Rise of AI Conference 2019 · Berlin · Unicorn Stage

ESA AI Demo: How to Apply AI on Earth Data

ESA, LiveEO, Birds.ai

At Rise of AI 2019 in Berlin, a representative of the European Space Agency introduces ESA's free Copernicus Earth-observation data before two startups demo how they apply AI to it. The ESA speaker explains that the Sentinel satellite constellation delivers petabytes of optical, radar and atmospheric data at 10-meter resolution roughly twice a week, free and guaranteed for decades, and describes ESA's challenge programs and plans to run AI directly on satellites. Daniel Seidel, founder and CTO of LiveEO, then shows how his company combines satellite imagery with AI to monitor infrastructure like electricity grids and railway vegetation, using Deutsche Bahn as its founding case before expanding to US and Austrian grids. Anouk Visser, co-founder and CTO of Birds.ai, closes by demonstrating how her company uses drone RGB and thermal imagery with supervised machine learning to detect defects on solar farms and wind turbines, delivered through a subscription platform called Bird's Eye View. A Q&A covers workflow ownership and each startup's path from an earlier, different idea to its current business.

Key takeaways

  • ESA's Copernicus Sentinel satellites provide free optical and radar Earth-observation data at 10-meter resolution roughly twice a week.
  • LiveEO found that vegetation exposure along Deutsche Bahn's rail network ranged from about 20% to 70% depending on the region.
  • LiveEO scaled the same satellite-based monitoring pipeline from a single German customer to infrastructure grids in the US and Austria.
  • Birds.ai combines drone RGB and thermal imagery with supervised machine learning to detect solar panel defects like broken substrings.
  • Birds.ai charges a monthly software-as-a-service fee for platform access plus a fixed fee per new drone-based inspection.

Detailed notes

ESA's free Earth-observation data

The ESA speaker introduces Earth Observation as the agency's use of satellites to monitor the planet, comparing the Copernicus program's Sentinel satellites to a doctor tracking vital signs, covering vegetation health, the ozone hole, clouds and atmospheric chemistry. The data, at 10-meter resolution and refreshed about twice a week in both optical and radar domains, is free, publicly funded and guaranteed for continuity over the next 50 years, alongside a newer commercial layer of small satellites offering daily, higher-resolution but paid imagery. He notes the scale, tens of petabytes, makes machine learning necessary, though a scientist is still needed to interpret results correctly, illustrated by AI-generated radar-to-optical imagery that looks convincing but can misrepresent reality. He describes a project with the World Food Programme and UNICEF that used a drone campaign to map crops in Malawi, then used transfer learning to scale the model from high-resolution drone data down to Sentinel's 10-meter resolution so results could extend from a field to an entire region or continent. ESA also runs challenge programs like FDL to pair Earth-observation and AI specialists, and is moving toward running AI directly on satellites, including a chip-equipped satellite launching in September and a planned pair of satellites that will share computing power over a laser link to decide what data is worth downloading.

LiveEO: monitoring infrastructure grids from space

Daniel Seidel describes LiveEO's mission to monitor what he calls the biggest machine on earth, the US electricity grid, which alone spans around 9 million kilometers, as part of a wider category of infrastructure grids, electricity, rail and pipelines, that together span more than 30 million kilometers globally, about 80 times the distance to the moon. He says this infrastructure is typically inspected manually, often by aerial survey or people walking the grid with paper and pencil. LiveEO combines free Copernicus data with higher-resolution private satellite imagery to classify vegetation risk along infrastructure automatically. Its founding case was Deutsche Bahn, which had suffered more than 800 train collisions with fallen trees over three years and did not know how much of its roughly 33,000 kilometers of track was exposed to vegetation; its original plan to send around a thousand inspectors out over five years was abandoned once LiveEO's satellite-based classification, later scaled to cover all of Germany, showed that vegetation exposure ranged from about 20% to 70% of track length depending on the region, letting the company reallocate budget instead of spreading it evenly. LiveEO has since applied the same pipeline to monitor electricity grids for customers in the US and Austria.

Birds.ai: drone and AI inspection of solar farms

Anouk Visser explains that Birds.ai, working first with the Austrian utility she refers to as Wien Energie, built a platform giving operations and maintenance managers a bird's-eye view of solar farms. Drone operators capture RGB and thermal images at about five-centimeter resolution, since thermal cameras reveal defects, such as hot spots from a broken substring or a faulty junction box, invisible in visible light. Supervised machine learning models, trained with input from inspection experts, detect issues like bird droppings, broken glass, overgrown vegetation, missing panels and thermal hot spots, and bounding boxes locate individual panels to build a blueprint of the farm. The findings are enriched with severity ratings and recommended actions and shown to customers on the Bird's Eye View platform or in PDF reports. The business model combines a monthly software-as-a-service fee for platform access with a fixed fee per new drone-based inspection. Visser says the company inspected 9.7 megawatts of solar capacity, about 40,000 panels, in its first year, and is expanding the same defect-detection approach to wind turbines and exploring uses like coatings on sea vessels.

Q&A highlights

Asked about workflow ownership, Visser says Birds.ai always sets the flight protocol even when customers run their own drones, and that usage of its app by maintenance teams in the field is more frequent than the underlying satellite or drone analysis cycle. Asked how they moved from idea to startup, Visser recounts starting with a challenge to protect game reserves from poachers, then testing several industries such as agriculture before settling on infrastructure inspection. Seidel recalls starting with hardware ideas for small satellite components, realizing LiveEO was about five years too late to build hardware, briefly considering a tourism app to measure beach crowding from space, and pivoting after market analysis toward the infrastructure and utility sector.

Quotes

We are around five years too late if we want to build hardware, so that's why we wanted to build a software solution.

Daniel Seidel

Artificial inspectors are a lot more cheaper than human inspectors.

Anouk Visser

Questions answered in this talk

What satellite data does ESA provide for free for AI applications?

According to the ESA speaker at Rise of AI 2019, the Copernicus program's Sentinel satellites provide free optical and radar Earth-observation data at 10-meter resolution, updated about twice a week, with continuity planned for the next 50 years.

How does LiveEO help Deutsche Bahn manage vegetation risk along railways?

According to Daniel Seidel of LiveEO at Rise of AI 2019, LiveEO combined satellite imagery with AI to classify vegetation along Germany's roughly 33,000 kilometers of track, after Deutsche Bahn had no data on how much track was exposed and had planned to send inspection teams out over five years.

How does Birds.ai detect defects in solar farms?

According to Anouk Visser of Birds.ai at Rise of AI 2019, Birds.ai uses drones to capture RGB and thermal images of solar farms, then applies supervised machine learning trained with inspection experts to detect defects such as bird droppings, broken substrings and hot junction boxes, shown on a platform called Bird's Eye View.

Mentioned

Sentinel satellites, transfer learning, super-resolution, hyperspectral imaging, thermal imagery, bounding boxes, software-as-a-service

Keep exploring

More from Rise of AI Conference 2019

More on Enterprise AI & Adoption

Transcript

Auto-generated transcript, not edited. · 8,130 words

Show the full transcript

good morning everybody my name is Philip I'm working in the European Space Agency this is a huge stuff and today I will only focus on this which is also big our planet because that ISA we do a lot of things from launchers to exploring the universe to transferring technologies but my main focus is what we call Earth Observation this is using satellite to look at our planet and creating information services and also science with this data and this is really the focus all the satellites is just a technology to help monitor and manage this this is our habitat we have only one I'm always starting talks with this picture because I think it's emotional basically and it created when it was taken a wave of emotions people realizing they were on a small planet in the blackness here and that it was limited in terms of resources and you see here the atmosphere the size of it extremely thin it's like a hair on a big football if you want which is enabling us to

live and to protect us from radiation and to breathe etc so it's very fragile and we really need to care properly about it so what we do in ISA is to launch satellite that monitored is on a routine basis and I'm gonna talk about a constellation called Copernicus with satellite called the sentinels and they monitor different parameter of the earth like a doctor is monitoring not only temperature but pressure different what we call the body the body metrics or the vital signs in order to understand the health of the planet and this is not only data there is a whole processing chain behind it to transform this data into information to understand the health of the vegetation the ozone hole the cloud for meteorology the vegetation extension atmospheric chemistry etc so you you can basically see everything from space so what we try to do is to provide you guys with a kind of data set a kind of digital planet on which you can build your own services with your startups and your analytics

and we talked about Big Data we are at the order of petabytes of data so we need techniques like machine learning these days to do this otherwise if you look at each pixel which is 10 meter by 10 meter it would take you modern hundred years to understand what's happening on the planet so we need to make it automatic and we will see example later on making it automatic is also challenging because you don't know what's in the black box so we still need the scientist in the equations I'm afraid to really interpret correctly what's happening and this is the infrastructure we put in space for you this is totally free for you the data it's paid by public money up front and it's totally unique Europe is actually leading on this because we launched a family here of satellites called the Sentinels monitoring the planet a 10 meter resolution twice per week both in the optical domain and the radar domain also looking at atmospheric chemistry with different resolution the new thing as well

is that this thing is likely to stay so you will have continuity of the data for the next 50 years always the same data plus the enhancement of technology so you can invest because one of the issue we had before is that we had research satellite and we didn't know if they will last so people didn't want to really get this into the decision process and with this we get metrology science mission and today the new commercial layer coming up with small satellite which is also disrupting this space providing very high temporal sampling like daily and very high resolution major difference is that you need to pay for it so there is a full synergy of the type of capability and we need to put this data in the context of new datasets coming up like drones IOT etc so for startups it's really an exciting time there is a convergence of technology's a convergence of capabilities in terms of computing open data and a growing demand for this because the earth is really in

a bad state and we need to manage it properly so business corporate society is calling for information services and the trick is of course to transform that data into the information this value chain can be very long it needs full ownership of the user so there is a lot of networking asking the user what he wants so that you tailor your service so that's where the added value is coming and most of this now is to be generated automatically through a machine learning and through a multi variable approach use if you remember the different signs of life the trick is to actually correlate everything's to understand the full health of the earth because people tend to be in silos they look at one aspect so that's why for me that the frontier is is really this integrated view and a is rising very fast in our field for automation of satellites for new services for creativity like this is Sentinel one as seen by van gock this also highlight for me the limitation of AI

because I call it the curse of beauty there are a lot of people now trying to make radar data look like optical data using gun a scheme it looks very good beautiful so people believe it the true it's the truth but when you compare to reality it's different stuff so you have also new issues they're linked with the ability of this to be so realistic okay and we are really in a state of flux and disruption in the space agency because of these new technologies and the convergence with other things such as we reusable launcher from Elon Musk venture capitalist focusing on your stuff space like seven years ago they didn't want to invest too risky then there was a wave in the Silicon Valley and now everybody wants to invest in space they think it's the new frontier not only the big guys like Jeff Bezos in on masks etc but also investor who thinks that you can have a 10x if not hundred ex back because it's it's already paid for by public

money and the marginal value you put on it can have a lot of commercial value this is an example we did with in the lab I'm working in it was a use case with the World Food Programme and UNICEF so we really tried to impact the life of people in the field and children and they wanted to know about the food resources in Africa you know in Malawi so they had a drone campaign because there is no regulation for drones there there is a corrido and you can try everything transport with drones except so they map the crops some farmer identified these grubs we put machine learning to expand this it work well for some crops less for some others discrimination is not always perfect and the key is to extend this to the resolution of the Sentinel which is 10 meter because if you can move the the learning from the field at high res with the drone to the medium res with the satellite you scale up and that's fantastic for them because

they can then optimize campaigns increase representativeness and then scale up at the region nation and continent level by a few clicks so this opens totally new door for them on how to do operation and this is enabled by the machine learning because this transfer learning is not the traditional way we also do challenges in order to get AI guys our big problem is to get AI guys that's why I'm here today they all fly away somewhere else but when you told them there are 15 petabytes of data which is free on the state of the planet weather environment climate all our economy depends on this and we need your help to actually mine this they get excited so that's my main message today is that it's all there free and we need new perspective on it to actually create new services so here we were putting together a team of five with almost half being ill or sub servation the other half being AI this guy come from the fki here in Germany and he

created a start-up after this challenge and two mentors and they stayed 8 weeks in Oxford and one week with us and they created a kind of super resolution scheme for disaster management and it has spillover effect because UNICEF who saw this wow that's impressive because it's like having superpower you have open data 10 meter resolution but by using temporal information you can zoom in it's amazing when you think about it and they want to make this operational so this year we have another FDL we have two teams one disaster and one climate unfortunately we select it already the participants but if you want to be a mentor of an observer let me know because fresh ideas always good so what we want to do now is to bring this capability in space directly on the satellite it might look not new to you because AI at the edge is routine possibly but for us we are really old because space is a very dirty environment with radiation so computer has have not evolved as fast

as on earth in fact we're still using technology of 10-15 years ago because they are not radiation check the new technology like the chip are so efficient that even if you upset them they still work it seems so now people are just trying out they put Nvidia chip they put we we will do a satellite in September we will launch a satellite with the movi juice chip which is one for your mobile phone it has a heritage of space in it so it will resist radiation it has been bombarded and we will do computer vision and we are bumping into new problems like computer vision is three bands and we have 50 bands so are we trying 47 bands away how can we actually make things happen by transferring technology from existing fields which had not the same problematic and again that's where I need your brain Deb this is the satellite it's actually amazing I myself being in the business for years but to see this is just amazing so six you it's this

big satellite is this peak there will be two satellites too many talking to each other to a laser and a radio link exchanging computing power one will have passive microwave and radar and what we call GPS reflectometry and the other one a hyperspectral and a thermal and putting this into that little box is mind-blowing in fact many engineers still thing is not possible the hyperspectral is creating so much data we don't have enough bandwidth to download it so we will use the chip to better understand what to download and that might change totally the way we do space in the next 10 years because if you start to put the intelligence on board you can make space for the stuff the antenna can be smaller because you can just download some small information directly and oh and you can get it someday on your mobile phone so that will disrupt the whole ground segment by bridging the last mile and this is the view did the path to mass market so what we try to

do now is to have all these sensors which are disconnected because they have been built with specific purposes to talk to each other and act as a kind of swarm intelligence this is a big job but this can unleash will either the the power of all this capability and that was my last light thank you very much thank you thanks a lot can you want because I have no idea what you're doing I have to say but that's that's amazing because you can just tell us thanks a lot I think what's also interesting people can apply constantly to your program and just talk to you about the data so in case you're doing that we also see a lot of startups looking at computer vision and Earth data yes Houston ISA is also a funding agency we try to cover the whole value chain of business from the incubation so you have a crazy idea you want to build a prototype there is an incubation Center there are actually 13 in different countries you can

apply for this but once you get a bit more mature we have also called for innovation proposals where you can test out stuff with user so if you want more information just talk to me but this is a really cool field to be in and we will see why you're on the whole day thank you very much there's also the Copernicus masters which is also partly financed by ISA where you can just apply it with the business idea we also did this two years ago when we started with life yo and actually I'm wondering that you didn't know us because we are venture capital back from dieter phone hard spring so hard spring they're two brothers in Germany and both have a venture capital firm so what are we doing with life yo before I tell you this I just want to get a feeling about the audience who of you already worked with data from space is there anyone life life you employees don't count okay and who a few works with AI okay

so so a lot it's great because we just here in the top before that the space industry it's it was very close in the past I'm also space engineer and I want you to go into the automotive industry because in the past it was very very conservative this is changing now so with data from space we have to get more people from other industries into the domain and in the silicon value of course in the past years this was already done now I will show you what we from life you are doing because we are solving problems here on earth for example our goal is to monitor the biggest machine on earth what is the biggest machine on earth well you may not believe it but it's the u.s.

American electricity grid the u.s. American electricity grid has a length of around 9 million kilometers and compared to our standard in Europe it's completely different on average every custom and the US has a downtime of the electricity grid of 480 minutes that's an average value this means that in some places for example last year in New York in the area of New York canet the energy provider they had around 10,000 people without electricity for a week we cannot believe it it's the US and it was around New York so not somewhere in the middle when no one lives and what we what we just we got very nice business numbers in the past month vegetation management only in the u.s.

is a market worth more than a billion US dollar and only one customer which we just convinced of a project two days ago only this singer customer is spending 300 million u.s. dollar on vegetation management a year I bet you didn't even know that they're vegetarian management we also were really impressed and this is just the US if you go to the to the entire earth they are more than 30 million kilometres of infrastructure grits and the verticals railways pipelines and electricity this is around 80 times the distance to the moon so it's incredible the interesting thing is that all this infrastructure grids they are facing the same challenges they are distributed over large areas and you have to monitor them this is today done in a very ineffective manual procedure so people are working alongside the grid 9 million kilometers no problem and sometimes they they use digital tools but they're also often using paper and pencil what they are also doing is they're doing aerial surveys with with planes or with helicopters which

is I bet a lot of fun but it's very costly and some gas providers for example they are flying every two weeks along the infrastructure and there are people sitting there and visually inspecting what's going on on the ground we all know I think here in the room that's if you use a I apply it on Earth Observation data you may can solve the problem and this is exactly what we believe we believe the solution to this incredible monitoring problem comes from space in particular from satellites we already hear that there was a big change we had the easer perspective but also in the private industry it is completely changing the industry and this is the hockey stick this is 2017 and this is continuing so the number of satellites which are launched into orbit are rising it's it's just incredible and I really as a space engineer I couldn't believe that this would have happened five years ago the interesting thing is that you have different data sources so you after after this meeting

after today you can go online and from the public program you can download data for free and work on it so the Copernicus satellites you can see them on the left they have very big machines and they are powerful monitoring earth and with a 10 meter resolution if you combine this then with private data set then the magic starts to happen for example this is a planet off from Planet laps from Silicon Valley it's 30 centimeters in length it can take images with a three meter resolution the interesting thing is there's more than hundred of them in orbit so you get it every day you can get another image from them and then you also have radar satellites which can look through clouds the cool thing about images from space this is how we see the colors with our human eye red green blue but actually data from space is not it's not really an image it's a measurement a measurement in different wavelength so this is for example the visualization of the infrareds world

which we cannot see but the interesting thing is that if you use those different wavelength you can even if you only have one pixel and the time series can differentiate between different tree species only with one pixel you don't need a CNN for this and then also with with radar data it's it's it's really incredible what you can measure with this the challenge is and was always in earth observation to really find the business models there is a lot going on in the agricultural domain because this is the oldest and it's also there are a lot of big challenges I will just show you how we are monitoring those grids regarding vegetation and actually it it starts with the problem from actually Deutsche Bahn in Germany so they in the media we read a lot of newspapers in the past three years before we started life here there was a lot of political pressure on Deutsche Bahn because in the past three years they had more than 800 collisions of trains with trees which fell

onto the way way in Germany alone and I will show you how we are solving this problem so we have a risk through vegetation we have satellite data we are playing AI and how are we doing this and this is a spot in in düsseldorf this is a high-resolution image with submitter resolution you can think that of course you can with image recognition ya detect the vegetation there we did this with a 10 meter resolution already and we know where the veget the railways are and just if you have this classification of the image based on AI you can already find risky spots of the way way so you can tell that Japan that they have a bigger risk for example here compared to here that sounds easy the interesting thing was fast that Deutsche Bahn they didn't even know how many kilometers are exposed to Airways they operating 33,000 kilometers and had no idea how to where to start and how to solve the problem their plan was actually to send their inspection teams

their vegetation guys around thousand for five years along side the grid with new tablets they never work with tablets before to get the information on the vegetation next to the way ways we were in the procedure with Deutsche Bahn we told them that this doesn't make sense the people are complaining because they have to cut back the vegetation now they also have to assess the vegetation and then after three months they recognized five years with thousand people that's not enough it will take longer so we scaled our architecture and did this for entire Germany already this was our our big big first customer last year and now we can tell though Japan exactly where they have the biggest issues because what's so powerful about the data from space is that you get the overview effect so we now can also aggregate information so we can tell the managers of Deutsche Bahn there for example 34 regions which the infrastructure in Germany is distributed and they all have their own budgets they all have to hire

people and there's the center and Frankfurt of Deutsche Bahn they have to allocate the resources and you just have to imagine they didn't even know before where there is the bigger problem and the differences are huge for example here you have around 70% of the way was exposed to vegetation if you drive to balloon you may see that there is a lot of vegetation and other spots it's only 20% so if you allocate the same budget per kilometer everywhere it just does not make sense and with the power from Earth Observation we can now solve those problems directly for the entire grid this is also a lot of people are working with UAVs but the challenge with your visas are always you know for example for the u.s.

grid of our customer they need around thousand UAVs and hundred days to monitor the grid and we can already monitor the integrity and you don't need millimeter resolution to solve those problems so yeah we also would do not only sell the maps so we aggregate the information and combine it with business information so I mean it's not I is not only applied in the image processing also in the in the in the completes pipeline so you know when you when you already have aggregated information onto every single meter of the way way Deutsche Bahn they have a lot of information hidden there the electricity providers also have a lot of information hidden there and we just want to gain more insights out of this by using IIM technologies and then what's also really powerful is what we from life you did is we developed a mobile application because I think the ones of you which are familiar with AI know that once you have a feedback loop of hundred hundreds of people or maybe thousands

of people in the field who are gathering the data and give you a feedback loop loop you then really have a login and can solve even more problems so this is what we're doing based on satellite images we can provide a full stack solution for infrastructure management now I just want to give you a few examples of some challenges that we had and then also some opportunities so what you can see here for example are the Sentinel tides from the public program and Earth Observation data is really really heavy we so we made some Amazon instances degrade ating so the hardware was crashed was not our fault just because the processing time and the cloud resources you need are just immense and to give you an example every square here is corresponding to and the optical domain one to two gigabyte so one time stem of this of each quarter quarter I wanted to gigabyte for Germany you need around 70 of those for just one time step when you monitor every month you can

see how those numbers go up and then of course the challenges for people who are not from the earth observation domain you have so many different data sources and they all be used to solve different problems but to get this understanding you will have to dive into the data for example what you can get out of 10 meter resolution is nice but what you can get out of 3 30 centimeter resolution is incredible but it costs a lot of money so yeah that's always a challenge for us also for customers what kind of data source do we want to buy maybe we can solve problems with 80% accuracy with the free data and then also if you if you can buy high resolution data of course those bounding boxes you all know from other computer vision tools this also works on satellite imagery and there is a lot going on there for example optin insight from the US they are monitoring how many cars are in front of Walmart's and can then make an estimate

of their quarterly results so in the finance industry there also a lot of different applications and this is also a very true problem earth is a sphere I'm so sorry flat earthers but the problem is the images are flat so and there are I don't know so thousands of researchers and people they are making their own projection so there are thousands of different projections how you make the 3d world into 2d images and yeah this is just from our Chet a small small example of our daily work those are labeled from trees in Berlin and another intern from us wanted to open the data and then he was not in Berlin he was in the United Arab Emirates so this is always a challenge you always have the geospatial component and then yes daily images are nice this is the Tiergarten it's really the tea garden and this is also challenge in winter why were you really by maybe in the future you will see more in radar data and this is also an opportunity

and you have different like in the pre-processing and processing the post-processing a lot of different you can apply technologies and this is an example of a radar signal of M and maybe this was also what the ether guy told you when you apply filtering methods this is a conventional filter without newer networks it already works pretty fine and this is a filter based on new and networks and so scientific papers now really seeing that it's outperforming the conventional filters so even like problems you couldn't solve in the past with radar data you may now consult and then there is another very powerful thing about Earth observation data when you now want to start a UAV company for example you have to fly and get the data if you want to start with Earth Observation data from space you have historical data over the last decades and over the past three to four years you have nearly daily images so you can already get historical data you can train your algorithms and this is very powerful

this for example it's over one and a half years just one tree one pixel so every colors corresponding to to one tree in one pixel and where the time series you can see how the tree is really living and then of course what is now changing after all those satellites on orbit you need access to the data and this is a very cool startup also from from Berlin together with herbals and BCG they created a 42 where you can run and build geospatial applications and get access to high-resolution data there are also others but I really if you want to start with this I really think you should have a look there it's it's very very cool and then regarding the applications it's you can solve so many problems this for example it's a spot in Leverkusen a construction site and you really can see where the daily images how this construction evolves and the past days I was in in Frankfurt until yesterday for one of the biggest construction projects from from the last

decades in Germany and we also want to monitor the progress from space for those construction sites so the first consolations are there you can already work with it but what's really unbelievable is what the future will bring to us there are actually two startups one from the US one from Finland which are building a radar constellation where you can get data every hour for every meter on the planet and this is then the next dimension so I really would love to see more startups in this domain because it's really it's unbelievable what you can what you can solve with this and the use cases they they scale without borders for us we started with Deutsche Bahn and now we are operating in the US we are monitoring grids in the u.s.

we are monitoring grids in Austria for us it just don't care we get the data it's the same data and actually also on the same business and problems and we already saw the the blue dot and it's just imagine how many problems you can solve if you have this overview effect and I think really for this century all like the biggest challenges we have climate change the natural environment I've been planning all those problems we can solve with technology from space we have to first know what are our problems how big our problems and there is no other possibility to do this with data from space so I presume you all take it for granted that we always have access to electricity well yeah we just learned that they're huge outages in the US I did I didn't know that like my last outage was at least two or three years ago I am used to having access to electricity at home in the office in cafes or even on a train yesterday from Amsterdam

to Berlin but until two years ago I never really thought about this huge infrastructure that was needed to actually make sure that we always have access to electricity until two years ago when I met with the team at a company called vienen aquí so via energy is an Austrian utility company and it is their their job to make sure that everyone in Vienna has 24/7 access to electricity and like we just learned there's a lot of infrastructure that you know comes looking around when you want to try such thing so this team in particular invests in renewable energy assets like solar farms and this is a pretty big investments these soda firms are huge they cost a lot of money and of course we want a healthy return on our investment but they're obviously outside and over time they will tear down and break down which is a problem for this team because it broke and sewed our farm or a part of the solar farm because obviously it doesn't break all at once is

no longer producing energy so also no longer producing revenue and while it's standing still it's also you know you're losing out of money but you have to invest more money to repair it to get it back online so this is not a great thing and preventing this is really hard because it's it's not quite as big as the electricity grid but these fields can be really really massive spread out across the landscape or scattered across the rooftops in the cities and right now if you want to inspect this order farm like this they indeed go out in the fields and they take pictures but since they're so big and and that takes a lot of time and money they will only sample a part of the field so they might look at the the area on the bottom right but they have no clue what's going on in the top left another problem is is that they're the they're measuring the energy production but that doesn't really tell you much about how it's doing so

you can look at the energy production you can see it going down but you don't know if maybe a panel is stolen if there's a cloud over casting which costs the shadows or there's less production or if it's an over growing plant which you can actually solve by taking away to plant and then the panel will produce more and more energy so we know that we need to do a better job because in the future we'll obviously have more renewable energy sources but right now we don't really know what kind of actions we can take to stop these things from degrading and that's what we're working on this off so together with Vienna Nagi Berta has developed their first product last year and we aim to give these you call them operation and maintenance managers a bird's-eye view of their solar farm so they want to see in one clan how the whole sewed our firm is doing they want to dance oom into individual modules individual panels and they want to track defects over

time so what you can see is that small damages might actually evolve into much bigger problems over time if you don't do anything so since this is a more technical talk I have the privilege of explaining how we do that and what kind of data we use so you're already seeing the drone in the previous image and obviously step one is to collect the data so today we're using drones because we do need a much higher resolution than I don't know what what Pierre mentioned ten meters per pixel we look at a resolution of I think five centimeters per pixel so we can really see all the actually bird droppings on the panels so we don't fly to drones ourselves but we find a drone operator in the country that we're inspecting and this drone operator will collects images not only in the official spectrum so regular RGB images but they also use thermal cameras because a lot of the defects on solar farms are only visible in the thermal spectrum that data then gets

up late uploaded to our cloud platform and that's where we get to work with processing the data using AI we specifically use supervised machine learning algorithms so we trained well algorithms by showing it many examples of defects that we've previously seen this has all been done together with actual inspection experts so the data is really high quality and we can get pretty good results out of that just to give you an idea of what we're looking at if you look at RGB images you'll indeed find bird droppings or glass damages or over growing plants and also missing panels or missing clamps which are the things that keep them into place if we look at thermal data you you see if there's a hot spot it's obviously that there's something wrong but we can also tell what's what is causing the hot spot so it might be a broken sub string or hot junction box to me one year ago this all sounded like what's what's the high junction box but there's very specific reasons of

why they break down and you can if you know the reason you know what to do that's not where data processing stops we also just simply look at where the panels are so we draw bounding boxes around all the panels and by doing this we can create a sort of a blueprint so that brings me to this next step is that we do data enrichment if you want to call it like that we add more information to the visual data that we collected so on the blueprint we use that to create a schematic overview we're able to position all the problems that we found we indicate with callers what the severity level of that is we attach recommended actions to all the things that we found and then in the final step we visualize all this information to the customer so this is in a very traditional PDF reports but we also build our own platform where which is more interactive where users can filter and zoom in to the data we collected and this

is a small screenshot of what that looks like so now to get a bird's-eye view the customer only needs to log in to our platform which is called bird's eye view and they get their view of their whole solar farm the data is complete we've seen the whole farm this is possible because we use drone operators and we use automatic image classification because this this wouldn't be possible if we would look at the images manually there are also recommended actions attached attached to every defect that was found so with this information the customer may actually choose to repair something or send someone out to fix a panel replace a panel or file warranty claim so in this way they they know better what kind of actions to take to prevent losing money on their investments so today we're talking about what we are doing with that and how our business model works so I want to quickly explain what that looks like for us so we sell the access to the platform as a software-as-a-service

license so customers just pay a monthly fee to get access to that state of their asset which keeps updating every month or a year or however often they want and to actually update it and create a new inspection they pay a fixed fee for the analysis where we actually go through these data pros and enrichment steps to keep the asset up-to-date what we are doing is we're using AI to actually increase our personal profit margin so we can choose to do everything manually which will often viously cost a lot of time and money but by applying a I were actually able to turn this into a scalable business because we don't have to hire ten more people if we want to analyze them where images and honestly artificial inspectors are a lot more cheaper than human inspectors so far we have been working with customers in the Netherlands and in Austria so examples in the Netherlands are oXXO Nobel and zebra which is a solar farm consultancy company and our first customer which I already

mentioned at the beginning is v9g in Austria we have detected last year many many defects in many solar farms so last year we were still fine-tuning our technology and building your products we looked at nine point seven megawatts of solar farm to me that didn't mean anything so I just listed the amount of panels which is forty thousand and a little bit more actually today we're flying another solar farm with twenty thousand panels so that will be a fun fun week next week to add them to the platform and luckily I'm not doing that alone there's a great team so mike my CEO and co-founder cameo is currently overseeing the inspection of this among this solar farm and there's also a bunch of developers on AI but also full-stack business developers and production production managers my name is Anouk I've been working in AI for over nine years a lot has changed since I started my studies at the University of Amsterdam and I was actually working on student projects together with Camille from five

years ago today as well so it's an exciting anniversary for us and from that student project we actually got the idea of we were doing something similar with drones and solving a cool challenge and we found out that people need a bird's-eye view of what they own and they don't need only need the bird's eye view of solar farms they also need a bird's eye view wind turbines of cell towers sea vessels chimneys and indeed the electricity grid so we are really scaling up our team this year we're looking for funding because we believe there's many more assets that are not supposed to break down unexpectedly so right now I told you were using drones with thermal and RGB cameras to visualize how a solar farm is doing we're doing a similar thing for wind turbines where we also use drones because they're pretty large to reach manually there were not only predicting defects but also the severity and how that will develop over time and new business cases that we're looking into might also

include coatings on things like sea vessels so there it's more of a manual approach where if you dry dock a boat like that you send out a couple of inspectors with iPads they take a photo but it's really hard to estimate but kind of what kind of damage is happening and coating so we want to use AI to make sure that all the data gets labeled in the same way so that's a different use case but the same problem you have a big asset you want to know how it's doing and you can scale it up way easier if you use AI so we're really excited to be here today and these are my contact details if you want to help give our customers a bird's-eye view of their assets so they can actually take it for granted that they don't break down I'm really happy to talk to see how we can work together thank you may I ask you to stay on stage maybe you can join us for some Q&A we still

have some time so thanks for making up the time Daniel was talking so much no but we still have a bit of time for the other speaker and I think some of you guys might have some questions so please just raise your hands and then so yeah we probably you don't have one yeah I'll get you one can we have a microphone up here ah cool shall we start here yeah perfect there's the microphone coming get to the point thank you for the talk it's very inspiring indeed I have the question about the workflow so what where's your engagement starts and where does it stop do you already operate the drones and do the pre-processing and all the work flow till the analysis and yes so it defers a little bit better customer so some customers are already experimenting with drones themselves but we are always involved in the process so we created a flight protocol that ensures that we're collecting your right kind of data and that the quality is high enough and if

the customer is not working with drone operator we'll just find one in in their area which is quite easy nowadays and then yeah we take the data from the drone operator we go through that whole process always if it's a new asset we may ask for some feedback on the kind of defects and the damages that we need to find we showed a visualization and that's sort of where it stops but it would be really interesting in a future to also really communicate with the asset maintenance team of what kind of actions are they taking how are they responding to their problems and then do they fail you our recommended actions at the aesthetic problems like you know the trees they are not growing that fast so you you take a picture and you know probably once a year or half a year where you have to cut them do we also do dynamic analysis for for construction sites like in the nature of a construction it's very dynamic that's why we are also like

working on this in this domain and then another thing which we realized which also took a lot of time that the frequency of the usage of our product does not depend on the frequency of our analysis because we have work force management integrated people are using our app in their daily life there's like they have their tasks they feed information back that's why we don't have to analyze every week with the satellite imagery they're using our software every day of analysis and most dynamic for construction site once a week so every week an update about the state of a construction I think we need the video so we have thank you you have solution for a special industry can you tell us something about the way from from the idea to the startup so for us for me it started like I said five years ago we were joining a global challenge to protect game reserves from poachers this was a very cool problem but protecting rhinos instead of poaching rhinos there's a lot more

money on the the wrong site so we were looking for a more profitable business case than a non profits goal which is still part of the vision of the company that's where we want to go back to so we took this idea if we want a bird's-eye view of the Game Reserve when we started looking at industries that also want that so we passed game towns livestock management agriculture but as we know from from the first talk agriculture benefits from using a lot of fence and it was really not our expertise so then we looked at other assets and we found that the inspection industry with drones was really taking off at a time so there was there was a lot of steps in between to finding our product market fits it has been a good journey yeah for me it was in yeah in adventure I wasn't Munich three years ago like in the automotive industry at BMW and I recognized I studied space engineer I wanted to go into the space industry so

I said no I'm gonna do a startup in the space industry and my first ideas because I was in the wind tunnel there were some hardware ideas so I wanted to build small set components or even a small set systems so I met my co-founder like tomorrow exactly three years ago who also wanted to build a startup in the space industry and then I moved to Berlin we shared our ideas of our Hardware ideas and also for our software ideas and we recognized we are around five years too late if we want to build hardware so that's why we wanted to build a software solution our first solution actually was completely different so we wanted to build a tourism application so we wanted to classify beeches and build an affiliate based model so you could find a beach which is crowded because you can actually measure how crowded the beaches are from space I thought this may be nice we talked to Tim we talked to trivago they liked the idea and but then when

and I we said maybe we can with the technology soft bigger problems and this is then we made some analysis of some market and markets and this the infrastructure utility sector was our our market which we are now targeting you

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.