DEEP LEARNING IN AGRICULTURE Erik Andrejko Head, Data Science — The Climate Corporation Silicon Valley Machine Learning Meetup Mar 28 2014
Sep 14, 2014
DEEP LEARNING IN AGRICULTURE
Erik Andrejko Head, Data Science — The Climate Corporation Silicon Valley Machine Learning Meetup Mar 28 2014
OUTLINE• The Climate Corporation
• The Agricultural Challenge
• The Role of Deep Learning
THE CLIMATE CORPORATION
OUR MISSION
To help all the world’s people and businesses manage and adapt to climate change…
starting with agriculture
PROTECT & IMPROVE
MP
TWI
GR
OW
ER A
PPLI
CAT
ION
S
Yield Potential
Profitability
Yield Expectation
TWITotal Weather Insurance: parametric supplemental crop insurance product
MPGovernment subsidized loss-adjusted insurance program, integrated risk management
GROWER APPLICATIONSCollection of grower management advisors using agronomic and climatological models
PROTECT
IMPROVE
The Climate Corporation aims to help farmers around the world protect and improve their farming operations & profitability.
THE AGRICULTURAL CHALLENGE
WORLD POPULATION
Alexandratos N, Bruinsma J. 2012. World agriculture towards 2030/2050, the 2012 revision. ESA Working Paper No. 12-03, June 2012. Rome: Food and Agriculture Organization of the United Nations (FAO)
AGRICULTURAL DEMAND
POPULATIONMORE PEOPLE
CONSUMPTIONHIGHER DEMAND / CAPITA
NORMAN BORLAUGIS CREDITED WITH SAVING
BILLION LIVES1
GREEN REVOLUTION
YIELD INCREASES
Ray DK, Mueller ND, West PC, Foley JA. 2013. Yield Trends Are Insufficient to Double Global Crop Production by 2050. PLoS ONE 8(6), doi:10.1371/journal.pone.0066428.
crop yield mustI N C R E A S E
60% to meet demand
b y 2 0 5 0Ray DK, Mueller ND, West PC, Foley JA. 2013. Yield Trends Are Insufficient to Double Global Crop Production by 2050. PLoS ONE 8(6), doi:10.1371/journal.pone.0066428.
*
EXAMPLE: BEEF
2,500M 700M
Source: FAO and USDA (assuming 2kg of cereal of 500g of beef)
Current worldwide cereal production (tonnes)
Corn demand to support current US per-capita beef consumption for a population of 7B (tonnes)
IT’S POSSIBLE
I now say that the world has the technology… to feed on a sustainable basis a population of 10 billion people. ”
“
Normal Borlaug
DATA SCIENCE MEETS AGRICULTURE
NEXT REVOLUTION ?
INTENSIFYApply breeding, fertilization to increase yields.
OPTIMIZEApply data science to optimize management.
GREEN REVOLUTION
GREEN DATA REVOLUTION
1960 –
2010 –
BIOTECHMarker assisted selection.
BIOTECH REVOLUTION1980 –
DATA SCIENCE
Computer Science
Domain Science Statistics
What is important?
How can it be built?
How can predictions be made?
SCIENTIFIC DATA SCIENCE
use software engineering to enable domain science maximizing use of data
MACHINE LEARNINGGENETICS PRACTICES ENVIRONMENT YIELD
T R A I
N
T E S T
predictive model
problem: curse of dimensionality
YIELD OPTIMIZATION
OPTIMIZED YIELD
Yield optimized for environment by optimization of genetics and management using predictive model.
YIELDYield optimized for environment by optimization of genetics and management traditional practices.
EASY! RIGHT?unfortunately, no
CHALLENGES
Spatio-‐Temporal Data
Sparse Data
Latent Features
Curse of Dimensionality
DATA CHALLENGES LEARNING CHALLENGES
Missing Data Multi-‐task Learning
Noisy Data
DATA POTENTIALYIELD MONITOR DATA
14B OBSERVATIONS
REMOTE SENSING DATA
260B OBSERVATIONS
WEATHER DATA
20B OBSERVATIONS
one season, one crop, one country
FEATURE ENGINEERING
Zea mays (corn)
Genetics, Environment, Practices
Soil Processes
Nutrient Processes
Crop Processes
Yield
LATENT SPACE
LATENT FEATURE SPACE
Environment, Genetics and Practices
Physical Processes
Yield Outcome
}{Engineered Features
Learned Features
CAUSAL DESCRIPTION
THE ROLE OF DEEP LEARNING
FEATURE LEARNINGgenetics, environment and practices
soil processes
nutrient processes
crop processes
yield
Hierarchical Dimensionality
ReductionDeep Neural
Network
yield
genetics, environment and practices
hidde
n lay
ers
phys
ical m
odels
SPATIAL DATA
Lee, Honglak, et al. "Convolutional deep belief networks for scalable unsupervised learning of hierarchical representations." Proceedings of the 26th Annual International Conference on Machine Learning. ACM, 2009.
CONVOLUTIONAL DBN
Hierarchical representation of spatial data !High-dimensional, scalable visible layer !Unsupervised hierarchical learning
MULTI-TASK LEARNINGDEEP NEURAL NETWORK
Hidden layers (latent features) shared across tasks !Multi-task informs latent features
deep neural network in multi-task setting
y
genetics, environment and practices
Hid
den
Laye
rs
w
Deng, Li, Geoffrey Hinton, and Brian Kingsbury. "New types of deep neural network learning for speech recognition and related applications: An overview." Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on. IEEE, 2013.
MISSING DATA
Hinton, Geoffrey E., Simon Osindero, and Yee-Whye Teh. "A fast learning algorithm for deep belief nets." Neural computation 18.7 (2006): 1527-1554.
DEEP BELIEF NETWORK
Greedy layer-wise training algorithm !Robust to noisy inputs !Generative process (MRF)
Alternating Gibbs sampling in deep belief network
additional applications of deep-learning in agriculture
OTHER APPLICATIONSCrop identification
Disease detection
Practice classifications
Remote sensing
Image segmentation / clustering
Nutrient deficiency detection
Cloud detection Environment classification
PROTECT & IMPROVEREDUCE RISK INCREASE YIELDS
Goal: optimize global food production
POSSIBILITIES
QUESTIONS [email protected]