Automated Phenotyping of Plant Root Systems

EPSO: The European Plant Science OrganisationEPSO Workshop on Plant PhenotypingNovember 02-03, 2009Forschungszentrum Jülich, Germany

Forschungszentrum Jülich, GermanyICG-3: PhytosphereJülich Plant Phenotyping Centre (JPPC)Website: http://www.jppc.de

http://www.plantphenomics.com/phenotyping2009

Joshua Weitz, Georgia Institute of Technology, Georgia, USA

Automatic phenotyping of plant root systems

Automated phenotyping ofplant root systems

Joshua Weitz, Ph.D.School of Biology & School of Physics

Georgia Institute of Technologyhttp://ecotheory.biology.gatech.edu

EPSO Workshop on Plant PhenotypingNovember 2, 2009Forschungszentrum JülichGermany

Anjali Iyer-Pascuzzi(Duke)

Olga Symonova(Georgia Tech)

Taras Galkovsky (GeorgiaTech/Kiev University)

Yuriy Mileyko(Georgia Tech/Duke)

John Harer (Duke)

Philip Benfey (Duke)

Briefly:What and Where is Georgia Tech

Located in the heart of Atlanta, Georgia Techhas approximately 13,000 undergraduatestudents & 7,000 graduate students.

Georgia Tech is a leading technology researchinstitution, going through a period of rapidgrowth in the biological sciences.

Rice varieties: (i) 9311 - 9 individuals; (ii) Caipo - 6 individuals

Rice varieties: (i) 9311 - 9 individuals; (ii) Caipo - 6 individuals

Difficult to discern a clear pattern: fine roothairs, branching-ness, length, size?

Project GoalsLong term goal: Identify genes underlying root system

architecture traits

1. Implement “real-time” non-destructive imagingsystem for rice roots

2. Develop useful 2D and 3D descriptors for rootsystem architecture

3. Identify heritable variation within rice varieties &quantify with descriptors

4. Identify root architecture responses toenvironmental stress

Project GoalsLong term goal: Identify genes underlying root system

architecture traits

1. Implement “real-time” non-destructive imagingsystem for rice roots

2. Develop useful 2D and 3D descriptors for rootsystem architecture

3. Identify heritable variation within rice varieties &quantify with descriptors

4. Identify root architecture responses toenvironmental stress

Gel-based Imaging Platform to Non-destructively View Root System Architecture

• Rotating stage setup with standardized lighting

• Gellan gum transparent growth media (2L ungraduated cylinders)

• >18,000 images generated for 12 genotypes and approximately 120 individuals

Observed Phenotypic Variation is HigherBetween than Within Genotypes

Small-scale Differences ObservableLow salt High salt

Project GoalsLong term goal: Identify genes underlying root system

architecture traits

1. Implement “real-time” non-destructive imagingsystem for rice roots

2. Develop useful 2D and 3D descriptors for rootsystem architecture

3. Identify heritable variation within rice varieties &quantify with descriptors

4. Identify root architecture responses toenvironmental stress

Root System Descriptor Pipeline

Methods:

Global thresholding (initially)

Adaptive thresholding (current)

Double-adaptive thresholding (planned)

There is No Off-the-shelf Solution forAll Thresholding Problems

Mehmet Sezgin and Bülent Sankur

Over 40 methods evaluated (none isoptimal for all conditions)

• Histrogram-based methods

• Clustering methods

• Entropy methods

• Local thresholding

…

“Standard” RSA Traits1. Median number of roots2. Maximum number of roots3. Average width4. Specific root length5. Surface area6. Total root length7. System depth8. System volume9. Length distribution10. Maximum horizontal width11. Root system width-depth ratio

Step 2: Calculate RSA Traits from Thresholded Images

“Novel” RSA Traits1. Network perimeter2. Solidity3. Convex area4. Projected network area5. Bushiness index

Estimating 3D traits from 2D images:Using the Medial Axis

Novel trait: Calculating the network soliditygenerally, we estimate features from the b&w transformed image

Convex hull area: Area of the convex hull (see green enclosing lines)

Solidity: Ratio of the root network area to the convex hull area

Root System Architecture Traits

Traits/features: Projections of a complex 3-dimensional object --involves loss of information.

Structure & function:Traits are chosen because of theirhypothesized relevance to function, e.g., yield, anchorage, stressresistance, nutrient uptake, etc.

What we think is important depends, in part, on whatwe can measure: The link between RSA traits and plant function isoften correlative. Measuring comprehensive root traits take time (arewe not measuring other important traits?). Here, analysis of entireimage, including thresholding & phenotyping, takes <1 minute.

Comprehensive traits require computational analyses:Our targets include previously identified RSA traits and other traits thatdescribe how a plant utilizes space & that cannot be estimated withoutimage analysis.

Project GoalsLong term goal: Identify genes underlying root system

architecture traits

1. Implement “real-time” non-destructive imagingsystem for rice roots

2. Develop useful 2D and 3D descriptors for rootsystem architecture

3. Identify heritable variation within rice varieties &quantify with descriptors

4. Identify root architecture responses toenvironmental stress

Determine the Statistically Significant TraitDifferences Between Genotypes

Methods:

1. PCA analysis

2. Machine learning

3. Pair-wise analysis of significant differences

Machine learning approach to classification:Support Vector Machines

SVM based classification can separate all genotype pairs(when genotypes are dissimilar)

SVM based classification can separate all genotype pairs(and when they are highly similar to the eye)

SVM Classification Yields Feature Ranks

Automatic Extraction of Feature Ranks

Feature Ranks Can Be Automatically Extracted…Generating Horribly Long Tables, But They Are Significant!

(the point is that all pairs have significant trait differences, evenwhen accounting for multiple comparisons)

Summary of Analysis• Imaging & Analysis Platform for Rice Root System Architecture

– Less intra- than inter-phenotypic variation– All pairwise comparisons reveal statistically significant differences– Feature ranking system using SVM and other approaches– Suggests use of imaging system for QTL identification– Paper in press: “Imaging and Analysis Platform for Automatic

Phenotyping and Trait Ranking of Plant Root Systems”

• First prototype of pipeline developed and implemented in Matlab– Image cleaning/thresholding– Feature extraction– Statistical analyses and classification– Requires minimal user input (at outset) and then fully automated– Usable by at least 2 people

• Follow-up manuscript on root growth dynamics• 3D reconstruction software & trait analysis in development• General framework for biological network processing

Next Challenge

Next Challenge

x 105 - 106 or more images

Both rice and maize

In 3D

Phenomics

“Over 1 km of conveyors delivering radio-tagged plants automatically to state-of-theart imaging stations controlled by highcapacity computing equipment.”

“The quantitative, high-throughput study of organismal phenotypes”

Application areas:

1. High-throughput plant breeding

2. Biofuels & bioenergy

3. Carbon sequestration

Broad Challenges in Root Phenomics

Broad Challenges in Root Phenomics• Imaging platforms

– Non-invasive & non-destructive– Field vs. laboratory based– Speed vs. accuracy– Replication at large scales and small scales

Fang et al. (Plant Journal, 2009)Laser based approach (3Danalyses, though slower thancamera-based reconstruction)

Phenotype Screening CorporationX-Rays (noisy images, but growthconditions can be modified

Brady et al. (Science, 2007)Combining expression &morphology for dicots

Broad Challenges in Root Phenomics• Image analysis

– Phenotyping roots in 2D• Optimizing current features & understanding trait correlations• Tracking dynamics & linking 2D -> 3D traits

– Phenotyping roots in 3D• 3D reconstruction (current technologies produce root systems

suitable for limited statistical, but not all relevant analyses)• Analyzing the geometric utilization of space by roots• In situ vs. in vitro challenges are distinct

• Downstream analysis– SVM or other classification to rapidly assess feasibility of QTL

studies with mapping population– Platform for sharing/searching/storing phenotype data within the

community• The equivalent of PDB or GenBank?• Diverse stakeholders from academia, government & industry

• Developing broadly use-able tools for “phenomics” across thespectrum of needs/interactivity:– User-supervised (DART, WinRhizo etc.)– Fully automated

GiARoots is a scalable architecture and processing framework for the high-throughput analysis of rootnetworks. The objective of GiARoots is to create a flexible set of software tools for fully automated, batchprocessing of root network data. Contact for more details on upcoming release ([email protected]).

Data view Configuration view

Output of processingWorkflow computed by GiARoots

Data view

Automatically calculated workflow

User-modifiable parameters, in interactive mode

Output view

Weitz Group @ Georgia Techhttp://ecotheory.biology.gatech.edu

Email: [email protected]

Research themes

1. Viral dynamics and evolution

2. Structure of plant vascular networks

3. Systems biology and bioinformatics

4. Quantitative disease dynamics

Funding Papers on plant structure & design:Weitz et al. (2006) Ontogenetically stable hydraulicdesign in woody plants Fun. Ecol. 20: 191-199.

Weitz et al. (2007) Evolution, interactions and biologicalnetworks. PLoS Biol. 5: 10-12.

Price et al. (2009) Evaluating scaling models in biologyusing hierarchical Bayesian approaches. Ecol. Let. 12:641-651.

Iyer-Pascuzzi et al. (in press) Imaging and analysisplatform for automatic phenotyping and trait ranking ofplant root systems. Plant Physiology. (Updated Jan 25,2010)

Price & Weitz (submitted) Zero sum allocationalstrategies determine the allometry of specific leaf area.

Summer 2009

Questions?