Combining the generalisability and accuracy of machine learning with physical knowledge to constrain the solution space and improve the interpretability of remote sensing data.
river, grass, field, tree, nature, dry, sky, landscape, blue, water, forest, meadow, cloud, yellow, beautiful, rural, scene, summer, season, park, scenic, lake, green, day, travel, autumn, outdoors, sun, view, background, land, spring, beauty, environment, horizon, sunlight, country, outdoor, clouds, white, road, branch, hill, mountain, countryside, panorama, fall, plant, valley, winter
Helmholtz AI project call showcase: Improving remote sensing data through Deep Learning
This weeks’ Helmholtz AI project showcase is a collaboration between the German Aerospace Center (DLR), Alfred Wegener Institute (AWI), and Helmholtz Centre for Environmental Research (UFZ). DeepSAR will improve remote sensing data through Deep Learning. Read more to find out.
Could you introduce yourself, giving your affiliation, area of work, and of course, the project title?
My name is Ronny Hänsch. I am the leader of the Machine Learning team in the SAR Technology department of the German Aerospace Center (DLR). My main research interest is the automatic interpretation of remote sensing imagery via machine learning, i.e. the intersection of computer vision and machine learning with Earth observation and remote sensing. Our project "DeepSAR: Improving remote sensing data through Deep Learning" lies at the centre of this intersection.
In simple words, what specifically is your project about? And, how and why do you think it is a high risk, high gain endeavour?
Currently, there are two main streams of work in Earth observation and remote sensing: On the one hand, traditional approaches that use increasingly complex models to try to invert the imaging process to derive geo-/bio-physical parameters. These models have a solid foundation in physics and statistics which often makes them interpretable. However, they require strong domain-specific expert knowledge and depend on many parameters that are often unknown and scene-specific. On the other hand, data-driven machine learning aims to find a direct mapping from the observation to the target variable. Since these models are general function approximators, they often lack both interpretability and clear physical meaning. However, modern approaches are extremely successful and lead to very accurate results. Our DeepSAR project is about merging both worlds to derive hybrid models that maintain the generalisation capabilities and accuracy of machine learning while including physical knowledge to constrain the solution space and increase interpretability.
How important has the Helmholtz AI funding and platform been to carry out this project?
Helmholtz AI is the optimal funding for this project, since it explicitly requires the collaboration of multiple Helmholtz Centers. For DeepSAR, this opened up the opportunity to bring experts from two different application domains, i.e. forest and ice, into the project.
Figure: The image shows an artistic transition from a polarimetric Pauli RGB image (Red: HH-VV, Green: 2HV, Blue: HH+VV) at L band to a forest height map obtained from the inversion of L band Pol-InSAR data over the Pongara national park in Gabon. The national park consists mainly of mangrove forests and contains some of the tallest mangrove populations in the world towering up to 60 meters.