Grounded in light: photonics and agriculture through the eyes of Gerrit Polder

Once every two months, we introduce you to a specialist. We give an insight into the person, their research and their expectations. This time, we talk to Gerrit Polder, senior research associate  Image Analysis and Machine Vision of Vision+Robotics at WUR Greenhouse Horticulture and the Agricultural Biosystems Engineering chair group. About the new world of Terahertz and Photonic Integrated Circuits, and how Raman spectroscopy and quantum dots will take agriculture by storm.

“In simple terms, photonics is just about manipulating photons, particles of light,” says Gerrit Polder. “A bit like processing electrons in electronics.” Another important element of photonics is the interaction between photons and electrons, he says. For example, you can convert electric current into photons in lasers, and convert photons into electric current in optical sensors, think of the camera in your smartphone. “Photonics is playing a growing role when it comes to more efficient, advanced and sustainable agriculture.” He gives some examples: “Multispectral cameras under a drone to monitor the growth and productivity of a potato or other crop; 3D cameras on harvesting robots directing grabs to harvest-ready apples. Or lasers that burn away weeds in the field.”

Tinkering with transistors

Gerrit Polder is senior research associate for Image Analysis and Machine Vision at the WUR business unit Greenhouse Horticulture and recently also in the Agricultural Biosystems Engineering chair group, where he supervises Master’s students and PhD candidates. His background: HTS Electrical Engineering at the university of applied sciences in Arnhem. No one was surprised that he chose to study engineering after secondary school. Even at primary school, he had been interested in electronics, radio engineering and things like that. He was one of those lads who takes transistor radios apart only to put them back together again. Or he would use parts from discarded appliances to make something else. “I still love tinkering with receivers and robots. First with my children, now with my grandchildren.”

During his HTS studies, he did an internship in the Department of Nature and Meteorology at Wageningen University & Research, then called ‘Landbouwhogeschool’. Nice and close to his hometown of Veenendaal, unlike his second internship in Indonesia, where he repaired aircraft radios for a missionary organisation. For his thesis research, he returned to Wageningen, to the same department. “My thesis was about automating agricultural machinery. So, the distant predecessor of precision agriculture, now an important theme within Vision+Robotics.”

Interface cards and analogue cameras

Just before graduating, Polder heard about a vacancy in the department of molecular physics at the Agricultural University and liked the idea. “Lasers and NMR. Magnetic resonance with large helium-cooled magnets. All very interesting.” His job was to ensure that all the equipment in the lab was working and give students a practical course. After a few years, however, he decided he needed a change. He ended up in the world of telecom. But that wasn’t for him. So he returned – over 30 years ago – to Wageningen. To support breeders’ rights research, where hundreds of measurements had to be done manually for new crops, using image processing.. A novelty at a time when cameras weren’t yet digital. Polder: “We used frame grabbers, interface cards in the computer that converted the analogue video signal from a video camera into a digital image.”

The University of Agriculture was a pioneer when it came to applying spectral image processing in agriculture in the late 1990s. As soon as the first imaging spectrographs hit the market, they were keen to start research on ‘spectral imaging for measuring bio chemicals in plant material. In collaboration with Delft University of Technology, funded by what was then the Technology Foundation STW. A great job for a PhD candidate. Polder put his hand up and said he would like to be that PhD candidate. This took him into ‘the next step’ in photonics. From black-and-white and colour images to spectral images. “I still have those first spectrographs in the lab’s attic. A piece of history. I should actually display them in a showcase.”

A lab setup with Gerrit Polder’s spectrographs in 2004

You should think of a spectrograph as a device that uses a prism to split light into different wavelengths. “You can use it to create a spectral image with a complete reflection spectrum on each image element,” says Polder. “A rainbow of colours corresponding to different wavelengths of light. When you analyse this spectrum, you get information about the physical and chemical properties of the object or material you capture with this camera.”

Lycopene in tomatoes

During his PhD, Polder worked on measuring lycopene in tomatoes. A powerful antioxidant, and one of the ingredients in tomatoes alongside beta carotene, chlorophyll, vitamins and minerals, among others. The spectrograph allowed him to trace lycopene pixel by pixel. Even more interesting: “You can use that camera to pinpoint where there are abnormalities within a tomato and other crops.” He obtained his doctorate from Delft University of Technology in 2004. The publications in his thesis are among the first on the application of spectral image processing in agriculture. Proud: “In particular, the article ‘Spectral image analysis for measuring ripeness of tomatoes’, published in ‘Transactions of the ASAE has been cited very often.”

Integrated Pest Management

Since his PhD graduation in 2004, Polder has mainly focused on spectral image processing. Particularly in relation to disease detection. “There are various diseases that are difficult to detect at an early stage. The symptoms are not yet visible, but you do want to identify them.” With spectral image processing or spectroscopy, you can see more than with the human eye. Which is great because the sooner you detect a disease, the sooner you can intervene. He mentions Integrated Pest Management (IPM). A hot topic. Farmers are becoming more and more limited in the number of pesticides they may use. But they do want to keep control over their crops. To prevent a possible plague and thus the failure of their harvest. Polder: “By spraying very specifically, only on that spot where the disease is, you use fewer pesticides. Which is better for the environment and saves money. And the sooner you spray, the sooner the disease is eradicated.”

Polder refers to the EU’s OPTIMA project: integrated crop protection, holistic and environmentally friendly. This is based on a smart camera system that allows you to find different diseases in a crop at an early stage. By doing so, you can significantly reduce the use of crop protection agents. “OPTIMA consists of a combination of machine learning, colour cameras and spectral cameras. We started by collecting images of diseases in the field. Downy mildew in vineyards in Piedmont in Italy, scab in apple orchards in Epila, Spain, and Alternaria in carrots in the Bordeaux region. Based on these images, we used deep learning to train a disease detection model. When you present an image to that algorithm, it can immediately establish whether there is a disease and where it is in that image.”

Gerrit Polder and colleagues during a field trial for the OPTIMA project

The camera with that disease detection system is installed on a tractor equipped with GPS and a 4G antenna. This crawls through the field recording images in rapid succession. Polder: “We know where the camera is when it takes an image. It’s a smart camera, so it doesn’t just take the images but also processes them.” The decision whether to spray or not is made by a decision support system in the cloud. “Based on that information, you can create a task map that you link to a field sprayer. This enables you to use the right amount of crop protection agents in the right place and at the right time.”

The more spectral colours, the higher the price

To record the spectral images in the field, line scan cameras are used. These are awkward in windy conditions because then the leaves move, making scanning tricky. A spectral snapshot camera is a great solution, says Polder. “Multi-spectral cameras with a couple of bands have been on the market for some time, but you are also seeing more and more snapshot cameras with high spectral resolution. These allow you to capture multiple spectral bands in a single click. Other spectral cameras need several shots or scans to do that.”

The problem with all these fancy spectral cameras: they are expensive. Polder: “That first spectrograph I used several years ago is still about the same price.” Spectral cameras simply don’t sell very well. That makes them expensive. Starting price 20,000 euros. “The more spectral colours the camera can detect – that is, the more chemicals, substances and components you can distinguish in an image – the higher the price.”

Multispectral and InGaAs

A lot of research focuses on finding wavelengths. Which camera do you need for which task in this process? Polder: “From a high spectral resolution camera with hundreds of wavelength bands, you go back to a single-band multispectral camera that works cheaper and faster.” Many of these cameras operate in the visible and near infrared region (400-1000 nm). Silicon is the base material of the sensor in those cameras and it works well up to 1000 nm. Above that wavelength, you’re in the short-wave infrared region (SWIR, 900-1700 nm), which is very interesting for many applications. Silicon no longer works here, so you need other material. He mentions the InGaAs image sensor, which stands for Indium Gallium Arsenide. And no, at three or four times the price of silicon-based cameras, it certainly isn’t a bargain.

Quantum dots

Nevertheless, Polder sees hopeful developments here, also in terms of price. He talks about quantum dots. Tiny semiconductor particles, usually only a few nanometres in size, that can convert light with very high precision. “You can tune these to absorb and emit specific wavelengths of light and so greatly improve the sensitivity and range of silicon sensors. Quantum dots therefore offer a cost-efficient and alternative technology for SWIR spectroscopy.”

Photonic integrated circuits

Polder is interested in developments relating to photonic integrated circuits. Lasers, sensors, spectrographs and other components which you can bake cheaply into a chip. He cites Raman spectroscopy as an example. Polder: “In this process, you shine a laser on a material. Measure how the light spectrum changes as it passes through, which teaches you all sorts of things about the composition of that material.” In his department, it was used on chrysanthemum leaves. To examine them for the presence of mildew. “With this technique, we were able to properly quantify the difference between healthy and diseased leaves.” Raman spectroscopes are currently still bulky and expensive, mainly because they require sensitive sensors and lasers. In the future, you’ll be able to integrate Raman spectroscopy as Photonic Integrated Circuit (PIC) on a single chip, says Polder. A laser together with a sensor. “You can then make a chip that can measure a Raman signal in one go. It’s not there yet, but potentially, Raman spectroscopes could be produced smaller and cheaper with these PICs.”

Terahertz

At WUR, together with OnePlanet and Eindhoven University of Technology (TU/e), they are testing new techniques, in the field of Terahertz (THz). That’s the area between radio and optics, between microwaves and infrared radiation. An empty area that had not been explored because it was not possible to make or detect signals here. Polder: “With radio signals, you talk about Megahertz and Gigahertz. This is followed by Terahertz. If you go even higher, you’re dealing with light frequencies. That’s tens if not hundreds of Terahertz.”

Researchers have found it hard to open the door to Terahertz. Neither with optical nor with radio techniques. That’s changed. Polder: “From radio, they can multiply that Gigahertz signal several more times to get there. Via the other optical side, it can also be done with photonics. With a laser and ZnSe or GaSe crystal.” Photonic- or radio-powered, Polder doesn’t know which device will be launched on the market to measure Terahertz. At WUR, they are not waiting for that to happen and are already looking at how to apply Terahertz in agriculture. “You can use it to examine the amount of water on the surface of leaves, which in turn says something about susceptibility to fungal diseases.” With infrared spectroscopy, you can also measure the presence of water on leaves, but the presence of sunlight and infrared radiation in grow lights disrupts those measurements. There is no Terahertz in sunlight and grow lights. So if you use them to take measurements, your research won’t be disrupted. That could be very interesting.” He doesn’t know what else you will be able to measure with Terahertz. “But it’s important to get a feel for that technique now.”

Small, smaller, smallest

He doesn’t like crystal ball gazing. Technical developments are moving so fast that it’s hard to predict the future. Nevertheless, Polder expects more and more photonics techniques in sensors and we really will get those cheap spectral cameras. Small, smaller, smallest is his motto. “Look at the first GPS receiver, which was a bulky device. Now it’s a mini chip.” And yes, there’s Artificial Intelligence (AI), of course. “For the plant breeders’ rights research, we used to write our own algorithms to determine features such as the tip of a bean or the curvature of a carrot. Today, modern image processing relies heavily on AI where you train deep neural networks to measure objects. “The results of these neural networks are certainly impressive, but it’s important to use common sense with AI.”

At odds with information theory

He sees publications where writers claim you can generate spectral images from RGB colour images using generative AI. Fine, but what’s the point, Polder wonders. Moreover, such a step is at odds with information theory, he says. “You can’t generate information from nothing or generate high-dimensional information from low-dimensional data.” An RGB image consists of three bands (red, green and blue), a spectral image consists of more bands, each representing a specific wavelength of light. A spectral image therefore contains more information than a standard RGB image. When you try to generate a spectral image from an RGB image using generative AI, you’re trying to create additional information that does not exist in the original RGB image. This means that the spectral image generated probably does not correspond with the actual reflectance spectrum of the object in question. So how reliable and useful is such AI research then? Besides, what good does it do? “It’s much harder to visualise a spectral image than a colour image, and for classifying objects, this seems like a needless detour.” Polder prefers to keep both feet on the ground. That’s the only way you can really make a difference, he feels. “When you see something useful emerge after years of research to which you have contributed, that’s always special.”

Senior research associate  Image Analysis and Machine Vision