Seminar #20 of the International Tree Mortality Network seminar series.
Speaker: Dr. Philippe Ciais
Full title: Harnessing big data and artificial intelligence for forest monitoring
Abstract: The fusion of satellite data from different sensors allows to create new maps of forest biomass and height, and their changes over time, down to the level of individual trees. Philippe will present new deep learning models of forest attributes and applications, including an attribution of disturbances over France.
Okay I think we can get going uh welcome once more uh this is the 20th international tree morality seminar I’m really really happy to see you all here um and it’s a great great pleasure to introduce Philip C to you Philip thanks so much for making time um it’s uh
Really great to have you and um you know it’s it’s really difficult also to introduce you because you’ve been doing influential work in in a such breadth of fields that it’s really hard to to where to start and where to stop I will just be very brief uh Philip has a
Background in physics actually um has worked quite extensively on the global carbon cycle and carbon budget uh has worked uh for and with the ipcc um has uh co-authored uh a vast number of papers so it’s I think it’s more than 1,000 uh and just in the last year uh
These Publications have garnered more than 30,000 citations in a single year so Philip you’re definitely uh what constitutes a rock star I guess in in our in our field here so um it is wonderful to have you and I am particularly excited because uh I mean there’s so many things that you could
Talk about today uh to us but you’ll be talking about uh as illustrated by this nice image also a very uh topical uh uh um subject specifically harnessing big data and artificial intelligence for Forest monitoring and I think this is uh monitoring forests and monitoring the changes in Forest this is something that
Is uh um crucial to a lot of the things that we all do so uh I’m really really curious to uh learn about uh what you’re doing and uh what you’re proposing uh and before giving the floor to you I will just also say to the audience that
Uh there will be ample time for asking questions uh you can do this already during the talk uh during Philip’s talk you can post your questions in the question and answer field which you find below uh in your Zoom panel uh so post your questions there uh and we will then
Relay your questions to Philip and he will answer them after the talk um and with that um Philip uh thanks for being here once more and the floor is yours looking forward to your talk okay so thank you very much and for the invitation and for the introduction and
Uh I’ve been working a lot uh on with the process oriented models but I find that it’s always frustrating because their level of development is always a little bit pathetic compared to the reality we are facing and that’s the reason why since maybe two or three
Years I’ve started to look more at uh the problem of understanding and monitoring Forest carbon with data in particular with satellite dat data and uh it’s clear that the kind of societal and the climate Stakes uh related to Forest are very high uh in the global budget forest alone are estimated to
Make up most of the landn itself absorbing about 30% of the emissions and uh the most important thing is that many countries have uh are counting on forign to meet their climate Target so basically they are counting on their Forest things or even on improving their Forest carbon uptake in order to meet
Their neutrality goals this is the case of Europe Europe uh wanted uh to cut their emissions by 40% in 2030 compared to 1990 and they decided two years ago to open the Pandora box instead of only looking at industrial emissions they said okay let’s give us a more ambitious
Target we want to cut their our emissions by 55% instead of 40% but this time we will use uh the land use sector and in the land use sector uh the sync is mainly in the forest sector so basically the EU for instance as many other countries are counting on Forest
Carbon up take to improve their commitments to neutrality uh in fact in Europe uh the situation is that we have seen from Grand data quite uh a persistent increase of uh disturbances and as you know in this group disturbances are a very good example of fast and sometimes
Massive losses of carbon that can uh you know in one event on a few year negate the slow carbon accumulation uh exposure to climate risk are in have increased over the last 40 years and I take the example of France for instance uh we see that over the
Last 10 years from the NFI data the mortality of trees has increased by 80% and talking here about the natural mortality not the Harvest it was I think 4% as a background value including a few big storms and now it’s about 8% what is even more worrisome is that
Many Forest are planted they are not self regenerated and we observe 54% failure of the new plantations so under this question uh the EU has committed uh to create some s to maintain some in the L use sector and when they took the decision if you look at the numbers basically negative are
Uptake and positive are sources basically cropland and emission of greenhous gases and the if you look at the green curve the green curve is the European Forest carbon syn as given by national inventories uh and it was pretty stable it was about 450 million ton of CO2 and and uh starting between
20210 and 2015 the Sync has gone down substantially by almost 20% and we are now below the target uh that was taken by the EU so when they choose the Target in 2030 we’ll have a carbon sync of 400 million ton CO2 per year it looked easy because the uh
Actual uptake was above this number but the Sy has decreased and this decrease as I said with the French example is mainly due to an increase in mortality and the slow down of tree growth some countries like I think the Czech Republic if you look at their National
Inventories have turned from being a syn to being a source because of massive back Bol attacks in France the if you plot the same diagram the national forest carbon syn has dropped by 40% so it’s three times more than the average over the EU so it’s clear that we are
Facing a problem and uh uh the question is what data do we have to at least monitor what is going on we have of course Forest inventories they are forming the backbone of the long-term monitoring because uh uh in many countries of of the countries of the north the the plot
Networks are dense inventories have been designed to give an average carbon stock of a Nation they have not been designed to give the average carbon stock of a small region because of the limited amount of plots uh it’s labor intensive work it’s accurate Tre level data uh and the idea
Is to make a pole of the forest uh for instance in France the national inventory is measuring every year uh 14 ,000 plots 7,000 are revisit of plots that were previously visited 5 years ago and 7,000 are new uh newly chosen plots uh so although this is sufficient
To have let’s say an average stock when the system is changing in a way that you have a small regions that can be affected by Massive losses like you know fires wind the insect disturbances the capacity of the sampling of the NFI to detect fast changes especially when you bring it
From the national scale to the regional and to the forest scale becomes very limited the second point is that intrinsically the sampling scheme is a kind of running mean over five years so if you have a shock year like in 2018 it was a very severe drought in Germany and
Poland you will only discover the impact of the shock when you will have passed your running mean through the shock here so it means that it’s only four years after the shock where you may have lost a lot of carbon that you will have a better idea of how much carbon has been
Lost but you still are not able to account for the instantaneous uh let’s say interannual change of carbon in the forest on the other hand we have also you know Edco variance measure it’s a completely different observing system much more expensive they do provide very accurate records of the net
Carbon exchange with the atmosphere they have some kind of upscaling properties and a lot of work has been done to use satellite data and climate data to provide greed estimates but they remain extensive rearch based mement and they do not give us information about the carbon stock directly that’s why the
Existing asset probably uh needs to be complemented with satellites and uh well I don’t need to explain that satellites offer a lot of advantages the advantage of a global coverage with consistent measurements the advantage of a relatively frequent revisit uh and also they measure different properties that are interesting for looking at
Carbon uh the classical sensors are Modis of course there is a compromise between spal and temporal resolution modies flies daily over the globe and has a resolution which is moderate lat has a higher resolution but it has less frequent images and what has been a bit
Of a game Cher and I’ll try to find the time to speak more about the use of those data is that thanks to some Californian innovators basically companies that have launched uh micro satellite or cube satellites uh we now have the possibility to use constellation of
Images that bring uh the best of the Two Worlds basically they provide a high spatial resolution down to three M even better for some of the sensors like worldview and uh they have a frequent revisit so now thanks to constellations like the planet data we can have an
Image of the earth uh every day at 3 me res solution the drawback is that the quality in particular the spectral quality of the data is not as good as the one of the research class satellites so satellites offer a new window to monitor wall to- wall carbon that can be
Like a propaganda thing for a space agency or research agenda but keep in mind that if we had the satellite that measure Forest carbon we would be in good shape but known of the uh observing uh stat satellites are giving you a carbon variable satellit are measuring colors spectral colors
Radar back scatter sometimes liar properties but uh these physical measurements have to be translated into carbon quantities and that’s what we Tred to do in collaboration with colleague from inra and Copenhagen University and now working also with Cornelius and his group we have been trying to make sense
Of of uh satellite data combined with ground observation uh to derive uh maps of biomass B maps of biomass change and maps of let’s say activity that can be used to attribute those changes uh we are facing the problem is that if you see that there is now an image of the
Earth at trim resolution every day it’s plus all the existing other sensors it’s a delus of data and uh the kind of classical way of doing remote sensing taking one satellite a couple of bounds forming some spectral indices uh is Impractical if you want to make use of all these data you have
To change a bit the kind of model that we are using to go from satellite data to carbon and that’s why machine learning and deep learning offer some Advantage they have also some inconvenient to convert a Del of satellite data that human brain cannot make sense easily anymore into a carbon
Quantity the disadvantage of machine learning method is that they are a bit like black boxes so they will give you an answer but uh they still have some kind of interpretability issues so to start with uh as a good example of what we did without machine learning with
Jeanpierre Von as you know monitoring biomass change is still something extremely difficult and uh jeanpier had the idea to use basically the noise of other satellite communities to derive an indicator of biomass uh the SM satellite and the smap satellite are lbound the passive microwave antenna they measure the kind
Of emissivity in in a microwave and usually people don’t like vegetation because they use a satellite to retrieve so moisture and vegetation contains water so it’s a it’s basically an annoying thing for that prevents the nice retrieval of Soul moisture but you can see it the other way around you can
Say well after all if vegetation water content influences the signal we should be able to look at the volume of water in the vegetation and to transfer that in biomass and it is this nice indicator this vegetation Optical depth which is nicely related to biomass with much less
Saturation than any Optical or shortwave microwave uh uh data that uh we used to uh create this global map of biomass change you see in the picture in green the areas where we think from this elban data that over 10 years there is a gain of biomass you
Recognize some of your pet sync areas like Southern China forest in dark green also this abandonment uh of Agriculture area in the south of Russia and you see also Brown areas where we have a loss of biomass in the the deforestation Frontier of Brazil also in central
Africa and indeed visually the the the map was kind of nicely correlated with the forest loss map so with this uh proxy we found three things the first one is that the global biomass stock increase on average over 10 years is only. five uh gigaton of carbon per year
And we said well it’s too small because there was this famous petal paper which says it could be two but then we looked at many other satellite biomass change data and they all give you uh a global biomass increment which is between 0.2 and 0.5 so we have here already an
Interesting question for the global budget maybe the satellite data have some saturation and they are missing some of the global carbon syn it’s possible maybe they are right and then it would mean that the global carbon SN only uh 25% of it is in biomass and the
Rest is in non- living things like course debris it’s possible for organic carbon sediments and so on uh so VOD is nice it’s one of the first record of 10 years to track biomass change of course it has some issues but it’s at a course resolution of 25 kilometer and it’s a
Bit course to understand the processes at play that is why in kind of second step and we are not the only one to do that in the world we started to uh see how could produce better biomass map from Jedi I don’t know if you know Jedi it’s a fantastic instrument it’s
Basically consist of four laser beams which are uh traveling across the Earth uh from the space station and Jedi is sending a beam of infrared photons uh so we have eight orbits and the beams are like a forest inventor basically it has a size of a forest plot diameter of 30 m
Spaced by 60 M from each other and the Jedi measures the waveform the waveform is showing the first return return of the photons on the on the tallest trees in the footprint then you have a vertical structure which depends on the position of branches leaves and
Under story trees so if it has a high peak it means that it’s a kind of you know modol layer the top canopy if it’s a very flat it means that it’s a kind of very dense and very thick canopy and it has a ground returns so from the
Difference between the top height and the ground return you can get uh a direct height measurement and we all know that height is related to volume is related to carbon the problem is that Jedi is an incomplete sampling it’s like an NFI but very much denser in France we
Have 14,000 NFI plot and there is a 100 million footprint of Jedi but still you know uh the question is uh there are still some gaps in between so what we have been trying to do is is to try to teach uh a deep learning model which is
Called a unet model to produce continuous high resolution maps of forest height using as an input data images which are coming from Optical data so 12 band from 10 band from Sentinel 2 and radar data for different images a day every five days from SEL one radar so uh imagine that problem is
Simple Y is a dependent variable it’s a liar height and X is one band of Sentinel one Sentinel 2 so you would write a regression model like y equal a X plus b and you find a and b the problem is that of course the relationship between height and spectral
Indices or rad is highly nonlinear and what the unit model which was derived from medical imagery is doing is that it’s providing convolution of images where the coefficient of the regression are the weights of the convolution and typically with a network like this you have about 17 million of U regression
Coefficients if you want trainable weights uh it’s still not overfitting because you have billions of satellite pixels in the input data and you have also millions of uh Jedi footprint so this is an example this is a forest that lives in the North of France it’s called see you have a Google
Earth image and this is what Jedi is showing you the row data uh row data of Jedi shows that there are some region of these Forest which are yellow so the for the trees are tall about 30 m some Forest which are very short but still
It’s difficult to give a you know an idea of the distribution of height and after training our unit model of course we verifi that there is no over fitting and so on and so forth you get get something which is looks like a nice picture of the forest you see the
Geometric variation of height because you have some Forest that have been cut and they are recovering and they have different height but you see in this Central Forest which has tall trees that you have also some kind of disturbance roads and gaps uh which are almost detectable with this high resolution
Height map uh so we produced a map uh for France and now for the globe at 10 m resolution and using a simple allometric relationship a stand level between height and biomass uh we derive the biomass map as you know biomass it does not make any sense to derive it at 10
Meter because you spot a trunk a stem you have 2,000 ton you move your pixel 10 m away you have a branch and you get 10 ton so it makes sense only to infer biomass at the spatial scale where you have many trees or to do it at tree
Level which we cannot do obviously with the 10 m satellite resolution the question is how good it is okay I produced nice maps you can see that the Alps and the east of France have a taller Forest than the Mediterranean and the leand Plantation uh so we did some evaluation
Using two observing systems which are completely independent they are not used in our model the NFI data so you see all the NFI plot in green some kind of uh permanent plot called ICP it’s a French ICP and some Airborne liar campaign and uh we were pleased to see that after
Several adjustment the model was able to predict quite accurately the values of height at the NFI plot which is on xaxis which is completely independent from the model so it’s possible yes from space born data to have a very accurate prediction of height with a mean error of about Less Than 3 m
Similar performance for the evaluation against Airborne lier data still some classical problems of those deep learning models they tend to underestimate the tallest trees because there is not enough training data uh and uh we said okay this is nice but we are not the only one to do this
And uh there are other products that produce uh land uh Forest height first of all we have poap of paper they used the similar approach with the lat instead of Sentinel and Jedi with random Forest nikang has been using Jedi data with Sentinel 2 only and we use sentinel
One and two and there was a recent paper by our colleagues from Copenhagen they use the planet data so it’s really the heavy artillery uh and the Airborne lar to produce a height map at 3 met resolution for Europe only so how does our height map compared to to other height products
First a visual inspection poo on the right long then Leu this planet product and our product they look similar but not quite the poap of map looks very flat L looks uh very more like fuzzy I mean it’s not as sharp as our map and
You has the same kind of contrast but it seems to be uh producing lower height in many parts where we diagnose the tall Forest which one is right or which one is better so this is an example of the comparison between the uh Airborne like the True Value the Airborne ligher
Height uh again the four products R from right to left and we found that potapov and L have low correlations with this uh kind of Ls data I was surprised because I would have expected that the planet product would be better than our Sentinel one and two based product but
Indeed uh it’s uh the error and the correlation are slightly lower than uh what we obtain with our deep learning model why is our map better for France well maybe because we are French and uh we designed the model for France so it’s easier to do it than the global product
Maybe because we use the better uh deep learning model maybe compared to long because we in injected The Sentinel One S data Maybe the Sentinel one data are making some difference and as I said maybe because our study was local and the other products at least two of them were
Global uh second thing once you have an accurate height map you can talk to Cornelius and rert and you can use their disturbance map to look at the height today or the biomass today and the age of the forest which is the time since the last disturbance occurred in that
Pixel and by putting these pairs of height or biomass and age together you can derive a what you call a growth curve which defines the recovery of height as a function of uh time since disturbance we did that here for a small region in France and uh with the lat
Data you get all the color curves which uh correspond to each species because we cross those recovery curve by a high resolution species map beyond the lat AA you cannot estimate a height for Forest which are older than 40 years so we completed them by the black dots which
Are the NFI H and height measurement there are several good news first of all these curves are important because they can be used for management and the expectation of Rec carbonization and so on and so forth secondly you see that there are some clear difference in grow rate between fast growing species like
Poplar and slow growing species like oak or fast growing species like duglas on average in this region we did not find any significant difference in the fit parameters with the Chapan reard equation between the Conifer rate of free growth and the broadlea ones and the third good news is that the
Satellite based So based on our height and the L side disturbance map from Cornelius and the NFI data the very consistent with each other so two different ways to retrieve gross C uh give a very different very similar result so once you have gross curve uh
And biomass data you can use that for a number of application one interesting application is that in 22 in France we experienced a completely incredible year with extreme fires that happened in regions where they do not happen before usually in France fires happen in the south of France in the Mediterranean
Biome with some internal variability so you have the burnt area at the bottom and suddenly in 22 because of a very long and dry summer we had two mega fires in the lon forest in the Atlantic part of France and a few fires in temperate forest and by crossing our
High resolution biomass map with the Sentinel two burnt polygons we found that this is really an extreme year because the Mediterranean did not lose more biomass than usual even less than usual but you have this new fire regime in particular those affected some old gross forest in the north of the Lan
Plantation that make up a very strong loss of biomass which is estimated to about two ton of carbon another interesting result is that we have very few fires in the temperate systems in Britany for instance or in Jura but these Forest have two times more biomass than the the Mediterranean forest and
Therefore even though the burnt area was small the biomass loss was even higher than in the Mediterranean regions because the impact system had a higher biomass density uh then we tried also to look uh a bit more closely at the map of disturbance because all this loss and
Regrowth depend on the disturbance agent and uh you know that there are different type of disturbances very different like you have whereare and widespread things like storms fires you have also a lot of cuts some biotic dieback just like defoliators for instance that we try to distinguish from biotic mortality which
Is mainly barbal so insects that defoliate but not kill an insect that kills and we tried to provide an attribution of the map of Cornelius by using also deep learning we use different data set we have have more than 10,000 field observation of Forest Health as part of the inventory but it’s
Not a national inventory it’s just Forest Health survey people go in the field and they say oh there is a bar Bol oh there is a defoliation blah blah blah secondly we have the NFI each five years they give you Tre level data and they tell you oh it’s a spruce forest and
There was some natural mortality or 10% of the trees have been chain sold and then we have another database which is a semantic database called DFD it’s basically a big Excel spreadsheet where you see oh there is fire in the region of mster in Germany in
21 but there is no geolocation it’s just a text so if you app text mining plus you kind of find some distance between each of the disturbance polygon and the site where you have either Health observation or NFI observation of course you know there is the disturbance but
There is the time uncertainty because NFI is only every 5 years plus the semantic data we are trying is not perfect yet to make a classification of the nice disturbance map that was produced from the lset data it’s working okay but it has a lot of margin for
Improvement one of the margin for improvement is that we know that the disturbance uh sometimes don’t come alone like you have a Dr and then you have a back Beal attack and then people are obliged to do Salvage cutting and they cut the forest so what is it is it a biotic mortality
Or is it a cut if you look at an NFI data you see the tree chain sold and nothing tells you that they were backs BTO before or not if you look at Satellite data you need to have high frequency to detect the kind of cascade
Of things that are going wrong but we have been trying to associate basically consecutive Forest disturbances to look at those cascading effects which is not completely easy but it’s work in progress uh we also use a disturbance map and our grow curve derived from satellite data in this case we don’t use
Our liar data but CCI biomass uh we noticed that there is an increase of the disturbances also an increase in their severity in many regions uh here Europe is divided in few few regions so we divide Europe into small squares of 18 kilometer in which you have many 30m disturbances and
Then our Po from has been trying to derive grow curve one every 18 kilometer so we know how much carbon is regained by the forest after disturbances on a GED here I show you only the aggregated regrow curve which are derived from satellite data for uh different regions so you
Have the Alpine region the Boreal the Atlantic you can see that the different regions differ by their asmt obviously you can get a much higher asymptotic biomass after full recovery in the Alps or in the Continental part of Europe compared to the Mediterranean or the Bal
Region and also the regrow rate is uh uh smaller slower in the Mediterranean compared to the other regions how good are those satellite product well we compiled with n a large database of Chronos sequence of forest plots where you have biomass and age there is a lot
Of noise of course because the plots include species soil differences they are very local measurements and our curves are a kind of average integrated recovery at 18 by 18 kilometer on a grid but still there was no significant bias across all the AG beans between our reg
Grow rate and the one you can get from a few plot data so then we said okay we know the recovery we still need to know how much biomass is lost for each type of disturbance and this was not easy because we have a map of disturbance
Severity but the world is not simple it’s not because disturbance severity is 0.5 that you lose 50% of your initial biomass or because it is one that you lose 100% the ratio between the disturbance severity and the biomass lost is not constant and we had no way to infer it
On a grid so basically what we did is that we defined this parameter this ratio of severity to biomass loss as a single scaling Factor per country that we tuned to match the NFI data so this is our magic tuning we tune it only one year and the good thing is that our
Model which now predicts the development of carbon because every year you read the disturbance map you know know how much is lost from your magic ratio and you know how much carbon is regrowing our black curve for each region but we can do that also for each country is now
Following relatively well the NFI trend of year-to-year changes of AGB which is a carbon sync in biomass that is good because if you try to do it directly from the CCI map of EA they have three consecutive Maps if you make a difference from year on year
Between between 2018 and 17 2019 and 18 between the maps the ISA map will tell you that Europe is a source of carbon in biomass which is not true we have enough NFI data to know that we are still a sync a decreasing sync as you can see
But still a sync we believe that this is because the E map is capturing very well the losses but it has some striping effect and it has also maybe some problem to captur the growth of forest that have a high biomass and if it is skewed towards SE see the loss is better
Than the sinks it makes sense that when you make the year-on-year difference of the biomass from the CCI Maps you are going to overestimate the losses and you get an unrealistic carbon budget so now we have the calibrated model that gives you gain and losses the only thing we
Did is we made the hypothesis that in each region the trend of disturbance severity can be linearly extrapolated from the last 30 years in the regions where you have a trend that’s what fran did so he can now calculate on agreed but this is shown for the broad part of
The European you know Forest it can calculate what will be the biomass carbon accumulation the net one including the loss from disturbances if disturbance have a trend of linearly increase in severity and the bad news here is that we predict with this very simple data driven model extrapolated to the next 10
Years that the European Forest Sy will continue to drop and uh it will drop to a much lower value than the political Target of the EU in most of the regions and indeed some regions like Borio Forest but also some parts of France and Eastern Europe are going to be carbon
Losses even if there are small declining carbon sinks today so now uh this is what we can do with the model basically a bookkeeping model with loss and gains another interesting uh approach to monitor carbon gains is to do it directly from the satellite data but the signal is
Small it’s easy to detect the height of 20 M out of zero but it’s very difficult to detect the height increase of maybe 1 m per year it’s signal to noise ratio is much smaller so we tried with the unit model to build uh basically uh a year agnostic
Model so we have the Sentinel data for five years we have the Jedi data we don’t want to be sensitive to phenology or drought or interal variation so we teach the model how to model height in a robust way whatever the darkness or pale color of the veget a due to drought and
Phenology and by building a year agnostic model we can then use it to predict year toye changes of height this is an example of the model result so we now have a map of height every year uh this is a small plot in leand and uh if you make the difference between the
Final and the first date you see a lot of red areas that have lost their height and this is obviously because they have been in Clear Cut as you can see from their geometric forms what is more interesting is that you have a lot of plots that are green and they have
Gained height and the question is that is this carbon gain or height gain signal realistic or not is it does it make sense or not maybe it’s just an artifact of the trend of greenness or radar data and the Jedi noise so indeed it seems to make sense when you consider
The map in 2018 as a Baseline and you make a difference uh between the height map you get in 2019 then 2020 and so on for the forest in leand you see that the trees that are 5 m tall they seem to grow quite well as you would expect by
Almost 1 meter per year but the trees which are 20 M tall they grow very slowly and that makes sense but it’s still interesting to see that using these deep learning models we seem to be able to calculate uh a growth signal from year on year even if it’s a very
Small signal compared to uh uh uh basically uh having just a height map this is another type of forest the one on the left is a Pine Forest maritim Pine Plantation they grow very quickly as you can see the other one is a alipop pine it’s a completely different species
Slow species in the Mediterranean region and you can see here that the grow rate are in fact very flow and that they tend to be higher at taller height older trees compared to uh uh younger trees and uh the last curve is a duglas fur duglas fer is a monster
Species it’s planted over clear cut in the center of France and when the plantation closes you see the trees the duglas which are 10 met high they have a crazy high grow rate and then of course the grow rate declines when the get a little older so with this new map of
Annual height change we can produce some kind of height on biomass increment at the scale of those relevant management and climate region of France so now we said okay we have done some work for France we had some papers where we applied similar model for Ghana for
Tanzania we are working for Spain and on the other hand there is some Global productor from nikol long and Peter puto and others and they seem to have some kind of quality difference so the question is why uh is there a difference between the global product and the local
Product and uh using the same model than the one we deployed for France we uh found that if you look at our map for France for instance it has much more you know contrast and it looks to be of better quality than and the global map so we trained basically uh a
Global model uh with the same kind of technology and we found that the global model based on Jedi data and trained by Sentinel one and two can give a result of the same quality than our kind of super tune model for France which looks much better than the quality of previous
Global model the same for the Amazon uh it’s quite amazing compared to the long map that is very uh kind of as low contrast you can almost see group of trees in the Amazon at 10 m resolution with this new high resolution map it has a good cross validation performance how
Do we do that well you just need to have a very good computer equipment and that’s what they have in the uh mster University they have also very good skills in machine learning they divided the world into time zones and they have basically processed all the Sentinel
Data each snapshot is 45 terabyte of data the Jedi raw data itself is a 30 terab of data when you filter the bad data with clouds and fogs and you know no height you get still a 200 gigabyte data set we use a Jedi unet model I pass on the technology but we
Are aware of the fact that there are some geolocation error in Jedi so we use an adaptative loss function basically we move the ji granules by a certain error which is the declar the geolocation error of Jedi about 10 m and we pick up the shift of the Jedi footprint that are
In best agreement with the data so this allows us to correct problem like this one you see that the measurement of Jedi are supposed to be in a forest but they give you a height of 2 m so obviously they are shifted North in the reality but the geolocation tells you that they
Are in the middle of the forest and using this kind of shiftable uh data of Jedi you can partly account for the low correlation geolocation High geoc ER of the Jedi data okay and then you just basically need to have a big computer what is the amazing is that maybe five
Years ago you need a team of 50 people to do that and now a graduate student very smart one can do the entire training of the deep learning model on 3 GPU for two days and then it can produce a global map provided a lot of gpus 40
Gpus in only two or three days which mean that today a grad student can produce a global map of forest height or Forest biomass at 10 meter spal resolution uh in one week then you can change the parameters adjust and so on we have to validate the map against liar
Data but we have for instance a product for Congo if you are interested and the Amazon of height not yet biomass you can see that there is apparently over Congo very tall forest in the western part of the Congo Basin or the Central African Forest I should say compared to the
Eastern part and when you compare the quality of this height map to the alss data there are lot of ALS data in DRC in camon so we did a systematic comparison you can see that it’s not as good as that we get for France but we still get
A mineral of about 4 to 5 m and the forest are completely different from European forest and they have a height of up to 60 M and the comparison between the satellite model and the kind of ALS data you can compare the red curve which
Is the ALS and the green curve which is the model is very good so the model is quite capable to capture the small scale erogeneity of the forest I think I’m running out of time right so I should maybe conclude I just wanted maybe to show show some of the work in three
Slides that we are doing with our colleagues from Copenhagen they are doing amazing things with the planet data they have trained the unet model uh to try to do height and trick over a Trier basically by using as an input the planet trim images and trying to ask
Them to mimic the alss height of our compaign regions so they produced a very accurate 3M Tri over of Europe and the CM biomass map this trees outside Forest is an interesting topic because in some countries like the UK we found that 25% of the tree cover is not in Forest
Because it’s very bare it’s very fragmented you have very few like big Forest chunks and the large fraction of the canopy cover is in land cover classes that are called crop land grasslands or Urban trees and in on the Atlantic part of Europe like britainy Netherland Denmark it’s about the same
Story uh using this time uh even higher resolution data so we’re talking about aerial photos with a resolution of 20 cm it was possible to segment all the trees of Randa and here of Denmark using a double task Unit Model that does both height and canopy segmentation um this paper you may have
Seen same techn ology using I think World View satellite at5 M you ask post dos or humans to uh delineate trees but of course if you they have to do it manually over such a big region it will take them several decades and then you train machines to infer the height and
The area and from this the biomass and the carbon of those um arid trees and we counted about 11 billion of trees from the wed sea to the Atlantic Ocean this is an example for Wanda another example which is very cool let’s say cool or bad
News I don’t know India is the biggest countes for Agro forestry it has a huge amount of trees over cropland and the group of Copenhagen they make the difference between the large trees so we are sure that there is no false positive in the detection between the rapid ey satellite images uh 2010
And the planet scope data average over 201822 and to our surprise in this study we found that in most of the regions where there is Agro forestry in India you see that many trees which are indicated by the pink dots have disappeared over one decade so why do we
See a disappearance of large trees over agricultural landscape of India we tried to do some interviews and surveys of local rural communities we thought it would be due to Drought or maybe some disease but in fact it’s just that people are getting richer they buy machines to do Harvest
And they don’t like the trees so they are cutting the trees and they are cutting the large trees unfortunately uh one last thing is that can we do better with high resolution images and in this new study we try to ask a deep learning model to do biomass
At plot level so you take an NFI plot and inside the NFI plot you have a very high resolution 20 cm image and you ask to a computer to use the texture and the information contained in this very high resolution image to create a single dependent variable which is the biomass
At the plot level is it better to do it by trying to segment all the trees but then you miss the under story trees in dense forest or just to exploit with the Magics of deep learning the texture of the images themselves and the answer is here is quite amazing you can compare
The performance of a Brute Force segmentation so you Contour every tree you have the height because you have some lighter data we can get a fair estimate of biomass at plot level with a still you know a correlation coefficient of 7 that’s the best we can get but by
Using just the textural information of the Airborne images you can get a coefficient of coration which is close to 0.9 so there is something contained uh in the texture of the very high resolution Airborne photos that the model can learn with some kind of activation to produce biomass at plot
Level and this is very promising because if we can get access to the plot data with their accurate geolocation we have a lot of high resolution images and we could make basically a national forest inventory with a very very high accuracy much higher than everything that exist so
Far um so in the end I don’t need to convince you that we need an integrated Forest monitoring system that artificial intelligence is very exciting uh we have big data from satellite but we are still in a world where the product you can derive from
Satellite and AI are still as good as uh the quality of the ground based data you have to evaluate them so we are still very much dependent on the availability and the quality of the groundbased measurement maybe in one decade the AI will do the job of forest inventories
And we don’t need Forest inventories anymore but we are still on The Fringe of relying very much on accurate uh ground data so conclusions there is a dele of satellite data from different sensors that no research group no post do no brain can make sense of uh previous physically based retrieval
Models uh because you have different wavelengths different uh observation system it’s more and more difficult to use physical retrieval models to achieve a fusion of this data this is the reason why some people have taken advantage of deep learning to try to move deep learning models into Forest monitoring
Systems uh we think that Tre level data is made possible Now using very high resolution imager or Airborne data and as I said ground based data remain critical for the evolution of satellite product and I want to say it’s not a call for arms but uh that if you
Have plot data since we now have all these annual change of of height we see the degradation we see the growth we would be very interested to analyze the mortality and Recovery with your help to try to understand better the carbon balance of forest so thank you very much
For your attention I hope I was not too long Philip thank you so much I think I speak for us all that this was quite impressive and uh really uh the the density and the the the amount of work that you’ve been doing on this is really just amazing so really really impressive
There have been a number of questions coming in and I will try to digest and relate them to you with the help of my colleagues from the triality network here the the first group of questions sort of clustered around the use of Jedi and particularly the limitation of Jedi
In the high northern and southern latitudes so when you do your Global U mapping and your Global products using Jedi how do you deal with the with the uh with these the limitation and this came both from Rico Fisher and Eric willan it’s a good point I was a bit
Quick is that Jedi is no data north of south of 52 degree so we just train a model where we have Jedi data and because we have Sentinel data everywhere we can produce a global map in Canada and Russia but how good it is when we basically uh got out of the model
Training for application it is something we need to find out so we are now trying to evaluate the Jedi based map in regions where there is Jedi data and outside regions where there is Jedi data against a large data set of ALS data from our colleagues from Bristol in
Order to have a better idea of the performance of the model if the performance is a disaster in the high latitudes we can still use I set two and that’s what we would like to do is we would like to combine both Jedi and IAT 2 which is a photon counting instrument
In the generation of the maps where we have no Jedi data wonderful thank you so much um I guess well there’s a lot of questions I try to I’ll try to uh get to all of them but or at least most of them um one that
Comes from Tom and maybe Tom you can jump in if I’m not if I’m not relaying it correctly I mean you’ve what we’re essentially doing is mapping States but of course from the from the idea of for Dynamics the underlying fluxes so growth and mortality uh are the interesting uh
Constituent drivers and then you know looking at Satellite data particularly in in dense forests do you think that we can eventually get to disentangling these fluxes that that make up the signal that we’re seeing well depends on the flux if it’s a huge stand replacing disturbance uh we will see a very clear
Height loss if it’s a cryptic mortality event affecting only the under storage Tre uh it will be much more difficult uh we see degradation features in the Amazon we have this annual map at 10 meter for 5 years and when we look at the ation features in Mato like
Selective logging or things like that or ground fire we see a drop of height so this is detectable I guess if we want to do better we can use the planet data there was a nice paper by Ricardo dalol over the Mato where they use the planet
Mosaic at 5 met the niik data and they were able to detect and to classify all the degradation features at least for the anthropogenic degradation ones uh we we can also exploit the Jedi waveform because in everything that I showed we were going for the low hunging fruit to
Predict the maximum height if we try to use deep learning to predict like rh40 it doesn’t work very well because the Sentinel are seeing the top of the canopy and the Jedi is measuring the depth of the canopy so what we are trying to do now is to build a not deep
Learning but it’s called contrastive learning you can try to look at the similarity between the texture of the Sentinel data and the complete shape of the Jedi waveform so instead of trying to regress height at each level you try to make some similarity or dissimilarity classification between the waveform of
Jedi and the Sentinel images I don’t know if it’s working yet but if it works it’s very promising because we can then model the entire waveform of Jedi which is much richer in term of information on the canopy depths to infer the biomass and the biomass changes than just the top
Height wonderful I mean this would basically allow you to go at structural changes as well I mean if you have the top canop be dying but sort of an a second layer persisting in the disturbance or the other way around exactly we would love to be able to uh
Accurately reproduce the Jedi waveforms uh but I don’t know yet it seems to be promising but I don’t know yet if we’re able to do it as well as for the top height I I have a related question to that one um you said it depends on the
Type of the disturbance one thing that we struggle with or struggle with that we encounter quite a lot in the European context is that a lot of disturbed areas are Salvage loged so basically you have a disturbance and then the managers come in and they put another human
Disturbance basically on top of it and that makes it sometimes really hard to disentangle what’s human what’s natural what type of natural disturbance was the trigger in the first place how how do you address this in your work in France or Europe or maybe even we are facing
The same problem it’s also a big political problem because if you say your harvested area increases okay it’s because you take more wood than needed but if you say okay I need to increase the Harvest because there is bar Beal and this Salvage logging it’s a very different political story so the forest
Industry is saying uh it’s not us it’s the back Beatles and the way we can find out to let’s say deconvolute this is that if we have Sentinel data every five days we can see some precursors of disturbance like Bol signal is relatively easy to detect which is the you know near
Infrared band and then we will see eventually that uh the forest has been marked to cut and it is cut because in France they Mark the forest to cut like the trees are damaged but uh they only have the resources and the manpower to cut it maybe two years after so they
Don’t cut like firemen you know immediately after there is a bar be detection they cut when they go to the place and they know that this place has to be cut so by looking at the high temporal frequency I hope that we will be able to see this cascading effect of
Disturbances uh but it’s quite challenging for instance I was in the field and I learned that uh I thought that when the beol attack the trees become red and brown but in fact if the beol attack in the fall they stay green and it’s only when the sa wants to go up
In the next spring that the next spring they become red and brown so the spru trees can still look green and healthy if the beors have attacked them in aurn but they only start to show sign of you know being killed a few months later so I start to realize the complexity of
This the thing that you guys know very well and we need to be quite careful in this kind of temporal detection signal yeah wonderful thank you I mean we’re basically out of time but I think there are so much more questions that I would uh take some more of those and I
Go a little bit over time if that is okay with everybody um there’s a number of technical questions and I i’ start with one that I’m also interested in uh from Power arano who asks about more specifics on the machine learning models that you use so which type of machine
Learning and probably also which types of Frameworks you guys are using in the work that you’re doing uh well as I said we use unet model so unet uh takes an image it makes a series of convolution uh to reach some kind of latent images and then it goes back to
Reassemble the latent images to match the data with the loss function uh and then of course it adjust the weight of the convolution which are like the coefficient of linear regression if you had a linear problem unet is nice and flexible but we have been uh finding better performances with the vision
Transformer model there was a group from meta from Facebook who also use Vision Transformer with very high resolution radar data in the tropical forest Vision Transformers have less of a problem in the underestimation of tall trees but you seems to be quite satisfactory although I’m not an expert of all this
And then this contrastive learning basically contrastive learning is methods used for facial recognition to recognize cats and dogs if I ask you to distinguish a cat and a dog if you look only at the details like the shape of the ears unfortunately some dogs have a
You know sharp ears like cats so you cannot just look at details but you can transform the image into something that looks similar like the shape of the noise and the ears and then you can associate things as being similar or dissimilar and that’s what we try to do
Uh to basically classify the Jedi we form as a function of the Sentinel properties so we use unet we are happy with unet but there are you know better models it’s always a tradeoff between computing power and uh performance without overfitting wonderful another more technical but it’s also interesting is
From Martin Turner and he asks uh how do you separate the influence of the the influence of moisture from the biomass signal in the VOD yeah yeah that’s we have been killed by Rie since we started to publish on VOD because he’s right VOD in a sense
It’s a water content so uh we can try to do some statistical models to use water proxies like lswi to try to get a part which is biomass and the part which which is water content the other thing we did is that we take the difference of
VOD from the wet season to the wet season like you are in the Amazon and you think that during the wet season there is so much moisture that the trees have reset completely their water content so if you make the difference between a wet season and a wet season
You have something which is close to Delta biomass rather than Delta water content so my question to you guys is that okay it works for the Amazon because there is always a very wet season but is there some regions where during one year the trees do not restore their complete water content so
They shrink and then it’s only one year later then they get back to their maximum water content if this is the case of course then the VOD would be affected by some problem if there is interannual changes in water content because we don’t use a seasonal signal we use the interal
Change uh the other problem with VOD is that when you have a disturbance like bar beetles or fires what is the water content of the dead trees I looked I found one paper from Canada and I thought that the dead trees they are black they are dead they are not moist
In fact they have 37% of moisture so that trees keep the moisture there is no sap but they keep the moisture so all the radar measurement should also be sensitive to the biomass or the moisture contained in the biomass of dead trees so it’s very difficult using I think
Radar or VOD data to separate uh the biom which is dead from the biomass which is alive because they both contain water and honestly I don’t have enough knowledge on this but this can be an issue wonderful thank you Philip um I I think there’s there’s two more things
That I that I want to pick out and apologies to all the participants who whose questions we couldn’t address because I mean there’s so much great stuff there that Philip presented and I’m sure there’s so much questions we could unities for long but there’s two more things and I think one that is
Interesting for a large majority of people and a Milo Viano asked this uh um is the hyp map publicly available and I think the question that is more broadly I mean you know the stuff that you’ve presented I mean is this available for the community to the height map for front is available
On zenor the global one we need to evaluate it against the ALS and as soon as it’s publish in the data paper it’s globally freely available the planet data from Martin Brown group it’s more of a problem because planet is a private company and basically they forbid you to
To sell the derivative of the derivative of the third derivative of their product so the map of Martin Brun are available on request for research which doesn’t mean that he’s defensive that you send him an email you get the map but planet is careful that he doesn’t give the map
For any commercial application because the agreement is signed to with planet to have a cheap price basically prevents him from making the map really available on Z noo repository otherwise all the data from Martin br’s group paper are really freely available on request and
It’s not a trick to not to give the data like sometimes it’s they are really freely available and and and the last one is more of a comment but I think it’s it’s an important one also because you’ve you’ve stressed the value of inventory and it’s from uberto fros from the
Brazilian NFI uh who basically says that with the approaches that you presented what we can basically do is uh reduce a little bit our densities in the NFI but Focus also some of the fieldwork on on aspects that probably remain still hard to uh measure from uh from remote sensing and basically
Develop you know a different type of inventory that’s informed or there is a feedback between the remotely sensed uh uh monitoring and the inventory that we do on the field in on the ground and I think this is a well at least you know personally as somebody also that I will
Be out in the field tomorrow looking at some of our field plots in the snow this is something that I find that I find quite quite uh interesting and relevant and so I’d be interested also to hear your thoughts probably as a concluding remark on this relationship between you
Know what we can do with remote sensing artificial intelligence but then also how the the the boots on the ground work fits into all of this well uh I think of course the remote sensing has such a good coverage that when the data of good enough quality they provide the better
Budgets it’s clear but uh on the other hand uh it’s easy to produce a very accurate height map but if I show you the accuracy of my biomass map the coration is2 and the height map is 7 so the allometry in a forest between top height and biomass is dramatic so if you
Don’t have Forest inventories with the different structure of forest to have a very good conversion of height to biomass uh you’re not going to do biomass just from remote sensing uh so we I would argue that we can do height very easily and now I believe accurately
But to go from height to biomass especially in tropical systems uh you need absolutely the inventory plot because there are so many combination that nature can find between height and biomass uh that the same height can have completely different biomath that in my opinion is super important to keep and
Even to reinforce the uh the the field measurements wonderful one problem also is the density of the wood because even when you have a very good volume okay you have a perfect inventory with the TLs and so on you are very happy you shake hands with the satellite guys you
Have solved the problem and still you have wood density That Vary between 04 and 0.9 so then you are still stuck to do the biomass wonderful and just open up another dimension I think that’s the beauty of science right I mean we’re not running out of questions anytime soon
Regardless of the big leaps that we are making and particularly Philip you and your group have been making thank you so much for making the time I I’m well aware that your presentation schedule you very well in demand and your schedule is quite dense thanks for
Making time for this I think this was a wonderful seminar thanks everybody for tuning in and I’m looking forward to seeing you at the next seminar and stay tuned for more goodbye byebye