Machine learning techniques to help better understand plant health.
Helmholtz AI project call showcase: ML-based Fluorescence Retrieval from Optical Satellite Data
In the second Helmholtz AI project showcase, we look at how machine learning techniques can help better understand plant health. The work is being carried out by Miguel Figueiredo Vaz Pato from the German Aerospace Center (DLR) in collaboration with Forschungszentrum Jülich (FZJ).
Could you introduce yourself, giving your affiliation, area of work, and of course, the project title?
My name is Miguel Pato and I am a research scientist at the Remote Sensing Technology Institute, German Aerospace Center (DLR) in Oberpfaffenhofen, Germany. I am part of the team dedicated to hyperspectral remote sensing (aka, imaging spectroscopy), where my work focuses on the processing of data from Earth observation satellites and the development of machine learning (ML) algorithms for the use of hyperspectral data in different applications. Our Helmholtz AI project is entitled “ML-based Fluorescence Retrieval from Optical Satellite Data” (abbreviated FluoMap) and is a joint effort between a team at DLR and a team at Forschungszentrum Jülich (FZJ) from the Institute of Bio- and Geosciences (IBG-2, Plant Sciences) and the Institute for Advanced Simulation (IAS-8, Data Analytics and Machine Learning).
In simple words, what specifically is your project about? And, how and why do you think it is a high risk, high gain endeavour?
Plants convert sunlight into chemical energy through the complex process of photosynthesis. One of the by-products of photosynthesis is the re-emission of part of the absorbed light as a fluorescence glow at red and near-infrared wavelengths. This fluorescence signal is an important clue to the internal health of the plant. In our project FluoMap, we propose to develop machine learning algorithms to derive fluorescence maps directly from hyperspectral data collected from space with the DESIS sensor and from the airborne sensor HyPlant. This is particularly challenging due to the ‘smallness’ of the fluorescence signal, which is frequently overshadowed by other signals, and by the need to model precisely the spectral characteristics of the sensor. It is far from certain that a suitable machine learning algorithm for fluorescence retrieval can be devised, which makes the project of high risk. But, in case of success, fluorescence maps will be able to be derived readily and across the globe, which represents a high gain for the monitoring of plant health worldwide.
How important has the Helmholtz AI funding and platform been to carry out this project?
Our project relies on the complementary expertise contributed by the teams at DLR and FZJ, two Helmholtz partners. The Helmholtz AI platform and funding were fundamental to put these two groups in contact and make the carrying out of the project realistic in a short time frame. This is not only crucial for our current joint project, but also opens the way for future collaborations between the two Helmholtz partners.