
benefit from such kinds of products for climate modelling, urban planning, crop monitoring or impact assessment of surface changes.
The output of the MAESTRIA project will be two-fold: (i) methods that leverage current challenges in Earth Observation image analysis; (ii) a large range of precise and up-to-date land-cover maps available over very large scales from 2m to 100m. Both will be made freely available so as to stimulate
research and commercial services built upon such maps.
Many global and land-cover geodatabases have been established during the last two decades. However, they do not meet the current requirements in terms of semantic and spatial accuracy and updateness. In parallel, a large body of literature has tackled automatic EO data exploitation.
However, most of the existing papers are limited to a specific environment, site or sensor, and a specific need. They are not flexible and not adapted to the new paradigm in EO with the advent of satellite missions with short revisit time and increased spatial resolutions.
What makes the analysis task challenging now is the heterogeneous physical nature of such images. One has now to design adequate methods to optimally exploit the complementary information provided by images acquired from a large variety of sensors. The current situation is exacerbated
when addressing the upscaling issue, i.e., when classifying this amount of images at large scales while trying to guarantee a homogeneous accuracy in all areas of interest.
Three main methodological challenges will be tackled. First, it deals with multiple sensor fusion so as to extract the most meaningful knowledge from the huge amount of heterogeneous data and subsequently its sparse representation. This will ensure to limit the amount of information required to generate a consistent land-cover map in a timely manner, by facilitating and improving the underlying supervised classification task.
Secondly, no very large scale learning techniques have been proposed so far so as to deal with noisy information. Noise may come from the images themselves but also from the input labels. Two main tasks are considered: the design of (i) semi-supervised learning strategies that rely both on unlabelled and labelled samples to alleviate the needs of large/well-balanced and accurate training set, and (ii) efficient optimization procedures over millions of samples with thousands of features.
Thirdly, we aim to develop methods that derive automatically new land-cover products with different spatial and semantic resolutions out of those produced in the two first steps. As a consequence, we target to obtain a continuum of adapted land-cover layers, both in terms of spatial scales (2->50-100m) and semantics. The core of this task relies on multi-scale semantic segmentation that can handle uncertainties and inconsistencies between scales. In addition, we will link our generic national land-cover maps with various end users' needs, both at scientific and institutional levels, and eventually, integrate the developed algorithms into a common open-source framework for further dissemination and data exploitation.
MAESTRIA is embedded into the Theia Land Data Centre initiative: we will benefit from the existing infrastructures, datasets, institutional partners, and first connections with private companies through regional animation networks and booster initiatives.