Embed Size (px)
Citation preview
Data Science u poljoprivredi: tehnološka platforma za
automatizaciju daljinske detekcije pomoću bespilotnih letelica
Gostujuće predavanje
18.12.2015
Milan Dobrota
2
Globalni izazovi u poljoprivredi ...
... proizvesti više hrane sa manje obradivog zemljišta !
2050
9 milijardi ljudi,
70 % povećanja proizvodnje
hrane
Većina poljoprivrednog zemljišta se već
obrađuje
Gubici zbog korova, bolesti i štetočina su veći
od 20 % godišnje
3
Uvod
Precizna poljoprivreda (Precision Agriculture) is the application of geospatial techniques and sensors (e.g. GIS, Remote sensing, GPS) to identify variations in the field and to deal with them. High-resolution satellite imagery was more commonly used. Small unmanned aerial systems (UAS), are shown to be a potential alternative given.
Daljinska detekcija (Remote Sensing) predstavlja metod prikupljanja informacija putem sistema koji nisu u direktnom, fizičkom kontaktu sa ispitivanim objektom. U užem smislu obuhvata analizu i interpretaciju različitih snimaka zemljišta. Informacije se prikupljaju registrovanjem i snimanjem odbijene ili emitovane energije objekta i obradom, analiziranjem i korišćenjem tih podatka.
Bespilotne letelice (UAV, UAS) platforms offer new possibilities to agriculture in order to obtain high spatial resolution imagery delivered in near-real time.
The increase of spatial and temporal resolution of the geomatic products obtained with UAVs should be accompanied with the use of new algorithms and techniques for information abstraction from these products. A clear example of this fact is the use of vegetation indices such as NDVI, which can be substituted by computer vision techniques or other indices based on RGB bands information, which can be obtained with inexpensive sensors.
4
Šta radimo
AgriSens Tehnologija je deo Precizne poljoprivrede koja koristi:
Daljinsku detekciju za snimanje velikih obradivih površina
Prikupljanje velike količine podataka u realnom vremenu
Obradu i analizu podataka koristeći statističke metode i algoritme sposobne da uče
Razvioj tehnološke platforme za daljinsku detekciju
Izlazi iz sistema su geo-referencirane mape područja sa procesiranim i analiziranim
podacima od interesa za specifični zahtev u poljoprivredi
5
Kako to radimo
Snimanje i obrada slika visoke rezolucije i
odgovarajućeg spektra (RGB, NIR...), snimljene uz
pomoć bespilotnih letelica i pre-procesiranje algoritmima
obrade slika.
Obrada podataka izvlačenjem skrivenih šablona u
podacima dobijenim iz slike, koristeći analitičke
algoritme, što takođe uključuje samoučeće algoritme,
odlučivanje pomoću neuronskih mreža i sl.
Analiza korišćenjem Vegetativnih Indeksa (VI) dobijenih
obradom podataka rezultat će biti geo-referencirana
mapa posmatranog polja. Algoritmi će u prvim
iteracijama ulazne parametre dobijati od eksperata, da bi
kasnije sami učili i ispravljali se
6
Inovacija
UAV (bespilotne letelice) uz niže troškove omogućavaju
češća snimanja, visoku rezoluciju zahvaljujućim niskim
visinama i malim brzinama leta i značajno manje potrebne
obuke za korišćenje. Satelitski snimci i fotografije iz vazduha
su ograničeni vremenom potrebnim za snimanje, niskom
rezolucijom, zavisnošću od oblaka i visokim troškovima za
ažurne slike.
Integrisanost i sveobuhvatnost alata za prikupljanje i
obradu velike količine podataka, omogućava krajnjim
korisnicima gotove rezultate, snimanjem u različitim
spektrima, data mining-om, mašinskim učenjem i
automatizacijom čitavog procesa, bez potrebe za
ekpertskim znanjem korisnika.
Očuvanje životne sredine primenom SSWC (site specific
weed control) principa ima značajne ekološke prednosti
smanjenom upotrebom pesticida i hebicida.
7
Primena tehnologije – u obuhvatu projekta
Identifikacija korova, u prvoj fazi kod široko-rednih zasada
(kukuruz, suncokret, šećerna repa, itd.)
Detekcija stresa koji može biti posledica oboljenja, štetočina ili
suše, praćenjem promena na listovima useva. Razlike refleksije
među različitim delovima EM spektra se koriste za razlikovanje
zdrave vegetacije od uvenule ili bolesne.
Brojanje biljaka i procena prinosa, naročito kod široko-rednih i
višegodišnjih zasada
8
Drugi primeri primena
Detekcija hlorofila: EM energija emitovana
od useva varira tokom cele sezone i tokom
dana u zavisnosti od sunčevog zračenja.
Detekcija nedostatka azota: distributeri
azotnog đubrivo nemaju algoritam po kome
upravljaju količinom đubriva distribuiranom
na pojedinim delovima zemljišta što može
dovesti do povećanja troškova i smanjenja
prinosa.
Klasifikacija zemljišta: fizičke osobine
zemljišta su u korelacijama sa reflektovanim
elektromagnetnim talasima određenih
talasnih dužina i zbog toga slike imaju
potencijal u automatskoj klasifikaciji vrsta
zemljišta i njihovom mapiranju
9
Korisnici tehnologije
10
Koncepet rešenja
11
Video
12
Proces, ogledi, analize...
13
Prikupljanje slika
Prikupljanje slika može biti podeljeno u tri faze:
Planiranje misije
Letenje UAV-om i slikanje (RGB & NDVI, Normalized Difference Vegetation Index)
Spajanje-mozaiking orthophoto slika
14
Dokumentovanje ogleda
15
Ekstrakcija vegetacijskih indeksa
Vegetation interacts with solar
radiation in a different way than
other natural materials. The
vegetation spectrum (figure 3)
typically absorbs in the red and blue
wavelengths, reflects in the green
wavelength, strongly reflects in the
near infrared (NIR) wavelength, and
displays strong absorption features
in wavelengths where atmospheric
water is present.
Different plant materials, water
content, pigment, carbon content,
nitrogen content, and other
properties cause further variation
across the spectrum
16
Automatic labelling Provides the automatic proposal of the OOI on the image.
Various clustering methods are be used for this task, namely k-means and its modifications,
DBSCAN and its modifications, OPTICS and its modifications.
17
Automatic labelling Method name Parameters Scalability Usecase Geometry (metric used)
K-Means number of clusters Very large n_samples, medium n_clusterswith MiniBatch code
General-purpose, even cluster size, flat geometry, not too many clusters
Distances between points
Affinity propagation damping, sample preference
Not scalable with n_samples
Many clusters, uneven cluster size, non-flat geometry
Graph distance (e.g. nearest-neighbor graph)
Mean-shift bandwidth Not scalable withn_samples Many clusters, uneven cluster size, non-flat geometry
Distances between points
Spectral clustering number of clusters Medium n_samples, small n_clusters
Few clusters, even cluster size, non-flat geometry
Graph distance (e.g. nearest-neighbor graph)
Ward hierarchical clustering number of clusters Large n_samples andn_clusters
Many clusters, possibly connectivity constraints
Distances between points
Agglomerative clustering number of clusters, linkage type, distance
Large n_samples andn_clusters
Many clusters, possibly connectivity constraints, non Euclidean distances
Any pairwise distance
DBSCAN neighborhood size Very large n_samples, medium n_clusters
Non-flat geometry, uneven cluster sizes
Distances between nearest points
Gaussian mixtures many Not scalable Flat geometry, good for density estimation
Mahalanobis distances to centers
Birch branching factor, threshold, optional global clusterer.
Large n_clusters andn_samples
Large dataset, outlier removal, data reduction.
Euclidean distance between points
18
Automatic labelling – stress monitoring
19
Image Recognition
Counting:
Template matching, a is a technique in digital image processing
for finding small parts of an image which match a template image
Haar-like feature based cascade sums up the pixel intensities in
each region and calculates the difference between these sums.
This difference is then used to categorize subsections of an images
are digital image features used in object recognition.
Stress identification and Yield estimation:
Histogram matching is the transformation of an image so that its
histogram matches a specified histogram. An image histogram is a
type of histogram that acts as a graphical representation of the
tonal distribution in a digital image
20
Software
21
Poslovni model
Key Partners Key Activities
Key Resources
Value Propositions Customer Relationships
Channels
Customer Segments
Cost Structure Revenue Streams
Vendors of equipment
Vendors of software tools
External consultants (know-how or sales activities)
Development of technology (integration of HW and SW) and know-how
Sales activities
Skilled experts in area of IT, data science, agriculture, mechatronics...
Funding for development and marketing activities
Raising initial awareness through internet, fairs and exhibitions
Channels of sales is under development at this point (direct sales)
Costs for developers and engineers to develop technologyFees for external consultants (not part of the core project)Purchase of hardware equipment (UAV, camera, etc.)Expenses related to sales activitiesExpenses related to field research
Services of crops examinations provided as a serviceSold out technology, either fully or partiallyIt is expected that customers currently pay more expensive technologyTechnology selling would be less frequent but generating bigger revenue at one shot
Direct contacts and networking with potential customers
Weak relations exists so far, at the level of pilot projects
Delivery of cost-effective intelligence about crops
Increase crops yields and reduction of risks from pests
Comprehensive technology integrated with advanced know-how in use for the benefit of customer
Satisfaction of customer needs of improving their farming in cost-effective manner
Environmental care (decrease of pollution)
Individual agricultural producers, using survey services
Large companies in agriculture who wish to implement technology and perform surveys
Government agriculture sectors
Insurance companies
Large technology companies interested in buying technology
