Center for Biosystems

Center for Information Technologies

Project aim: AITool4WYP project is dedicated to the digital transformation of a traditional way of early prediction wheat yield, with the potential to expand usage on other types of crops.
&nbsp;
About the project: This project will bring and make accessible smart digital technology even to small producers. The result of the project will be the implementation of a prototype of mobile applications for early prediction of wheat yield based on images of a part of the plot from a mobile phone and UAVs.
&nbsp;
This application offers to:
&nbsp;
• Producers: 
-continuous crop development monitoring 
-a possibility of timely action in the case of malnutrition, a lack of water, or occurrence of 
diseases 
-improvement in logistics and business management (plan harvest workloads and machinery 
capacity) 
-pre-harvest selling (price spike) 
-sustainable production
&nbsp;
• Silo-owners: 
-improvement in logistics and management of processes of the supply chain (plan harvest crop 
storage and drying capacity)
&nbsp;
• Seed companies: 
-faster determination of genetic seed potential and organization of a larger number of wheat variety trials
&nbsp;
• Insurance companies: 
&#8211; an objective damage assessment system that will secure trust between the farmers and the 
insurance company. Currently, is done by field experts and is prone to errors, corruptive 
activities and increases the cost of insurance.
&nbsp;

Research associate in the field of technical/technological sciences- information technologies

Marko Panić

Željana Grbović is a young researcher and PhD student in electrical engineering at University of Novi Sad, Faculty of Technical Sciences. She is highly interested in the application of machine learning algorithms and deep learning in the field of agriculture. Her beginnings in machine learning started 2016 when her team ranked in the top 5 among 52 teams in the prestigious competition SP Cup organized by IEEE. She received her master’s degree with the thesis about ear detection in wheat thermal images using deep neural networks. She has expanded the field experiment on the wheat involving images from mobile phones to provide early prediction. She has got funding from Innovation Fund, for The Proof of Concept for her PhD research topic. She is a guest researcher at Wageningen University and Research (WUR) in the group of Computer Vision and Agro-Food Robotics. She was involved in a Humistatus project about the prediction of spreading fungal infection on tomatoes based on hyperspectral images funded by the company Green Yard and WUR. She was a member of the team who tested PerClass MIRA Software for hyperspectral image processing before it was launched as public. Her main interests are image processing and application of transfer deep learning in plant phenotyping, food quality and safety. She is an active member of the artificial intelligence community and was invited speaker on several meetups and conferences where she promoted obtained results on phenotyping. She was a leader of a team that won third prize in the national competition „Space4Women“. Also, she led a team that participated at Alibaba Get Global Challenge final competition. She is actively involved in a few Horizon 2020 projects including ANTARES, Cybele, Dragon (serves as Dissemination Manager) and AgROBOfood and a few industrial projects with partners.

Željana Grbović

Sanja Brdar is a senior researcher and head of the Centre for Information Technologies at BioSense Institute. She received a PhD degree from the University of Novi Sad in 2016. Prior to starting her PhD studies, she spent two years working in the software industry with a major in databases design and development. In 2010 she was awarded a ten-month Basileus fellowship for joining the Bioinformatics Laboratory at University of Ljubljana which set the foundations for the PhD thesis on data/knowledge fusion in bioinformatics. Currently, she is actively involved in many projects including H2020 Antares, Dragon, Bestmap and Flexigrobots and teaches two master level courses at Faculty of Sciences. Her research interests include machine learning, explainable AI and bioinformatics. She works on ensemble methods, data fusion, clustering, and predictive modelling with applications in biology, agriculture, environmental sciences and health. She enjoys participating in worldwide data science challenges and with the team of researchers made significant results: placed third in Nokia Mobile Data Challenge (2012), finalist of Orange, France, Data for Development Challenge (2013), finalist/winner/third place of Syngenta Crop Challenge (2016, 2017, 2019) and finalist of UNDP Depopulation challenge (2020).

Sanja Brdar

Oskar Marko, Ph.D. is the Head of the Center for Information Technologies and the Assistant Director for Innovation and Cooperation with the Economy at the BioSense Institute. His research focuses on applying advanced machine learning and evolutionary algorithms in agriculture. He spent the 3rd year of his undergraduate studies at City University London, where he did his final BEng project in signal processing. He led BioSense’s team that developed novel Big Data algorithms for yield prediction, seed distribution, and optimization of sowing strategies, which secured the 1st prize for BioSense at the Syngenta Crop Challenge and CGIAR Inspire Challenge. He is actively involved in many Horizon 2020 projects including Antares, Cybele, Dragon, and Flexigrobots, and industrial projects with partners such as Krivaja, MK Agriculture, Delta, and other farming and insurance companies.

Oskar Marko

Dr. Marko Panić is a research associate in the field of technical sciences and information technologies at the BioSense Institute. His doctoral dissertation is related to the development of algorithms for accelerated image reconstruction from incomplete measurements using a sparse representation of signals and statistical Markov random models. His research direction includes statistical modeling of multidimensional data obtained from different sensors. He is currently working on the analysis and processing of images obtained from hyperspectral and multispectral cameras using advanced machine learning methods with application in assessing the quality of fruits and vegetables. As a member of the BioSense team, he participated in Syngenta's "Crop Challenge" competition, which resulted in winning first place. He is actively involved in many projects from the Horizon 2020 call, including Antares, Cybele, Dragon and Flexirobots, as well as in projects with companies such as Krivaja, MK Agri and Delta Agrar.

Bojana Ivošević is working on utilizing Unmanned Aerial Vehicles (UAV) for efficient monitoring of croplands, freshwaters, and terrains. She hails from a background in biology and ecology. She initiated her PhD in Kyungpook National University, South Korea, where she gained 2 years of experience with UAVs. She continued to pursue her field of interest at the Faculty of Sciences, University of Novi Sad, Serbia. Her current research focus is to implement UAV technology in precision agriculture to calculate vegetation indices important for agricultural crop yield production, disease detection and crop management. She is also interested in enforcing UAVs for animal detection and counting, monitoring algae bloom, land cover classification, and seasonal effects on the ecosystems.

Bojana Ivošević

logo

This project with grant number 5801 is funded by the Innovation Fund within the program named  „Proof of Concept“  designed for researchers who require support to prove that there is a new product, emerged from research and for which there is a need in the market.


National

BioSense’s research in the domain of imaging technologies relies on different devices for acquisition of data. They vary in the parts of the spectrum they cover and the number of bands in which they operate. These devices include visible near infrared (VNIR) and near infrared (NIR) hyperspectral cameras for lab research, and multispectral and thermal cameras which can be also used for the in-field purposes.

&nbsp;

Laboratory analysis is more accurate as hyperspectral cameras perform line scanning in more than 200 spectral bands and to further increase the reliability of measurements, they need to be calibrated using the reference sheets. However, high accuracy requires a large storage and very efficient image processing algorithms, since a hyperspectral image can take up more than 2 GB of space. For reduction of dimensionality we are selecting the most discriminative spectral bands, bands with high signal-to-noise ratio or applying the standard techniques such as PCA and ICA. Thermal and multispectral images also require pre-processing, but it is more focused on noise reduction using filtration. In the next step, depending on the problem, we are relying on feature engineering, clustering techniques and deep learning to segment the image into the regions of interest. Afterwards, depending on the problem, we are using the techniques of post-processing for final representation of results, or we are feeding the results into the machine learning models for higher-level analysis.

&nbsp;

<img class="alignnone wp-image-1492 size-large" src="https://api.biosens.rs/wp-content/uploads/2021/06/50-1024x544.jpg" alt="" width="1024" height="544" />

&nbsp;

When it comes to applications, we used these advanced cameras to develop algorithms for detection of fungal infections in vegetables and for revealing bruises in fruit days before visible symptoms occur, but they have also proven extremely valuable in ecosystems monitoring and yield prediction. These algorithms allowed us to get more information about the plants, estimate shelf-life of agricultural products, predict the crop yield and provide objective recommendations in post-harvest decision-making.

Hyperspectral, Multispectral and Thermal Imaging

At Biosense Institute we are using quadcopter DJI Inspire 1 to gain better insight into crop production and plan and manage agricultural operations more accurately. The technology can collect very high-resolution imagery below cloud level, with much more detail than the satellite imagery. The datasets include images captured with RGB and modified NDVI camera, but we are also using Flir thermal camera to make temperature measurements over large areas of heterogeneous terrain. Unmanned aerial vehicles are easy to use, as most mapping and data-collection missions are now conducted autonomously, by feeding the shape file of the desired area into the dedicated software.

&nbsp;

For image stitching we use Pix4D and AgisoftPhotoscan software to generate high-resolution orthomosaics, vegetation index maps and DSMs (Digital Surface Models). These orthomaps are important layers of information and they can be integrated into GIS systems along with other layers derived from additional sources. These datasets are then processed and used for training of advanced machine learning algorithms.

&nbsp;

BioSense uses UAS (Unmanned Aerial Systems) to detect nutrient deficiencies, pest infestations, water stress or excess and other conditions that affect crop development and improve the management of crops. They are especially useful for fields of a few dozens of hectares, such as orchards or vineyards, as scanning large fields of arable crops requires a lot of time. UAVs have also proven as effective tools for environmental monitoring, as they can cover medium-sized habitats and tell us more about the type of surroundings that insects and mammals grow in. The use of UAVs revolutionized spatial ecology and agriculture and allowed us to get high-resolution data and make optimal decisions with high accuracy.

&nbsp;

<img class="alignnone wp-image-1853 size-full" src="https://api.biosens.rs/wp-content/uploads/2021/06/slika5.png" alt="" width="775" height="444" />

NDVI map of sugar beet crop

Drone Imaging

The goal of BioSense’s AI Group is to get a comprehensive and accurate overview of biosystems, make phenotyping and plant breeding more efficient, understand molecular biology better and offer optimal and automatic decision-making to the farmer. Our research starts with the problem and requires innovative application of state-of-the-art algorithms and development of the novel ones. Machine learning algorithms extensively used in BioSense applications include Random Forest, SVM, Decision Trees, Ridge Regression, Spectral Clustering and many others including special family of machine learning algorithms- Deep Learning, and where the complexity of the problem goes beyond the available algorithms, BioSense researchers develop the novel ones.

&nbsp;

Example of novel algorithm is include weighted histograms regression (WHR). WHR is based on a voting mechanism, where training samples are voting for the output value of a test sample. The votes are weighted according to the similarity of features and used to form a histogram that essentially represents the PDF of the output, while the expectation of the PDF is essentially the predicted value. This algorithm proved very useful for yield prediction, where it outperformed random forest, SVM and artificial neural networks.

&nbsp;

&nbsp;

<img class="alignnone wp-image-1478 size-full" src="https://api.biosens.rs/wp-content/uploads/2021/06/42.png" alt="" width="503" height="504" />

&nbsp;

&nbsp;

Another novel algorithm is integrative clustering based on Nonnegative matrix factorization. Combining results of individual clusterings together into an ensemble leverages evidence accumulation in order to improve the results of clustering. Individual clustering results are merged into a joint cluster membership matrix that is further factorized into matrices of encodings and basis. Algorithm can integrate clusterings stemming from different data sets, different data preprocessing steps or different subsamples of features or objects.

&nbsp;

This algorithm proved very useful for functional genomics. Proposed approach based on nonnegative matrix factorization can fuse diverse data sources and infer gene clusters with high functional enrichment and improved gene coverage.

&nbsp;

&nbsp;

<img class="alignnone wp-image-1480 size-large" src="https://api.biosens.rs/wp-content/uploads/2021/06/43-1024x419.png" alt="" width="1024" height="419" />

Machine Learning

Deep learning is a rather new method of machine learning, that is due to its specifics and excellent performance in image processing often considered as a separate field within artificial intelligence. It works best with big datasets which have multi-dimensional characteristics, and there are a number of such datasets processed at BioSense.

&nbsp;

The first group of datasets are hyperspectral, multispectral and thermal images acquired in the lab or during the outdoor field experiments. Here, the the goal was to segment the images and recognise characteristic patterns within them, such as bruises and scratches on apples, infected tomato sepals or ears in images of wheat. Different architectures were tested with various parameters, which included the number of convolutional and fully connected layers, type of activation functions and position of dropout layers. They all were optimised for a particular problem, concerning the size of input images and the number of samples in the database.

&nbsp;

<img class="alignnone wp-image-1469 size-full" src="https://api.biosens.rs/wp-content/uploads/2021/06/337.jpg" alt="" width="640" height="480" />

&nbsp;

The second group consists of airborne particle measurements acquired with Rapid-E, the ﬁrst automated particle detector based on laser-induced ﬂuorescence technology, which records scattered light and laser-induced fluorescence patterns, representing the shape and chemical footprint of the particles. This is one of the examples of deep neural networks (DNNs) that relied heavily on convolutional layers, due to their ability to deal with datasets fused from different data sources.

&nbsp;

In both applications, preprocessing is perhaps the most vital task, as the noisy and non-standardised inputs are very difficult to analyse. Also, an important thing to note is that every problem requires a customised solution, whose complexity corresponds to the richness of the dataset and whose accuracy justifies its learning and execution times.

Deep Learning

AITool4WYP – Artificial intelligence-driven tool for early wheat yield prediction 

Centar za Biosisteme

Centar za Informacione Tehnologije

Cilj projekta: AITool4WYP projekat je posvećen digitalnoj transformaciji dosadašnje metode za procenu prinosa pšenice, sa potencijalom proširenja na druge kulture, na način koji će biti pristupačan i malim proizvođačima.
&nbsp;
&nbsp;
O projektu: AITool4WYP projekat će uvesti i približiti pametnu digitalnu tehnologiju malim proizvođačima, tako što će se kroz projekat potvrditi i realizovati prototip mobilne aplikacije za ranu predikciju prinosa pšenice na osnovu slike dela parcele sa mobilnog telefona i bespilotnih letelica tj. drona. Ova aplikacija će omogućiti proizvođačima da kontinuirano nadgledaju rast i razvoj useva. Osim toga, mogućnost pravovremenog delovanja u slučaju nedostatka nutrijenata, vode ili pojave bolesti, rezultiratiće unapređenjem logistike i rukovođenja farmi. Sa ovom informacijom, proizvođači će moći bolje da planiraju žetvu i kapacitete za skladištenje i sušenje, kao i da bolje odgovore na potrebe tržišta. Vlasnicima silosa se omogućava pregled očekivanog prinosa na regionalnom ili nacionalnom nivou, pa samim tim se i poboljšava i olakšava upravljanje procesima u celokupnom lancu snabdevanja. Semenskim kompanijama korišćenje ove aplikacije olakšava i ubrzava određivanje genetskog potencijala semena, kao i organizaciju velikog broja eksperimenata sa različitim vrstama pšenice. Ova aplikacija će osiguravajućim kompanijama omogućiti objektivan sistem za procenu štete i osnažiti poverenje sa proizvođačima, a štitiće interese obe strane.

Naučni saradnik u oblasti tehničko-tehnoloških nauka- informacione tehnologije

 Naučni saradnik u oblasti tehničko-tehnoloških nauka- elektronika, telekomunikacije i informacione tehnologije

Željana Grbović je istraživač i student doktorskih studija na departmanu Energetika, elektronika i telekomunikacije na Fakultetu tehničkih nauka, Univerzitet u Novom Sadu. Njena oblast interesovanja obuhvata primenu mašinskog učenja u poljoprivredi. Algoritmima veštačke inteligencije je počela da se bavi 2016. godine kada je njen tim rangiran u TOP5 u konkurenciji od 52 tima na prestižnom takmičenju u obradi signala, SP Cup koji organizuje IEEE. Završila je master studije 2017. godine na temi detektovanja klasova pšenice u termalnim slikama korišćenjem dubokih konvolucionih neuronskih mreža. Istraživanje o pšenici je proširila i na ranu predikciju prinosa, na osnovu slika sa mobilnog telefona i drona, za šta je dobila finansiranje od Inovacionog Fonda, za Dokaz Koncepta. Gostujući je istraživač na Univerzitetu u Vageningenu (Wageningen University and Research (WUR)), u grupi za Kompijutersku viziju i robotiku u poljoprivredi. Učestvovala je u industrijskom projektu WUR-a i kompanije GreenYard, pod nazivom Humistatus, za predikciju širenja patogenih oboljenja na paradajizu, na osnovu hiperspektralnih slika. Glavni domen istraživanja je transfer dubokog učenja u oblasti procene prinosa, kvaliteta i bezbednosti useva u poljoprivredi. Predvodila je tim koji je osvojio nagradu na studentskom univerzitetskom takmičenju „Space4Women“ i koji je učestvovao u finalu svetskog „Alibaba Get Global Challenge“ takmičenja. Aktivni učesnik je na nekoliko Horizon 2020 projekata, uključujući ANTARES, Cybele, Dragon (Menadžer diseminacije) i AgROBOfood, kao i projekte sa kompanijama.

Dr Sanja Brdar je naučni saradnik i rukovodilac Centra za informacione tehnologije na Institutu BioSense. Doktorirala je na Univerzitetu u Novom Sadu 2016. Pre početka doktorata provela je dve godine u softverskoj kompaniji u timu za razvoj baza podataka. Desetomesečnu Basileus stipendiju za istražvački boravak u Laboratoriji za bioinformatiku na Univerzitetu u Ljubljani dobila je 2010. godine i time postavila osnove za doktorat u oblasti fuzije podataka/znanja u bioinformatici. Aktivno je uključena u mnoge projekte, uključujući H2020 Antares, Dragon, Bestmap i Flekigrobots i predavač je na dva master predmeta na Prirodno-matematičkom fakultetu, smer Nauka o podacima. Njena istraživačka interesovanja uključuju mašinsko učenje, objašnjivu veštačku inteligenciju i bioinformatiku. Radi na ansambalskim metodama, fuziji podataka, klasterizaciji i prediktivnom modelovanju sa primenom u biologiji, poljoprivredi, životnoj sredini i javnom zdravlju. Učesnica je na svetskim izazovima iz nauke o podacima i sa timom istraživača postigla je značajne rezultate: treće mesto na Nokia Mobile Data Challenge (2012), finalista Orange France, Data for Development Challenge (2013), finalista / pobednik / treće mesto na Singenta Crop Challenge (2016, 2017, 2019) i finalista UNDP izazova depopulacije Srbije (2020).

&nbsp;

Dr Oskar Marko je rukovodilac Centra za informacione tehnologije i pomoćnik direktora za inovacije i saradnju sa privredom. Njegovo istraživanje usmereno je ka primeni naprednih metoda mašinskog učenja i evolutivnih algoritama u poljoprivredi. Treću godinu osnovnih studija je pohađao je na “Siti univerzitetu” u Londonu, gde je odbranio diplomski rad u domenu obrade signala. Vodio je tim BioSensa koji je razvio nove algoritme iz oblasti obrade velikih podataka koje su za cilj imale predikciju prinosa, distribuciju semena i optimizaciju strategije setve, koje su Institutu osigurale prvo mesto na Sindžentinom “Izazovu za useve” i CGIAR-ovom “Inspirišućem izazovu”. Aktivno je uključen u mnoge projekte iz poziva Horizont 2020, među kojima su Antares, Cybele, Dragon i Flexirobots, kao i u projekte sa kompanijama poput Krivaje, MK Agri, Delte i drugih poljoprivrednih kompanija i osiguravajućih kuća.

Dr Marko Panić je naučni saradnik u oblasti tehnološko-tehničkih nauka – informacionih tehnologija na institutu BioSens. Tema doktorata mu je vezana za razvoj algoritama za ubrzanu rekonstrukciju slika iz nepotunih merenja korišćenjem retke reprezentacije signala I statističkih Markovljevih slučajnih modela Njegov pravac istraživanja obuhvata statističko modelovanje multidimenzionalnih podataka dobijenih sa različitih senzora. Trenutno radi na analizi I obradi slika dobijenih sa hiperspektralnih I multispektralnih kamera korišćenjenem naprednih metoda mašinskog učenja sa primenom u proceni kvaliteta voća I povrća. Kao član tima BioSens ucestvovao je na Sindžentinom takmičenju “Izazovu za useve” koje rezltiralo osvajanjem prvog mesta. Aktivno je uključen u mnoge projekte iz poziva Horizont 2020, među kojima su Antares, Cybele, Dragon i Flexirobots, kao i u projekte sa kompanijama poput Krivaje, MK Agri i Delta Agrar.

Bojana Ivošević se bavi upotrebom bespilotnih letelica za efikasno nadgledanje ratarskih kultura, slatkih voda i životne sredine. Ona je po struci biolog i ekolog. Doktorske studije je upisala na Kyungpook Nacionalnom Univerzitetu u Južnoj Koreji, gde je stekla dve godine iskustva rada na bespilotnim letelicama. Nastavila je studije na Prirodno-matematičkom fakultetu, Univerzitetu u Novom Sadu. Njen trenutni fokus istraživanja je primena bespilotnih letelica u preciznoj poljoprivredi za izračunavanje vegetacijskih indeksa važnih za procenu prinosa, detekciju bolesti i upravljanje usevima. Takođe je zainteresovana za primenu dronova u detekciji i prebrojavanju životinjskih vrsta, praćenju cvetanja algi, klasifikaciju zemljišnog pokrivača i sezonskih efekata na ekosisteme.

Ovaj projekat sa brojem granta 5801, finansiran je od strane Fonda za Inovacionu delatnost iz programa Dokaz koncepta, koji je namenjen istraživačima koji smatraju da njihova istraživanja imaju komercijalni potencijal za kojim postoji potreba na tržištu.

Nacionalni

Za analizu problema i prikupljanje podataka, BioSens koristi različite uredjaje za akviziciju slike. Razlike između uređaja tiču se širine spektra elektromagnetnog zračenja i broja spektralnih pojasa u kojima rade. Hiperspektralne kamere obuhvataju vidljivi i blisko-infracrveni deo spektra (VNIR) kao i isključivo blisko-infracrveni deo spektra (NIR) i koriste se za laboratorijske analize, dok se termalna i multispektralna kamera uglavnom koriste za terenski rad.

&nbsp;

Laboratorijske analize su preciznije jer hiperspektralne kamere mogu imati rezoluciju i od preko 200 kanala, a nakon kalibracije korišćenjem referentih ploča se dodatno povećava pouzdanost merenja. Ipak, kako jedna hiperspektralna slika može imati i do 2GB, ovakve slike zahtevaju velik memorijski prostor i veoma efikasne algoritme obrade slike. Za redukciju količine podataka biraju se najinformativniji spektralni kanali sa visokom odnosom snage signala i šuma ili se primenjuju standardni postupci analize glavnih komponenata (PCA ili ICA). Slike koji stižu sa termalnih i multispektralnih kamera takođe zahtevaju predobradu ali su ovi postupci više fokusirani na smanjenje šuma filtriranjem slika. Sledeći korak, zavisno od problema koji se razmatra, je analiza slike koja obuhvata tehnike izdvajanja obeležja, grupisanja podataka po sličnosti i primene dubokih neuralnih mreža radi segmentovanja slike na regione od interesa. Dalje se koriste dodatni algoritmi za finalnu prezentaciju rezultata i dobijeni rezultati se prosleđuju algoritmima mašinskog učenja za dalju analizu podataka.

&nbsp;

<img class="alignnone wp-image-1839 size-large" src="https://api.biosens.rs/wp-content/uploads/2021/06/100-1024x544.jpg" alt="" width="1024" height="544" />

&nbsp;

U praksi se ove napredne kamere koriste za razvoj algoritama za detekciju infekcija na povrću ili otkrivanja modrica na voću i to nekoliko dana pre nego što simptomi postanu vidljivi golim okom. Značajan je i njihov doprinos u praćenju uslova ekosistema i predikciji prinosa, a razvijeni algoritmi pružaju dodatnu informaciju o stanju biljaka, daju procenu roka trajanja poljoprivrednih proizvoda kao, predviđaju prinos i daju objektivne predloge za donošenje odluka nakon žetve.

Hiperspektralne, multispektralne i termalne slike

Na Institutu BioSens koristimo slike kvadkoptera DJI Inspire 1 kako bismo dobili bolji uvid u stanje useva i preciznije planirali i upravljali agroekološkim aktivnostima. Ova tehnologija omogućava prikupljanje podataka visoke rezolucije ispod nivoa oblaka sa mnogo više detalja nego što se može videti u satelitskim snimcima. Skup podataka uključuje slike iz vidljivog dela spektra kao i slike dobijene modifikovanom NDVI kamerom. Takođe koristimo i Flir termalnu kameru za merenje temperature na velikim površinama u hetoregenim staništima. Bespilotne letelice su jednostavne za upotrebu, jer se većina mapiranja terena i prikupljanja podataka sada sprovodi autonomno, prema zadatim granicama terena uz pomoć aplikacija za planiranje leta.

&nbsp;

Za preklapanje slika koristimo softvere Pix4D i AgisoftPhotoscan kako bismo generisali ortomozaike, mape vegetacionih indeksa i digitalne modele površine. Ove orto-mape su važni slojevi informacija i mogu se integrisati u GIS sisteme zajedno sa slojevima podataka dobijenih iz drugih izvora. Ovakvi skupovi podataka se zatim obrađuju i koriste za obuku naprednih algoritama mašinskog učenja.

&nbsp;

BioSens koristi bespilotne vazdušne sisteme za otkrivanje nedostatka hranljivih materija u biljkama, napada štetočina, prekomerne suše ili vlage kao i drugih faktora koji utiču na razvoj useva, kao i za povećanje efikasnosti pri donošenju odluka. Posebno su korisni za površine od nekoliko desetina hektara, poput voćnjaka i vinograda, jer skeniranje velikih ratarskih parcela zahteva mnogo vremena. Bespilotne letelice su se pokazale i kao efikasni alati za praćenje stanja ekosistema. Pokrivajući veća prostranstva, omogućavaju pristup informacijama o tipu staništa u kojem rastu i razvijaju se živi organizmi od insekata do sisara. Njihova upotreba dovela je do revolucionarnih otkrića na poljima prostorne ekologije i poljoprivrede, i omogućila nam dobijanje podataka visoke rezolucije za donošenje optimalnih odluka sa visokom preciznošću.

&nbsp;

&nbsp;

<img class="alignnone wp-image-1853 size-full" src="https://api.biosens.rs/wp-content/uploads/2021/06/slika5.png" alt="" width="775" height="444" />

NDVI map of sugar beet crop

Obrada slika sa dronova

Duboko učenje je novi metod mašinskog učenja. Zbog svoje specifičnosti i odličnih performansi u obradi slike često se posmatra kao zasebna oblast u okviru veštačke inteligencije. Najbolje radi na velikim skupovima podataka koji imaju multi-dimenzionalne karakteristike, sa kakvim se često susrećemo na BioSensu.

&nbsp;

U prvu grupu podataka koje posedujemo spadaju hiperspektralne, multispektralne i termalne slike, napravljene u laboratoriji ili na terenu. Cilj njihove obrade je segmentacija slike i prepoznavanje karakterističnih obrazaca, kao što su počeci truljenja i fizička oštećenja na jabukama, infekcija peteljke paradajza ili detektovanje klasova pšenice. Brojne arhitekture su testirane sa različitim parametrima, poput broja konvolucionih i potpuno povezanih slojeva, vrsta aktivacionih funkcija i slučajne eliminacije dela mreže da bi se izbeglo preučenje. Svaka arhitektura je optimizovana za specifičan problem, uzimajući u obzir veličinu slika i broj uzoraka u bazi podataka.

&nbsp;

<img class="alignnone wp-image-1469 size-full" src="https://api.biosens.rs/wp-content/uploads/2021/06/337.jpg" alt="" width="640" height="480" />

&nbsp;

Druga grupa skupa podataka se sastoji od merenja čestica u vazduhu pomoću uređaja Rapid-E, prvog automatskog detektora čestica zasnovanog na tzv. LIF (eng. laser-induced ﬂuorescence) tehnologiji. LIF tehnologija beleži raspršene fluroscentne obrasce svetlosti i lasera, koji predstavljaju oblik i hemijski otisak čestica. U ovom slučaju se koristi jedan od primera dubokih neuronskih mreža (DNN) koje se najviše oslanjaju na konvolucione slojeve, zbog njihove sposobnosti da obrade skupove podataka koji potiču iz različitih izvora.

&nbsp;

U oba slučaja, predobrada je jedan od najvažnijih koraka, jer su zašumljeni i nestandardni uzorci veoma teški za analizu. Takođe, važno je istaći da svaki problem zahteva drugačije, prilagođeno rešenje, čija složenost odgovara veličini i složenosti skupova podataka i čija tačnost mora opravdati vreme učenja i izvršavanja algoritma.

&nbsp;

Duboko učenje

Cilj BioSensove Grupe za veštačku inteligenciju je da obezbedi sveobuhvatan i precizan uvid u stanje biosistema, poveća efikasnost fenotipizacije i oplemenjivanja biljaka, ponudi bolje razumevanje molekularne biologije i pomogne poljoprivrednicima kroz optimalno i automatsko donešenje odluka. Naše istraživanje počinje od problema i zahteva inovativnu primenu i razvoj algoritama koji su poslednja reč nauke. Algoritmi mašinskog učenja koji se najčešće koriste za probleme kojima se BioSens bavi su stabla odluke, vektori nosači, ridž regresija, spektralno klasterovanje i mnogi drugi, uključujući i posebnu porodicu algoritama mašinskog učenja – duboko učenje, a tamo gde složenost problema prevazilazi postojeće algoritme, istraživači sa Instituta razvijaju nove.

&nbsp;

Primer algoritma koji je razvijen na BioSens Institutu je algoritam regresije težinskim histogramima. Dati algoritam je zasnovan na mehanizmu glasanja, gde uzorci za obuku glasaju za izlaznu vrednost test uzorka. Svaki glas dobija određenu težinu u zavisnosti od sličnosti uzorka koji glasa i uzorka iz test skupa. Pomoću dobijenih glasova, formira se histogram koji de fakto predstavlja funkciju gustine verovatnoće (PDF) izlaza test uzorka, dok matematičko očekivanje PDF-a predstavlja predikovanu vrednost datog uzorka. Ovaj algoritam se pokazao veoma korisnim za problem predikcije prinosa, gde je nadmašio algoritme stabla odluke, vektora nosača i veštačkih neuronskih mreža.

&nbsp;

&nbsp;

<img class="alignnone wp-image-1478 size-full" src="https://api.biosens.rs/wp-content/uploads/2021/06/42.png" alt="" width="503" height="504" />

&nbsp;

&nbsp;

Još jedan inovativni algoritam koji je razvijen na Institutu BioSens je algoritam integrisane klasterizacije zasnovan na nenegativnoj faktorizaciji matrica. Kombinovanjem individualnih rezultata klasterovanja u ansambl, dokazi se akumuliraju u cilju poboljšanja rezultata klasterizacije.

Individualni rezultati klasteriziacije spojeni su u zajedničku matricu koja se zatim faktorizuje u matrice kodovanja i bazne vektore. Algoritam ima mogućnost da integriše klastere koji potiču iz različitih baza podataka, različitih nivoa predprocesiranja ili poseduju različite podskupove obeležja.
Ovaj algoritam se pokazao kao izuzetno koristan za funkcionalnu genomiku. Dati pristup, zanosvan na nenegativnoj faktorizaciji matrica poseduje mogućnost fuzije različitih izvora podataka i rezultuje klasterima gena sa povećanom funkcionalnošću i pokrivenošću gena.

&nbsp;

&nbsp;

<img class="alignnone wp-image-1480 size-full" src="https://api.biosens.rs/wp-content/uploads/2021/06/43.png" alt="" width="1691" height="692" />

AITool4WYP – Artificial intelligence-driven tool for early wheat yield prediction

Themes:

People:

Centers:

CBS

CIT