Kristof van Tricht by @kokosiki_babosiki_OneSoil Blog

"The challenge is to embrace new technologies while not losing touch with the real world." Interview with Belgium researcher Kristof van Tricht

To understand global trends, we like to talk to our colleagues from different organizations. This time we turned to Kristof van Tricht, R&D professional from VITO, and independent Flemish research organization in the area of cleantech and sustainable development. What does the life of the European institute look like? What is the future of satellite images? And why everyone hunts for ground truth data?

"The challenge is to embrace new technologies while not losing touch with the real world." Interview with Belgium researcher Kristof van Tricht

To understand global trends, we like to talk to our colleagues from different organizations. This time we turned to Kristof van Tricht, R&D professional from VITO, and independent Flemish research organization in the area of cleantech and sustainable development. What does the life of the European institute look like? What is the future of satellite images? And why everyone hunts for ground truth data?

How to get involved with remote sensing

– Where does your interest for remote sensing come from?

– I've always been very interested in everything that happens in the world. I always wondered as a child where snow was coming from or how a little seed could turn into a living tree. So wonderful! I studied bioscience engineering; I did my Master's thesis using a hyperspectral sensor from space trying to detect invasive tree species in the Hawaiian rainforest. For me that was the first case when I touched on remote sensing imagery—and I was truly amazed.
Photo of Kristof van Tricht_OneSoil Blog
After that, I switched a bit and did a PhD on the role of clouds on the Greenland ice sheet. That was the first time I used active sensors, a lidar and radar from space. I characterized the cloud properties and how they interacted with the ice sheet. I learned a lot during those 4 years. We actually had really scary conclusions that those clouds were increasing the ice sheet meltwater by an extra 1/3 as opposed to clear-sky conditions. After my PhD, I decided that I wanted to leave academia but I still wanted to do something with remote sensing. And then I saw this vacancy at VITO Remote Sensing. I joined the team in 2016.

I'm very passionate about cryosphere, snow, ice and climate, but I felt that I was constantly ringing a bell without being able to make a difference. I thought that in agriculture I could really make a difference by improving yields and giving information to farmers. This is a more practical use for remote sensing. I still miss the cryosphere, but I feel that we're really contributing to the food security challenges in my current job.
I thought that in agriculture I could really make a difference by improving yields and giving information to farmers.

Life at a research institution

– How is VITO organised?

– VITO has around 800 employees. They are organized in different units and departments. One of the departments is Land Use and the two units addressing this topic are the Remote Sensing unit and the Spatial Modelling unit. At VITO Remote Sensing, we have an operations team that does the processing of the incoming Proba-V satellite data (operated under the authority of ESA—OneSoil), a technology team exploring new satellite missions, and an applications team. This last team focuses on applications for different themes; biodiversity, water and coast, and agricultural. I contribute to the agricultural applications team. Currently, I do automatic parcel delineation and crop type identification with deep learning, and I'm also involved in a bunch of other projects.

We have a horizontal structure. In our team we have one supervisor who is mostly responsible for networking and finding opportunities for projects. And all the others are involved in several projects: some of us do project management and some focus on the research. For example, a few people are in Spain now for a project in which we are trying to detect the red palm weevil. We want to detect individual trees on commercial Pleiades imagery, and in the pilot study we want to see if you can use these individually detected trees to upscale them to Sentinel-2.

– What types of projects do you do as a research institute?

– We have some big operational projects running, like the Copernicus Global Land Service or the Climate Change Service, Horizon 2020 projects, ESA (European Space Agency — OneSoil) projects or projects for FAO (Food and Agriculture Organization of the UN — OneSoil), like FRAME which is contributing to the WaPOR portal. A considerable part of the research we do with the applications team, we do on a global scale. Also, we have a few smaller commercial projects. One of them is the 'watchITgrow' online platform for Belgian farmers. One of the ambitions is to expand the commercial services, as they teach you what are the real needs of companies and society. However, our research projects will always remain our first activity.
Life at VITO
– How does the platform work? Do you plan to implement this platform in other countries or will it remain Belgium-oriented?

– Farmers can select or draw their fields, they can monitor their fields and enter other field information. Farmers get the images from Sentinel-2, which are screened for clouds, and then they get the curve which tells them whether their crop is in good condition or not. This curve in the background is combining optical and radar information, so it's complex to get to it. But we feel that you always need to combine technologies and data into easy to understand information.

'WatchITgrow' is a free-to-use platform for the farmers, and since this year they even get an incentive to upload their field data on the platform. This will allow us to improve the platform even more.

We already have it running in parts of Italy, we have some trials in the Netherlands and in Africa. Africa has always had our attention; we have lots of experience there. So we start from the well-known regions and then we transfer the technology to the more difficult regions where we need to retrain and improve.
Our basic idea is that a farmer doesn't want to know about clouds, deep learning, radars, optical—all he wants to know is how his field is doing.
– It resembles the way OneSoil does things quite a bit.

– From the moment the OneSoil web-platform was online, I thought "We need to see what they're doing and talk to them". I see our companies as two different worlds which need to be working closely together to move forward. I know that some other companies and institutes look at it differently, they might be a bit suspicious about small startups processing the whole world with seemingly limited efforts. But if you're suspicious about emerging technologies and startups, you quickly risk to lag behind. We should both learn from each other and some technologies from one world could go to the other.

One big difference I see between OneSoil and VITO is that you have a rather small team with a clear focus and the need to implement everything in the cloud. We have a very big team—only in the remote sensing unit we have over 100 people, and we have our own data and processing center. As a consequence, many different projects impose some challenges in getting a shared focus as you would have in a small team.
PROBA-V is the satellite for global vegetation monitoring, where VITO is responsible for the incoming data processing and distribution
Another difference I see is that we also work on the preprocessing of data and integration of datasets to make them radiometrically and geometrically consistent over time. We have many scientists who do extremely good research work. They thoroughly calibrate and validate algorithms, which of course takes considerable time, and this is something that I feel people tend to do less these days.

– Wait, why do people tend not to calibrate and validate algorithms?

— Nowadays, there are countless possibilities to quickly run algorithms on many different platforms with huge amounts of data sources. We all tend to apply these algorithms to various domains, without proper understanding, calibration and validation. It allows producing output quickly, but there is definitely a risk if we lose a sense of reality and forget to question the results. I see it as one of the biggest current challenges in our domain to embrace new technologies that make anything possible, while not losing touch with the real world.

– Coming back to 'watchITgrow', is your aim to earn money?

VITO is a non-profit organisation. But sometimes the research that we do results into spinoffs and the new company can start earning money at some point. Let's put it like this, we aim with our research to have an impact on a sustainable future and once it is clear how the results of our research activities are commercial viable—we transfer it.

In the search for ground truth

– Do you experience problems due to a lack of ground truth data? Is it an issue for you?

– This is a huge problem. I see the tendency in projects to spend more money to apply new technologies leaving less money available to collect field data.

As an example we've been using data from LPIS for a few years now. There is a common agricultural policy in Europe that all the member states collect field boundaries and crop type info. For some countries it is already open, in others it is supposed to become completely open in the future. This is a very rich data source against which to calibrate and validate algorithms and it would be great to have it all publically available. Also data collected in publicly funded research projects should be freely available to train and test algorithms. This would help solving the problem we are facing in lack of data in e.g. developing countries.

So we're doing fine for crop mapping in the Western countries, whereas for Africa ground truth data is basically non-existent. There, you either need to put 20 students in a digitizing exercise, or you need to talk to local agencies. But then there is the risk of getting this data in a format that you can't understand. And we're only speaking about crop type mapping, we're not even talking about crop condition mapping or yield forecasting.
Fields on the Sentinel-1 image_OneSoil Blog
Fields on the Sentinel-1 image
In the end, everyone is interested in what the yield of the field will be. Here I think that emerging technologies like deep learning can definitely help, but models that do yield predictions from scratch are still in the future because we don't have enough training data. Yield forecasting that doesn't need thorough region and crop specific calibration and therefore loads of data… I think it's not possible yet.

So for me, ground truth is still the holy grail for remote sensing.

– I see. Let's talk about modern technologies that you like using.

– I'm a fan of Google Earth Engine. I think it's safe to consider them as one of the pioneers in enabling remote sensing data mining at large scale without all the preprocessing hassle. They ingest the data from all these different sensors, so you don't have to take into account different projections, pixel sizes, etc. It's really awesome—you can quickly develop and test things with it, but it's also risky because it makes people a bit lazy. You just do stuff and don't take the time to try and understand it. We should use it wisely!

The other thing is the growing deep learning universe, which I think is essential to cope with the growing amounts of data. The times of using one single data source, or one simple regression model to translate from one source into another—those days are coming to an end. One reason is the ever growing availability of free data sources urging us to look beyond one single satellite sensor to create novel applications. The other reason is that there's just too much data because the resolution gets better all the time, so we can't handle it anymore.
Ground truth is still the holy grail for remote sensing.
Let's say I want to do global analyses with 10m Sentinel-1 and Sentinel-2 data. If I don't do prior data reduction, for example by calculating specific metrics, there will be too much data for our conventional algorithms. However, if I do data reduction, I take the risk of losing important information that I'm not aware of. Deep learning models are a good answer to this dilemma. They can handle huge amounts of data, looking for hidden patterns that help building your application.

In fact, it's one of the core research tracks I'm working on at VITO Remote Sensing. I'm building fused products that take advantage of the combined strength of optical Sentinel-2 data and the cloud-penetrating capacities of radar Sentinel-1 data. At first, I used a conventional random forest classification approach to create a Belgian crop type map using Sentinel-1 and Sentinel-2 information. Didn't work that bad, but I had to make important data reduction choices.

More recent efforts based on deep learning allowed me to automatically delineate the parcels and recognize the crop types using combined Sentinel-1 and Sentinel-2 time series without those prior data reduction choices. And the best thing: it turned out to give even much better results. With deep learning, we now even accomplished fusion at the product level itself.
A piece of cropmap for Belgium_OneSoil Blog
A piece of Belgium cropmap made by VITO
With our CropSAR technology, we create Sentinel-2 like time series of vegetation parameters which are free of clouds, using the original Sentinel-2 and Sentinel-1 data in a deep learning framework. The technology is operational on field level in the ' watchItgrow' platform (as a cloud-free 'greenness' curve), while we're currently working to develop it for images as well.

That's for me a key thing: we create real actionable information like one simple greenness curve which is easy to understand, and in the background this is accomplished by throwing together different data sources in a well trained deep learning framework. Creating simple information in a chaos of data!
But all of this with the same side note I mentioned earlier: you really need to understand what goes into such models and what needs to go out, no matter the application. Otherwise there's the risk of losing feeling with the real physical phenomena. As long as we pay attention to that culprit, deep learning lets us explore new grounds for sure!

Visions for the future

– Our R&D team asked me to discuss some questions with you. So there are publicly available satellite images out there and ones you have to pay for, provided by commercial companies. How do you think this will change in the future?

– Good question! It used to be all paid data, now if you want to go beyond 5 meters resolution there's a tendency for data to become and stay free. I think it is a good concept to offer a lower resolution for free, and if you want higher resolution, you'd have to pay. If you go below 5 meters, then this is commercial. I think it is impossible to get all the data only from publically funded projects. For instance, Copernicus cannot launch so many satellites to cover the high temporal frequencies, high spatial resolution data, etc. The commercial companies will always fill a gap in the free data and this is a healthy thing. I don't see this as an issue if you are able to come to an agreement with such companies. So I don't think that in 10 years we will not have commercial missions anymore or vice versa.

– Got it. And the last question is about hyperspectral images—do you use them now and how do you see their future?

At VITO, we have been working with hyperspectral for several decades. What's interesting about the hyperspectral technology is, for example, the ability to identify individual pigments of vegetation to study health conditions more closely. Or before the start of the growing season you'd be able to detect any shortage of specific components in the soil so you can advise a farmer to apply specific fertilizers. At the moment, however, for many agricultural applications the resolution of the sensors is an issue, hence we tend to work with drone data in the hyperspectral domain.
I think thermal is the next big thing that we will see for vegetation and agriculture.
Another thing worth mentioning for the future is thermal imaging. In agriculture, thermal imaging would allow to detect plant stress in early stages, enabling the farmer to take immediate measures before real damage occurs. Currently, however, thermal sensors are facing similar problems as the hyperspectral data, i.e. a resolution problem. Sentinel-2 doesn't have a thermal band, but there will be new thermal missions in the future with improved spatial resolution.

I've seen a nice presentation at a conference not long ago where combining images from thermal sensors with the higher resolution products of Sentinel-2 already provided some very nice results. Imagine how those results would improve if we had thermal information at the same resolution as Sentinel-2! I think thermal is the next big thing that we will see for vegetation and agriculture.
Olga Polevikova, the author_OneSoil Blog
Olga Polevikova
The text
Nastya Vishnevskaya, the illustrator_OneSoil Blog
The illustration
Do you like this post?
Related Articles