Modeling and prediction of wind damage in forest ecosystems of the Sudety Mountains, SW Poland
This is a modal window.
The media could not be loaded, either because the server or network failed or because the format is not supported.
Formal Metadata
| Title |
| |
| Title of Series | ||
| Number of Parts | 57 | |
| Author | ||
| License | CC Attribution 3.0 Germany: You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor. | |
| Identifiers | 10.5446/55258 (DOI) | |
| Publisher | ||
| Release Date | ||
| Language | ||
| Producer | ||
| Production Place | Wageningen |
Content Metadata
| Subject Area | ||
| Genre | ||
| Abstract |
| |
| Keywords |
00:00
Maxima and minimaPredictionGamma functionElectronic meeting systemIntegrated development environmentTesselationAssociative propertyMathematical modelMoment (mathematics)Reading (process)Projective planeArithmetic progressionPredictabilityMereologyForest
00:47
SpeciesDifferent (Kate Ryan album)MathematicsProcess (computing)ForestProjective planeBitCASE <Informatik>Complex (psychology)Network topologySoftware developerMereologyTerm (mathematics)Data managementVirtual machineForcing (mathematics)FrustrationCore dumpData structureMessage passing
02:30
Asynchronous Transfer ModeIntegrated development environmentLevel (video gaming)ForestDifferent (Kate Ryan album)Virtual machineVector potentialMathematical modelComputer animation
03:04
AreaSelf-organizationMereologyDemosceneSpecial unitary groupAreaNumberConnectivity (graph theory)System callMetreObservational study2 (number)PlotterEvent horizonRange (statistics)InformationForestComputer animation
04:13
ForestForestPlotterMappingMereologyMathematicsDigital photographyGraph (mathematics)Moment (mathematics)Constraint (mathematics)Forcing (mathematics)Computer animation
04:42
Set (mathematics)ForestMathematical modelPerformance appraisalPredictionObservational studyVector potentialIntegrated development environmentGame theoryMIDIMaxima and minimaDemo (music)Thomas KuhnArrow of timeAverageInformationCategory of beingComputer clusterFirst-person shooterMathematical modelWind tunnelOperational amplifierPlastikkarteSpeech synthesisQuantumState of matterForcing (mathematics)NumberNetwork topologyNeuroinformatikIntegrated development environmentGrand Unified TheoryGoodness of fitProjective planeMetreExecution unitElectronic mailing listVirtual machineSlide ruleGraph (mathematics)Uniqueness quantificationCubic graphCybersexForestProcess (computing)SpeciesMappingComputer animation
07:16
Network topologyVolumeFunction (mathematics)Price indexMathematical modelDependent and independent variablesGame theoryVariable (mathematics)Set (mathematics)Absolute valueWave packetSoftware testingAdditionMathematical modelPlotterValidity (statistics)Pearson product-moment correlation coefficientState observerCross-validation (statistics)NumberForm (programming)Musical ensembleResultantPoint (geometry)TwitterRule of inferenceComputer animation
08:47
Random variableBinomial heapSupport vector machineForestNegative numberOperator (mathematics)Kernel (computing)Computer networkMathematical modelProcess modelingoutputGradientBootingInformation securitySoftware testingResultantCurveForestPower (physics)Virtual machinePerformance appraisalMathematical modelAreaRandom variableMultiplication signReceiver operating characteristicNumberOptical disc driveComputer animation
10:08
Price indexFunction (mathematics)Network topologyVolumeIRIS-TVariable (mathematics)Graph (mathematics)Volume (thermodynamics)Rule of inferenceVirtual machineComputer animation
11:02
Plot (narrative)Network topologyVolumeFunction (mathematics)Price indexSurfaceSound effectSound effectGraph (mathematics)Variable (mathematics)Social classThresholding (image processing)Volume (thermodynamics)Subject indexingRhombusDiameterComputer animation
11:57
ForestMathematical modelMappingProduct (business)Different (Kate Ryan album)Random variableForestVirtual machineMathematical modelGraph (mathematics)Roundness (object)
12:55
Network topologySpeciesVolumeDistribution (mathematics)Population densityPlot (narrative)Boom (sailing)Mach's principleParameter (computer programming)Population densityMathematicsSocial classMultiplication signForcing (mathematics)Presentation of a groupCollisionRow (database)Computer animation
14:27
ArmMathematical modelGraph (mathematics)Shared memoryPower (physics)ForestVolume (thermodynamics)Virtual machinePredictionRandom variablePanel painting
15:05
Reading (process)Local area networkProbability density functionWindowObservational studyNormal (geometry)Formal languageData conversionSign (mathematics)Term (mathematics)Right angleComputer animation
Transcript: English(auto-generated)
00:10
My name is Luca Stavnik and I currently work as Associate Professor at the University of Silesia in Poland. And this is just a small project developed in cooperation with Professor Sandra Harrison from the University of Reading in the UK.
00:30
And the project is still in progress, so we hope to improve it in the near future. This is a modelling and prediction of wind damage in the forest ecosystem
00:42
of the Sudeti Mountains in one region in Poland, the southwestern part of Poland. So first of all, I'd like to tell you a bit about the motivations. This region is particularly susceptible to forest damage due to strong wind.
01:03
And one of the reasons for this particular susceptibility is complex forest history. Forests were largely changing this part of Poland due to human impact. First it was deforestation for agricultural purposes and then in the last two centuries
01:26
it was intensive forest management different practices, which resulted in changes in forest structure. So basically, this region's process was introduced in a large amount and is dominated for us.
01:46
This tree species is dominated for us. So basically, in this project we've managed forests. And this project is linked with the concept of the bio-morpho-dynamics. We try to look at how living organisms, and in this case trees, how they impact geomorphic processes on fish soaps and soil development.
02:15
And the last thing, no past attempts have been made in terms of forest damage more than using machine learning methods.
02:31
Some objectives. First, the most important question is where are forests most susceptible to hurricane winds of different origin? And, of course, it would be really useful to find potential strong predictor of forest damage caused by wind.
02:52
So we tested several machine learning methods and we selected the best models and then we tried to build a map of damage and probability.
03:05
Here's our study region. This southwestern part of Poland is Sudeten mountain range and part of this mountain range is on Polish territory, and it is a part of Bohemian Massif.
03:20
As you can see, the conifer forest dominates this area, especially in Krosys, a large part of this area. And some important information about climate, especially wind climate of this area. Here on this plot, you can see daily wind speed, so you can see some seasonal component, but it is even better visible on this plot
03:47
where you have number of hours per day with wind speed over 10 meters, over 10 meters per second. And here this seasonal component is even better visible. Basically, the highest wind speed is
04:04
winter season, it's new winter season, and it is correlated or related to atropical cycle events. Here you can see some examples from Thierry Jones. This is the highest part of the Sudeten mountain, the Kalkonosian mountain. You
04:22
can see some forest changes caused by strong wind and selected plots from 2001 to 2012 based on all the photo maps. And here another example from the middle part of the Sudeten mountains, and this damage was caused by the Kerry winds in 2007.
04:42
So our workflow was like combining three data sets. First one, forest features before damage obtained from forest days. And then based on unique ID number, we combine it with forest damage from 2004 and 2010.
05:03
And this information was in cubic meters per basic forest unit. And then we tried to fill it with some environmental predictors. And basically, bioclimate features from Charles' database, from a work claim that was in monthly wind speed, and then some properties.
05:25
We use shuttle rather topography mission product, and then we calculate some terrain properties in the South Nigeria. So basically, we tried to test it in four scenarios. First, we divided or we use all tree species versus honest process.
05:50
And then in another two scenarios, damage in cubic meters versus damage in cubic meter per hectare. And then these values, this information will transform to binary data, and we tried to use machine learning models to solve classification problem.
06:14
At first, mostly it was a corridor package, and then we perform maps in quantum GIS, and for some terrain features, we use cyber GIS.
06:28
And for some parallel computing, we use infrastructure from peer grid, because this is both computer clusters, and we use one of them with 24 cores, just to speed up.
06:49
So this is a list of machine learning models we tried to test, and some on this slide with some information about hyperparameters tuning. So we tried to try to apply grid search, in order to choose the best hyperparameters and then obtain the best model for our project.
07:17
Before model training, we removed the highly correlated variables, so when two
07:26
variables were highly correlated, so the absolute correlation coefficient was over 0.75. One of these variables was removed, but that was based on the relationship with other variables.
07:45
And then we split our data set to training and test set, but just for years 2007, 2010, and other years were used as additional test sets for model validation.
08:00
And depending on scenario, we could use from over 37,000 observations to 20 ,000 observations, and for model validation we use 10-fold cross validation repeated five times.
08:20
And here on this PCA plot you can see the final number and relationship between variables chosen for the first two scenarios, so we had like only 11 variables. And you can see that even after removing some highly correlated variables, still some of them were closely related.
08:49
Okay, now it's time for some results. This is for the first scenario, we use forest damage security test for all three species, and generally speaking,
09:02
this is a classification problem, so for evaluation of these models, we use accuracy and the curve, receiver operator characteristics curve, and generally speaking, the results are quite similar.
09:24
So for most of the models and accuracy, area under the curve is slightly below 0.7 or slightly over 0.7. And the best models, the best predictive power was offered by the random forest model and driving machine.
09:49
And in this plot, down here, you can see that the results were quite similar, and the best accuracy, the best predictive power was offered by random forest and driving boosting machine for the first scenario.
10:10
In terms of future importance, three volume and three edge were the most important features.
10:22
So, that was especially for, I mean, generally for all models, but especially for the random forest and for driving boosting machine, and some other variables like bioclimat variables were quite important.
10:44
For example, precipitation over the coldest quarter or Mention wind speed, and as you can see topographic variables were less important.
11:02
Okay, another slide. Using accumulated local effects plots, we can show how individual variables influence damage probability. So, as mentioned before, three volume and three edge is the most important variables, and damage probability increase according to this is
11:29
increased with three volume, three edge, but only for some edge class between maybe 70 years old, up to 110 years old.
11:40
And further damage probability increases with wind exposition index, topographic wetness index, adjustment, certain threshold level, and then valley gap. Okay, so the final product were maps of damage probability, and I'm not going to explain this in detail,
12:10
it's much, I think, much, much complicated, but what you can see that damage probability for years 1998 and 2016,
12:23
is much higher than for 2020. Those two maps are named based on the random forest model, and this difference between years is even more better visible for driving boosting machine.
12:42
Okay, so now the important question is why, why you have these differences, and we still try to find the best explanation, but maybe two reasons are important to highlight here. First of all, a proportion of spruces for 2020 was less, for the recent data was less than
13:08
for data between 1998 and 2006, but we know that spruces are more susceptible to damage to high winds. So it might be a good explanation that we have less, lower proportion of spruces, we can expect lower damage probability.
13:34
And another, another important argument would be changes in age classes, you can
13:44
see that was quite, on the density plot, it's quite well visible changes between, three age classes between previous, previous period, and recent data from 2020, and as accumulation of local, local
14:08
effectors show that damage probability was, was increasing, was increasing for three age class between 70 and 110.
14:20
And, yes. Okay, now it's time for some conclusions. We tested five machining models for four scenarios, and the most important predictors were 3 volume and 3 age.
14:44
Geomorphic and climate predictors were less important, and two models offer the best predictive power for random forest algorithm, and to that it was the same, and forest stand pictures might significantly influence the probability of forest damage.
15:05
Thank you for your attention.
15:46
Thank you. Yes. Yeah.
16:10
This is the
16:25
idea of course. Yeah. Right. Yeah, we didn't test it. In terms in terms of Yeah.
16:41
Yes. Yes. Yes.
17:01
Travis analyzes or Yeah, of course, I mean, yeah, I wanted to say that the next steps would, would
17:24
include some things we learned in this session so it would be great to do that.