We're sorry but this page doesn't work properly without JavaScript enabled. Please enable it to continue.
Feedback

CERES project- Earth Observation-based information for "smarter" agriculture and carbon farming

00:00

Formal Metadata

Title
CERES project- Earth Observation-based information for "smarter" agriculture and carbon farming
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
Publisher
Release Date
Language
Producer
Production PlaceWageningen

Content Metadata

Subject Area
Genre
Abstract
Branislav Bajat, Faculty Member of the University of Belgrade, illustrated the CERES project. In this talk, he explained how the use of AI in agriculture should be especially important in Serbia, as agriculture is one of the crucial sectors of Serbian economy. This project will be an important step forward in the application of a wide range of relevant data generated on a daily basis and offering a huge potential for improving agricultural production and developing the concept of smart and regenerative agriculture.
Keywords
Type theoryData modelInformationImage resolutionSource codeInterpolationEvent horizonElement (mathematics)Vector potentialInformation securityData managementDecision theoryLinear regressionTemporal logicScheduling (computing)ImplementationEndliche ModelltheorieTime domainPhysical systemOperator (mathematics)Flow separationForestThomas BayesGame controllerPredictabilityDifferent (Kate Ryan album)Moment (mathematics)Endliche ModelltheorieMappingWebsiteLinear regressionProjective planeBuildingClosed setState observerCASE <Informatik>Virtual machineNatural languageRight angleObject (grammar)Centralizer and normalizerLevel (video gaming)Service (economics)SpacetimeKey (cryptography)Process (computing)SurfaceArtificial neural networkTask (computing)Traffic reportingSampling (statistics)Presentation of a groupHypermediaCore dumpEstimatorMereologyMetreGraph coloringMatching (graph theory)OpticsSystem callLine (geometry)Repository (publishing)Multiplication signArchaeological field surveyDegree (graph theory)ResultantDeterminantLink (knot theory)Workstation <Musikinstrument>DatabaseSource codePressureLogic gateLocal ringPhysical systemBitEvent horizonSocial classLibrary (computing)Machine learningOpen sourceProduct (business)Price indexInformationNeuroinformatikMobile appType theoryDecision theorySet (mathematics)Latent heatCartesian coordinate systemWave packetAngular resolutionIntegrated development environmentRow (database)Condition numberMeasurementSpeech synthesisImage resolutionChainGroup actionInequality (mathematics)Medical imaging1 (number)File formatInternet service providerObservational studyPosition operatorGoodness of fitTotal S.A.Wage labourCoefficient of determinationParticle systemForcing (mathematics)Order (biology)Special unitary groupTransport Layer SecurityMetric systemUniverse (mathematics)State of matterAreaElectric generatorDigital libraryMathematicsGeometryHarmonic analysisoutputGradientFreewareOpen setFundamental theorem of algebraVapor barrierInterpolationSystem identificationTheory of relativitySoftware developerInformation extractionValidity (statistics)Temporal logicRule of inferenceAlgorithmPattern recognitionCivil engineeringEDIFParameter (computer programming)RandomizationVector spaceCharacteristic polynomialWeightField (computer science)Point (geometry)Mathematical analysisAssociative propertyPhotographic mosaicFaculty (division)Context awarenessSatelliteAutocovariancePreprocessorStatisticsDomain nameCubeData transmissionWeb 2.0Adaptive behaviorAdditionComputational scienceForm (programming)Independence (probability theory)Scaling (geometry)Heat transferMaxima and minimaElement (mathematics)DivisorComputer animation
Transcript: English(auto-generated)
I'm coming from the same institution like previous presenters that means from the university of barrier the faculty of civil engineering and the department of geodesy and the informatics
serious project is a project on the national level founded by the science found of the republic republic of serbia okay it's a project on the national level however we hope that some of the results will be valuable on the broader level so that's the good opportunity
to present also these results in this kind of the workshop first i would like to briefly present some goals of this project but first it's very important to say that agriculture is one of the most crucial sectors in serbian economy and that sector implies more than 20 percent of the total
labor force in serbia and also it produces a significant part of total export the goods from serbia also again more than 20 percent so the main problem here is that the obvious decrease of rural population causes the lack of labor force in agriculture and also inefficient practices
makes it barely profitable for the most of the farmers in serbia one solution to improve this situation is to use a novel technologies that provides timely information relevant for the decision making in agriculture and that will significantly decrease production
costs and increase yields and the quality of agricultural products another opportunity to increase profitability of serbian agri-food value chain is a wider adoption of carbon farming practices which result in a carbon sequestration and in soil and to combat with the climate changes
taking this in mind the overall objective of theres project is to develop a set of geospatial information products affordable to any type of end users that will be designed to support information driven decision in agriculture and to increase profitability
of agricultural sector in serbia and also to support carbon farming the set of products first identification of the crop growth disturbance second one is a yield estimation then estimation of soil organic carbon identification of tillage activities
extracting information from web textual resources and finally higher resolution daily temperature and precipitation prediction that will be exactly the input for the rest of the projects out but anyway we have already very good results in in this field
all information and all those products will be generated by applying artificial intelligence methodologies to big geospatial data from various sources primarily from earth observation data that means comparing transmissions meteorological data the global soil data like
land gis soil grids and also from the web textual sources and also from other available data sets okay i will pass briefly from through all those tasks the first one is the prediction of crop growth disturbance events the pattern recognition and early detection
of crop disturbance will be based on the artificial neural network the algorithm will be automated and it will deliver the results without need of any visual inspection of images that's our idea the earth observation core data that will be used are derived from freely
available comparing to certain sentinel missions and the training data in situ data will be collected in the context of serious project but also potentially competitive in situ data and as well from the other european projects that means from the previous one okay the second
one task is yield estimation currently crop yield estimations are mainly done through the time consuming and very expensive field sampling which includes
the measurement of biomass weights and grain size we know that and our idea is that commonly utilized by physical characteristics of vegetation in estimated crop yield will be calculated from vegetation indices derived from sentinel to data so far as you know a universal model that will
be applicable for all crops types is hardly to be achieved we know that but our aim is to develop machine learning crop specific model based on biophysical parameters those vegetation indices rather and optical data and also metallurgical data the next step is
edification of tillage activities we know that carbon farming includes the cultivation techniques and of regenerative agriculture that take carbon dioxide out of the atmosphere where it causes global warming we know that and to convert it into carbon-based compounds
in the soil that aid plant growth within this objective we will present an approach for mapping tillage activity changes by using sentinel 1 and sentinel 2 data at high spatial resolution
but that means 100 meters and by applying temporal change detection algorithm based on artificial neural networks as a based machine learning classifier in this case the next task is estimation of soil organic carbon there was a really a lot of presentation
related to soil organic carbon but we know that building up soil carbon is one the key to achieving high yields without chemical inputs we know that
the surface reflectance and vegetation indices derived from the optical satellite imaging that means from sentinel 2 soil texture from the radar imaging that climate variables and terrain factors will be used as covariates in this case to model to build a predictive model of
soil organic carbon for that purpose the performance of the several different machine learning algorithms based on in on in situ measure data values will be tested with the objective to design the best predictive model for a particular site a local site three main
subtasks are related to a metric of soil organic carbon that means to obtain all available training soil data to pre-process all predictors of all covariates which are important for this
model all predictors should be downscaled or upscaled to 30 meters resolution and the the last one sub task is to automate the the generation of soil organic carbon maps from point data with associate prediction uncertainty by
again by using a state-of-art machine learning methods the next one task is interesting in this case and it is related to natural language processing we know that information about undesirable agricultural events like calamity drought
disease pest attack and also yield estimates are often published in the numerous local or statewide news sites agricultural reports also on social media and
such information if it is properly properly recognized could complement the available earth observation data when when building and validate and validating during the process of building and validating different estimation models therefore it is necessary to apply different
artificial intelligence methods which include natural language processing of Serbian language in this case because it's a product that's the product of national level again expert defined rules and machine learning techniques to extract and to classify the information about
relevant time and space determined events the fulfillment of the objective could be conducted for interlink interlinked tasks the first one is the identification and the classification of the available text sources the second one is development of related data
colors again information extraction again geocoding and finally application of the extracted information in building and validating different estimation models the last one task is high resolution to produce high resolution climatological grids for Serbia our aim is to
generate high resolution daily temperature maps from freely available temperature observation and by using geostatistical and machine learning techniques also those data will be also used for precipitation maps and for this purpose a new method for spatial temporal
interpolation based on machine learning we call them random forest spatial interpolation was already developed in the next step I will introduce some I will inform you about some
ongoing activities related to each of those tasks the first one was a prediction not each one but some of those tasks the first one was predictions of crop growth disturbance events each the moment we are working on a spatial temporal data cubes with the Sentinel-1 and
Sentinel-2 data also with meteorological data for each parcel libraries for data download the pre-processing are already developed and at the moment
final preparation of predictors and the harmonization of a different in-situ data has been performed in this case we prepared four years of analysis for Sentinel-1 and Sentinel-2 data together with the better data for northern Serbia because our case study will be
the Voivodina that means the northern part of Serbia where agriculture is dominant agricultural area in in our country we collect already 3 000 fields with diligent yield
information and some initial modeling was performed on the on the part of the data but we are still working with the baseline models the next interesting task is estimation of soil organic carbon we know that quantity of publicly available data has been growing
at a massive amount and that will continue to to do so in the near future but however we have a problem in Serbia with the limited amount of the available data especially reliable data
and this challenge will be how to implement the knowledge based on global models from the wider areas to the case study of Serbia as a possible solution the domain adaptation the transfer learning technique in which the system aims to adopt the knowledge learned from the
source domain and applied it to another related domain will be utilized in this case considering various sources of relevant databases on a global scale will be also called multi-source domain adaptation as a additional form of domain adaptation technique
and we will practically use more than one source to to adapt it and to compensate compensate compensate the lack of data in our case in Serbia we are still waiting
just look at this map we are still waiting for Lucas 2018 data for the west Balkan countries
and just to to test some models we started with the Lucas 2015 we are interested only in the in situ data related to cropland classes and here you have a map with all those samples
those Lucas data but crop uh cropland data cropland samples over the Europe you know there's some very short statistic in Malta we have only two samples related to the croplands but in Spain we have almost 2000 of the samples and
we use some additional covariates some of those are soil chemical properties which are also presented in the Lucas data
again metrological data some digital elevation products and also satellite observation this moment we use lansat mosaics also modest data and so okay the the coordinates will be also
involved in this modeling and we we just try to do with some basic machine learning techniques like gradient boosting regression but first just to say that those maps are
really very interesting because we use a whole set of data for the training and testing the the model and the model is validate always with the leaving one country approach that means we use all data from Lucas we left data from one country just for validation and results are
really very similar in this case you you see some like some results like gradient boosting regression random forest results read regression support vector regression and also extreme
gradient boosting regression all those models were processed in python environment and you know with the with the numpy is a fundamental package for scientific computing but also we use cklearn as an open source machine learning
library also easily boost as an optimized distributed gradient boosting library and but as i told you we are still waiting for the results with the we are still waiting for
the inputs from the Lucas 2018 and also we are waiting for collected data from the serbia from institute data hopefully reliable data that's always a problem you know the next one
we're running in overtime okay i will finish in the next two minutes the next one is extracting information using nlp here you have the tasks and at the moment
we are still working on we have already some results related to embedding library for the data on serbian you can try this link but anyway they are especially related to the to the language serbian language so probably you can
just look how it looks like and the next one is or the final one result is something that is totally accomplished that means that we already prepare one kilometer and one day timers
maps or grids with the one kilometer spatial resolution and one day time resolution for the climate elements like minimum maximum temperatures mean atmospheric pressure and precipitation we also aggregate those data on monthly and annual level for this purpose
we already developed the random forest spatial interpolation methodology and it is already embedded in the package method you can look and find the information about this package in this link also all details are given in this paper we produce those data based of on
ojimet data 28 are in serbia we also use independent data for voivodina from some
automated methodological stations in voivodina and we produce those grids everyday grids for the 20 years there will be inputs in our models validation of this the grids are already done and finally you can also find that data in following your
repository so it's open and free for all people thank you branislav let's see we
we went a bit over time yeah so you're doing a lot of a lot of work mainly project is focused on serbia if i understand correctly yeah so and so that's also again interesting to see how will the predictions you know national level match the european level and how do we harmonize
that that that's why we call the project by the way geo harmonizer so thank you so much for your talk