Agroforestry in the Alas Mertajati of Bali, Indonesia
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 | 266 | |
| 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 | 10.5446/66342 (DOI) | |
| Publisher | ||
| Release Date | ||
| Language |
Content Metadata
| Subject Area | ||
| Genre | ||
| Abstract |
|
FOSS4G Prizren Kosovo 2023252 / 266
10
17
23
44
45
46
47
48
49
50
52
53
80
84
85
91
110
116
129
148
164
167
169
173
174
181
182
183
186
187
199
202
204
206
209
215
241
248
265
00:00
Scale (map)Flow separationProjective planeStudent's t-testExploratory data analysisCollaborationismComputer animation
00:30
Category of beingForestSheaf (mathematics)Source codeAreaCategory of beingImage resolutionCondition numberPoint cloudProjective planeContext awarenessFrequencyOpen sourceTemporal logicExploratory data analysisWebsiteDegree (graph theory)ForestSatelliteCovering spaceMoment (mathematics)Constraint (mathematics)Multiplication signMedical imagingObservational studyBitRemote procedure callComputer animation
02:42
Mathematical analysisProcess (computing)Group actionAreaMereologyRight angleCondition numberObservational studyMoment (mathematics)Covering spaceMultiplication signWebsite
03:51
Machine learningRule of inferenceInformationMoment (mathematics)Process (computing)AreaProjective planeVirtual machineLevel (video gaming)MereologyContext awarenessDifferent (Kate Ryan album)Integrated development environmentVotingMultiplication signDependent and independent variablesGroup action
05:07
Virtual machineMultiplicationSoftware frameworkMachine learningSatelliteIntegrated development environmentContext awarenessForm (programming)Set (mathematics)ResultantCollaborationismLibrary (computing)Product (business)Different (Kate Ryan album)Shared memoryPoint cloudParameter (computer programming)
06:40
Musical ensembleRight angleArchaeological field surveyForestElectric generator
07:07
Type theoryForm (programming)Level (video gaming)Network topologyDifferent (Kate Ryan album)Electric generatorCategory of beingMetric systemData structureMultiplication signPanel painting
07:57
Configuration spaceMusical ensemblePoint (geometry)Instance (computer science)Latent heatSatelliteMetreEvent horizonMapping
08:37
Matter waveGreen's functionSpectrum (functional analysis)Dependent and independent variablesProjective planePairwise comparisonComputer animationPanel painting
09:10
PlotterPairwise comparisonMusical ensembleAngular resolutionRevision controlSatelliteProjective planeMotion captureDesign of experiments
09:37
Mixed realityPoint cloudForestGrass (card game)Virtual machineCategory of beingHookingImage resolutionElectronic signaturePhysical systemInformationMusical ensembleRight angleMachine learningArithmetic meanComputer animation
10:11
Mixed realityForestWorld Wide Web ConsortiumSupport vector machineComputer animation
10:28
Musical ensembleCoefficient of determinationNetwork topologyElectric generatorComputer animation
11:23
ForestSupport vector machineComputer fileMixed realityField (computer science)MereologyRevision controlMusical ensembleComputer animation
11:40
ForestMixed realitySupport vector machineSupport vector machineVirtual machineMultiplication signVideoconferencingRevision controlResultantPerspective (visual)Category of beingParameter (computer programming)Computer animation
12:32
Different (Kate Ryan album)Electric power transmissionMassCategory of beingRight anglePopulation densityIntegrated development environmentMatrix (mathematics)Type theoryForestLine (geometry)Level (video gaming)Contingency tableTerm (mathematics)ResultantWeightStructural equation modelingCovering spaceRandomizationSupport vector machine
14:18
Right angleWebsiteProcess (computing)Error messageVirtual machinePhysical systemCovering spaceGoodness of fitRandomizationForestCategory of beingType theory
15:17
Source codeSupport vector machineType theoryGame controllerPoint (geometry)Different (Kate Ryan album)Error messageCovering spaceMappingPhysical systemMachine learningOperator (mathematics)Support vector machineLevel (video gaming)Virtual machineMetropolitan area network
16:25
AlgorithmAdditionMetadataRight angleAreaComplex (psychology)
17:03
Multiplication signAdditionInformationAreaSystem callGraphics tabletGroup actionService (economics)
18:07
Level (video gaming)Graph coloringProduct (business)Physical systemInformationMultiplication signMappingMixed realityField (computer science)Type theoryIdentical particlesSource code
18:39
Arithmetic meanResultantHookingMappingMusical ensembleMultiplication signField (computer science)AreaPosition operatorInformation
19:26
Installable File SystemMappingLevel (video gaming)Standard deviationWater vapor
19:58
WebsiteWater vaporRight angleField (computer science)Musical ensembleLevel (video gaming)
20:29
Software frameworkMathematicsLevel (video gaming)Arithmetic meanWater vaporComputer animation
20:56
Source codeMultiplication signGroup actionLevel (video gaming)VideoconferencingMereologyProjective plane
21:22
Computer animation
Transcript: English(auto-generated)
00:08
Thank you For having me here. This is a collaborative research project exploratory research project between the Indonesia Institute of Sciences and my home institution University of Buffalo
00:22
Rajiv and Tony are members of this team and several graduate students and so What we want to do is an exploratory research project in GOAI Specifically for the majority world
00:41
Resource constrained conditions. What is AI going to do expensive AI going to do when you apply it to? Topics in emerging economies and we have a study site in Indonesia and bodies particular Beautiful Island and body that some of you might know so we've been looking at Bali
01:01
for two or three years now I'm working with Satellite data specifically from Planet Labs and I'll get to the Justification for that and then open source conference in the moment But what we started to do was first look at what kind of land cover conditions can we actually
01:21
Detect with this new asset for those of you who are not familiar with Planet Lab imagery It is a high resolution imagery at a very high temporal Frequency yet you get a bit of a fresh image every day and that's meaningful in the tropics because of the cloud cover so if you work with a Sentinel-2 data where you have a
01:45
Revisited frequency of at best five days And you have clouds 80% of the time chances of getting a useful image is very small. So this is a new opportunity amongst other reasons for that particular feature
02:02
And so our first project was as it says you're looking for tropical forest categories In the highlands of Bali and this has simply not been done before so Indonesia has mapped the archipelago 11,000 islands Quite well existing a juice methodology, but they haven't used high resolution
02:26
Remote sensing to the degree that you could to the degree that we do in Europe and in North America So what happens when we apply these techniques? to Emerging economies and when we then put this into the context of AI classification systems
02:43
And so in particular to contextualize the kind of like the the reason for the study site. This is the Allah's matter Jati it is a area that is home to the tumbling on an indigenous group that has been arguing for
03:00
Land use rights of this area with the Indonesian government for a long time and that process is ongoing and part of The work that I'm doing is collaborating with a local NGO this new To find ways how we can document land
03:21
use historic land use practices based on land cover analysis And so this is meaningful because of the way the tumbling on have been using the land they have techniques of sustainable agriculture that are becoming very prominent now because they're resistant to
03:41
change in climate They've been adapting to low rainfall or high rainfall conditions in interesting ways and I'll show you that land condition in just a moment and so here is How the business foundation which is a local NGO that works with indigenous groups to to map areas and
04:01
Define problems environmental problems most of them in response to Land disputes oftentimes fueled by the tourism industry And so these new is our local partner so my team is kind of like the science part and we work with this new to interface with all the
04:22
Complicated Indonesian rules of how things have to happen and then we work with the tumbling on directly Who we confer with for? Piece of information just to understand what their concerns are and another part here that I'm trying to integrate in this project is to think about G oai as a project that is nascent right and do we have the opportunity to
04:45
Think about how Large-scale machine learning is going to happen differently if we bring it into different contexts And so so it's slow in that regard and it's purposely slow because I want to take the time see what happens as I'll show you in a moment when you
05:01
Look at different constituencies and different stakeholder interests as you go through the machine learning process and so What I built Kind of like haphazardly just because I didn't have other tools that seem to do everything I wanted was a little framework It's called cocktail that integrates headless QGIS or fail machine learning library G doll
05:26
P cloud all in a Python environment just for collaboration. And so it's a machine learning environment that we can train classifiers and We can work remotely so I have someone in Indonesia working and the u.s. Working and then they travel
05:43
and We can compare results. So one of the things that's important when you can machine learning is yes How did you set your parameters? For example, I want to replicate your your results, right? You have like 20 problems. So so we kind of like share that those brometer settings in a way that we can replicate results
06:00
We work with multiple forms of satellite data so we can ingest The Sentinel data or a lot of data and then importantly we take the Products we produce outside of the machine learning framework So it can be quite expensive to maintain a machine learning training environment I mean literally expensive is like these are high-end machines. So if you produce lots of data you want to move it out
06:25
You will be billed for it Otherwise and so to have a feature this moves things out seems trivial but makes a big difference If cost is an issue and if you're working in emerging economies and in education context cost is really important
06:40
Okay, so the first survey we did with the first generation of Planet Labs data, which was four bands RGB and infrared Allowed us to differentiate Forest cover quite well, right as I saying this is not perfect But it was better than anything that I've done before our goal
07:00
However was to look at these sustainable and loose practices and we were not able to do that with the four band data What we wanted to do was map agroforestry Agroforestry is important because it is a sustainable form of land use that is robust to climate fluctuations as I mentioned and has been used by the tumbling gun for generations and
07:24
So it's an interesting Category to capture but it's very difficult because it's very varied and so in showing you here just three three types the topology of agroforestry that that I found it and they basically vary by
07:40
Intensity of Structure if you just go by the vegetation, it's very hard to say what the difference is but if you go according to a metric of intensity of how the land is used some of them are Hardly managed and others very strongly imagined almost like the garden you can kind of like it gets some grasp on it and so last year
08:01
2010 to Planet labs came up with a new constellation of satellites that had four additional bands So we have a configuration that maps every point on the planet every 24 hours offers eight bands Specific events that are useful for
08:23
Vegetation and has a spatial resolution of below four meters bingo We're able to use this We're able to get instances of agrofoam now So this is this asset that the planet came up with last year. This is really new
08:46
And I was lucky to have access to that and so we could go back to this agroforestry topic that we were And Due to inability to actually capture one of the most interesting features of the landscape and one of the politically charged features
09:07
Do you want me to do something? Okay, and so I'll just show you here like some comparison The Sentinel to versus doe that was the four band version of the satellite
09:22
Planet lab data and then super doe and you see it in the paper better Really bad projection here, but the spatial resolution is such that we can capture Larger plots. All right, so it's like bingo we can we can get started here And from the spectral resolution you see kind of like we're starting to get spreads here in some of the bands meaning
09:45
We have enough information So said a machine learning system can latch on to something right this about just does do you have a signature in your information? That machine learning system is a statistical organizing machine right can latch on to and so and so we find that and so
10:02
We can go from the existing categories that we have Over to agroforestry, right? So so there's a technical hook into this so I'm going to Just show you again what?
10:23
This agroforestry business looks like give you an idea so agroforestry is really think of it as a three-dimensional
10:42
Okay, so lots of dogs So it's a three-dimensional garden so you have at the top you have clove trees then you have coffee plants and then you have Taro or fruits? No, it's fine or or Albasia flowers and and this all grows together and depending when one of them
11:03
Flowers or blooms you get that fruit or this fruit is constantly going everything is always growing It's not like a growing season just always growing always harvest something an agroforestry is robust it works with the climate it works with the knowledge of the people and And has been in place for for for generations. So this works
11:26
Okay Okay, so now this is kind of like the boring technical part We had to make so many versions of our data collection So I had people go out into the field to collect data
11:41
With videos send it back. We evaluate. What is this? Is that a garden is that agroforestry what and then I went down myself for two weeks to kind of like augment that Dataset after we had arguments. What is this is a turned out, you know, like version 13 But we start to get something where we can with some reasonable results
12:02
Get agroforestry at the same time the mixed forests and at the same time the rice patties So we were able to optimize any one of those categories, but we had a hard time doing all three of them All right, and so so there we go. That's our kind of like result. It's not good It's not perfect from a machine learning perspective But it's the first time that's been done is 62 percent accuracy
12:21
F1 score rather for for agroforestry for a category that has not been mapped in Indonesia before So and this is with support vector machines. I'm going to show you these the results that we have. So this is The first map of the Alice Marta Jati with live SVM and
12:42
13 land use land cover categories Little things here quite interesting these lines here. For example, that is actually a high voltage line covered in the forest in a protected forest as a whole story behind that and And so one of the things we found this was actually difficult to distinguish
13:01
If we used a low dimensional neural network and get to the differences later on like it actually misclassified that human intervention as agroforestry, right and so One of our one of the topics that we were looking at which is the best classifier, right? So when you think about the best classifier, you think about okay, which one has the highest f1 scores
13:23
Which performs best on a confusion matrix and so we wanted to see also what type of? Contingencies to these classifiers required like so for example Which one can we use with the lowest amount of data because data is hard to get right and neural nets are notorious for
13:42
requiring vast massive amounts of data So if we can use a classifier that is that works almost as well with much less data that is an interesting candidate for Resource constrained environments on top of that the neuron that just didn't quite get Its act together if you saw one with the settlements, so it tended to
14:04
overestimate the density of Settlements so for those two reasons and because it's too expensive in terms of data collection we moved away from shallow neural nets and continued to look at Random forests and support vector machines now
14:23
So that's kind of like where the topic and then when it stops we have a first Land cover map of Allah's matter Jati bingo good. Okay, great hasn't been done now What I want to talk about are the types of errors that we encountered and how? These errors are interesting opportunities to think about how we want to make machine learning happen
14:44
Right because we can tweak these systems And so one of the places where we started to look at the error categories was around this Lake. This is Lake Tumblingan It's a it's a sacred lake to the to the Tumblingan and there can be no economic activities in this lake, right?
15:01
They are in the process of arguing with the Indonesian government of how to manage that There are tourists who come here they leave a mess the Tumblingan are very concerned about this and They want to make sure that when they represent the land that this is a sacred site. So if I give them a Lance fancy AI land cover classification that shows rice patties
15:25
I've got a problem. I've got a problem with my eye. I have there's all kinds of slippages but this is a particular type of problem because it's Problematic for the stakeholders who are supposed to profit from this mapping operation. So
15:42
With The SVM approach we can simply take a few extra data points around the lake Run the classifier again and this particular type of error disappears. So we purposely Unbalanced the data which you're not supposed to do in machine learning We purposely unbalanced it to address one particular type of error
16:03
That we favor over other types of error Because of who the man is supposed to serve And I think that's while the whole details don't necessarily matter this notion of that We want to take these machine learning systems that otherwise would just run off do whatever they want We want to take control over them and taking control over these systems requires intervention in different places
16:27
Then I want to show you some of the problems we have with getting ground truth So this is an area of unfortunate fog, but it actually has the vegetation of agroforestry But it's inside of a hotel complex Right, so so it has the look of agroforestry, but it doesn't have the use now that the bad thing is that
16:48
Algorithms are blind to that right? So the algorithm even the best one is going to classify this as Agroforestry, so what kind of additional meta information do we have to add to address this? This is an open question
17:03
Likewise here a rice paddy that has now been abandoned and is going to be a construction site This is in January. So there's all kinds of like Transitional areas that that the classifier in its attempt to be very strict about what is what?
17:22
Makes calls on that actually don't coincide with the realities on the ground Moreover you have the rice patties Of course the famous rice patties of Bali that people go to see and see the pastoral landscape and how the farmers work the land And a lot of times for farmers, it's no longer Economically viable to produce rice. It's actually better to work in the service industry. So sometimes
17:46
Rich landowners will will pay farmers to maintain the rice patties Not for food production, but for the tourists for the tourists just to look at so here's another Misuse of land use if you want that
18:02
AI will never capture unless we find out how we can add additional pieces of information. Okay, so back to my kind of like products if you want we presented this to the tumbling gun and They loved it. It's great. This never happened, but they had a hard time seeing all the details right in this mix
18:22
so there's a so we make a product is AI produced system and If you're in the field you can read this but if you're not you actually have a harder time and one of the problems was they couldn't distinguish is this a photograph because of the color mapping sources a map or but it's got confusing so What we then tried to do was come up with ways
18:42
This is much better in the paper to look at is how we could Separate out some of the categories and put them in a neutral background meaning in the infrared band of the particular image such that one At a time would pop out and so here's the settlements of 20 I think it's 2020 here now the agroforestry of
19:02
2022 and when we did that We were much we were in a different position to communicate the results and we could start to talk about wait a minute This area here. Is it really that dense? I want to talk about it. We go back out in the field and check so this was the hook that allowed us to make our Information our results more meaningful to to the stakeholders and that's that's an interesting
19:26
Result, and then finally I want to spend just one minute if I have that to talk about bad maps in emerging economies Badness, so this is a hydrology map produced by the Indonesian government the year 2000 and it's the standard map They did all of Bali in that year
19:42
It's the standard map that people use to this day to demonstrate the water richness of the island But the problem is Bali is running out of water. This map is a fiction it doesn't it's this is not it doesn't represent reality anymore and and I know this because I Spoke with the tumbling on it says like what is the water situation says it's bad
20:04
and so what we did is we spent the day just going out into the field and looking at some of the Looking at some of the sites where they're supposed to be running water. This is during the rain season And they were dry and this was just a spot check out of I want to say 15 sites
20:24
20% were dry Right and so and so what I want to end with is this idea of of the bad map of the map That's an official map by a government in an emerging economy where you don't have the means to just update this right away and it persists and
20:42
Becomes a framework that that limits the opportunity to respond to these changes that are happening the lack of water That way differently when you don't have the resources to change it quickly So what we're doing now is with the tumbling on with the use of the tumbling up
21:00
We're doing a kind of like group action to go and find as many You know of these sources that we can take a video in the rain season and in the dry season We can make a new a new little map and then here's the other part of the project government now, okay Thank you very much for your time