Upload File

combining satellite imagery and machine learning to ......high-level overview transfer learning...

  • Home
  • Documents
  • Combining Satellite Imagery and Machine Learning to ......High-level overview Transfer learning approach Results § Global poverty line: $1.90/person/day § Almost 1 billion people
1
Mo#va#on Combining Satellite Imagery and Machine Learning to Predict Poverty Neal Jean, Michael Xie, Stefano Ermon {nealjean, sxie, ermon}@stanford.edu References [1] C. Elvidge, K. Baugh, et al. Proceedings of the Asia-Pacific Advanced Network, 2013. [2] J. Blumenstock, G. Cadamuro, and R. On, Science 350, 6264 (2015). [3] M. Xie, N. Jean, et al. CoRR 1510.00098 (2015), hXp://arxiv.org/abs/1510.00098. High-level overview Transfer learning approach Results § Global poverty line: $1.90/person/day § Almost 1 billion people live in abject poverty § Lack of reliable data in developing countries poses a major challenge for making informed policy decisions § Goal: Accurately and scalably predict poverty and wealth measures from day#me satellite images, using only publicly available data X 1 X 2 X 3 X 4 X n Convolu<onal Neural Network Y 1 Y 2 Y 3 Y 4 Y n Predic<ons: Economic indicators Inputs: Day#me satellite imagery Low-dimensional feature representa<on § Standard supervised learning won’t work—there isn’t enough ground truth poverty data § Key idea: Use nighgme light intensity as a proxy for economic development § Train a convolu#onal neural network (CNN) to extract image features relevant to nighgme lights § A subset of these features are also useful for predic#ng economic well-being North Korea vs. South Korea Learned features corresponding to buildings, undeveloped areas, water, and roads −2 −1 0 1 2 3 −1 0 1 2 predicted asset score observed asset score Nigeria, 2013 r 2 = 0.68 −1 0 1 2 3 4 0 1 2 predicted asset score observed asset score Tanzania, 2010 r 2 = 0.57 a b c d § Task: Predict consump#on expenditures and asset wealth in 5 African countries Asset-based wealth predic#ons for (a) Nigeria and (b) Tanzania Nigeria Tanzania Uganda Malawi Pooled Country Evaluated On Nigeria Tanzania Uganda Malawi Pooled Consumption Expenditures Country Trained On Nigeria Tanzania Uganda Malawi Rwanda Pooled Nigeria Tanzania Uganda Malawi Rwanda Pooled Country Evaluated On Assets Country Trained On 0.69 0.50 0.46 0.24 0.39 0.63 0.43 0.58 0.51 0.40 0.47 0.54 0.56 0.61 0.66 0.55 0.56 0.62 0.40 0.42 0.43 0.55 0.42 0.46 0.55 0.68 0.66 0.59 0.75 0.71 0.50 0.45 0.46 0.41 0.43 0.56 0.5 0.75 1 0.25 0 r 2 0.41 0.30 0.26 0.19 0.39 0.44 0.56 0.47 0.48 0.52 0.34 0.38 0.46 0.37 0.44 0.25 0.34 0.37 0.42 0.36 0.42 0.40 0.41 0.33 0.45 a b § Accuracy compares favorably with recent study in Science that uses mobile phone data to predict poverty § Models generalize well across borders, sugges#ng that satellite imagery reveals shared features of poverty Cross-border generaliza#on for (a) expenditure and (b) asset models

Upload: others

Post on 29-May-2020

4 views

Category:

Documents


0 download

Report

TRANSCRIPT

Page 1: Combining Satellite Imagery and Machine Learning to ......High-level overview Transfer learning approach Results § Global poverty line: $1.90/person/day § Almost 1 billion people

Mo#va#on

CombiningSatelliteImageryandMachineLearningtoPredictPoverty

NealJean,MichaelXie,StefanoErmon{nealjean,sxie,ermon}@stanford.edu

References[1]C.Elvidge,K.Baugh,etal.ProceedingsoftheAsia-PacificAdvancedNetwork,2013.[2]J.Blumenstock,G.Cadamuro,andR.On,Science350,6264(2015).[3]M.Xie,N.Jean,etal.CoRR1510.00098(2015),hXp://arxiv.org/abs/1510.00098.

High-leveloverview

Transferlearningapproach Results§  Globalpovertyline:$1.90/person/day

§  Almost1billionpeopleliveinabjectpoverty

§  Lackofreliabledataindevelopingcountriesposesamajorchallengeformakinginformedpolicydecisions

§  Goal:Accuratelyandscalablypredictpovertyandwealthmeasuresfromday#mesatelliteimages,usingonlypubliclyavailabledata

X1 X2 X3 X4 Xn

Convolu<onalNeuralNetwork

Y1 Y2 Y3 Y4 Yn

Predic<ons:Economicindicators

Inputs:Day#mesatelliteimagery

Low-dimensionalfeaturerepresenta<on

§  Standardsupervisedlearningwon’twork—thereisn’tenoughgroundtruthpovertydata

§  Keyidea:Usenighgmelightintensityasaproxyforeconomicdevelopment

§  Trainaconvolu#onalneuralnetwork(CNN)toextractimagefeaturesrelevanttonighgmelights

§  Asubsetofthesefeaturesarealsousefulforpredic#ngeconomicwell-being

NorthKoreavs.SouthKorea

Learnedfeaturescorrespondingtobuildings,undevelopedareas,water,androads

−2 −1 0 1 2 3

−1

0

1

2

pred

icte

d as

set s

core

observed asset score

Nigeria, 2013r2 = 0.68

−1 0 1 2 3 4

0

1

2

pred

icte

d as

set s

core

observed asset score

Tanzania, 2010r2 = 0.57

−2 −1 0 1 2 3

−1

0

1

pred

icte

d as

set s

core

observed asset score

Uganda, 2011r2 = 0.69

−1 0 1 2 3 4 5

−1

0

1

2

3

pred

icte

d as

set s

core

observed asset score

Malawi, 2010r2 = 0.55

−1 0 1 2 3 4 5

0

1

2

pred

icte

d as

set s

core

observed asset score

Rwanda, 2010r2 = 0.75

a b

c d

e

§  Task:Predictconsump#onexpendituresandassetwealthin5Africancountries

Asset-basedwealthpredic#onsfor(a)Nigeriaand(b)Tanzania

Nigeria Tanzania Uganda Malawi Pooled

Co

un

try E

valu

ate

d O

n

Nigeria

Tanzania

Uganda

Malaw

iPooled

Consumption Expenditures

Country Trained On

Nigeria Tanzania Uganda Malawi Rwanda Pooled

Nigeria

Tanzania

Uganda

Malaw

iRwanda

Pooled

Co

un

try E

valu

ate

d O

n

Assets

Country Trained On

0.69 0.50 0.46 0.24 0.39 0.63

0.43 0.58 0.51 0.40 0.47 0.54

0.56 0.61 0.66 0.55 0.56 0.62

0.40 0.42 0.43 0.55 0.42 0.46

0.55 0.68 0.66 0.59 0.75 0.71

0.50 0.45 0.46 0.41 0.43 0.56

0.5

0.75

1

0.25

0

r2

0.41 0.30 0.26 0.19 0.39

0.44 0.56 0.47 0.48 0.52

0.34 0.38 0.46 0.37 0.44

0.25 0.34 0.37 0.42 0.36

0.42 0.40 0.41 0.33 0.45

a b

§  AccuracycomparesfavorablywithrecentstudyinSciencethatusesmobilephonedatatopredictpoverty

§  Modelsgeneralizewellacrossborders,sugges#ngthatsatelliteimageryrevealssharedfeaturesofpoverty

Cross-bordergeneraliza#onfor(a)expenditureand(b)assetmodels

SHALLOW WATER BATHYMETRY MAPPING FROM UAV IMAGERY … · SHALLOW WATER BATHYMETRY MAPPING FROM UAV IMAGERY BASED ON MACHINE LEARNING P. Agrafiotis 1,2, D. Skarlatos 2, A. Georgopoulos

Spatial and machine learning methods of satellite imagery

Classification of Hyperspectral Imagery with Neural …fschleif/mlr/mlr...MACHINE LEARNING REPORTS Classification of Hyperspectral Imagery with Neural Networks: Comparison to Conventional

Using Machine Learning & Deep Learning With ArcGIS Imagery · 2020. 2. 24. · Deep Learning with Imagery in ArcGIS ArcGIS supports end-to-end deep learning workflows •Tools for:

IMAGERY University of Idaho. Imagery Theory Imagery Theory Questions of Interest What is imagery? Does imagery work? How does imagery work?

DeepSat – A Learning framework for Satellite Imagerybit.csc.lsu.edu/~saikat/publications/sigproc-sp.pdfDeepSat – A Learning framework for Satellite Imagery Saikat Basu Department

Integrating camera imagery, crowdsourcing, and deep ...ecoss.nau.edu/wp-content/uploads/2019/05/Kosmala... · Integrating camera imagery, crowdsourcing, and deep learning to improve

food menu drink - The Raven · SOFT DRINKS £ cans coca cola 2.20 diet coke 1.90 coke zero 1.90 irn bru 2.20 diet irn bru 1.90 san pellegrino orange 1.95 san pellegrino lImonata 1.95

DeepSat – A Learning framework for Satellite Imagery

Deep object classification in low resolution LWIR imagery ... · Deep object classi cation in low resolution LWIR imagery via transfer learning 4 dimension and training is continued

Enzymatic Assay of L-GLUTAMATE OXIDASE (EC 1.4.3.11) … 1.4.3.11 l-glutamate oxidase.pdf · Reagent D (Acetate) 1.90 1.90 1.90 1.90 1.90 1.90 Reagent E (MBTH) 0.80 0.80 0.80 0.80

VILLA CASTILLA LA MANCHA - Events of Excellence€¦ · villa castilla la mancha . 16 suites 1 room 1.90 x 1.90 bed individual 3 rooms 1.60 x 1,90 beds bathrooms 12 rooms 1,05 x 1.90

Spectral spatial feature learning for hyperspectral ... · Spectral–spatial feature learning for hyperspectral imagery classification using deep stacked sparse autoencoder Ghasem

Imagery, text and geospatial Machine Learning applications ... · Imagery, text and geospatial Machine Learning applications in Montreal's booming ML landscape Tom Landry 1, Samuel

Spatial Representations and Imagery in Learning · 2017. 4. 24. · in their mind’s eye. The process of working with these mental spatial representations is called, “imagery.”

Learning Sparse Codes for Hyperspectral Imagery€¦ · 1 Learning Sparse Codes for Hyperspectral Imagery Adam S. Charles, Bruno A. Olshausen, and Christopher J. Rozell Abstract—The

HEEG Conference - Visual Imagery and Digital Media in Learning

Connotation learning sequence English Stage 1 · Web vieweducation.nsw.gov.au. English S1 learning sequence – Connotation, imagery and symbol. Learning sequence description. Connotation,

Imagery Higher Analysis Learning Intention: I will understand how to write a comment on a writer’s use of imagery

TYPES OF DIGITAL VISUALS in E-Learning. Objectives Discuss dimensionality (1D to 4D) in digital imagery Highlight some affordances of digital imagery

Using publicly available satellite imagery and deep ...web.stanford.edu/~mburke/papers/yeh_et_al_2020.pdf · ARTICLE Using publicly available satellite imagery and deep learning to

Crime Mapping from Satellite Imagery via Deep …Crime Mapping from Satellite Imagery via Deep Learning Alameen Najjar Shun’ichi Kaneko Yoshikazu Miyanaga Graduate School of Information

LP & HP6 Blow Out Preventers - Double E, Inc. · LP Ram LP-DF Ram 2 2 2 1.90/2.44 1.90/2.44 1.90/2.44 Standard Standard Standard H2S Standard Standard Standard H2S Standard Standard

Building Damage Detection in Satellite Imagery Using ...imagery. Cooner et al. [3] compared the performance of multiple machine learning methods in building damage detection with both

Guided Imagery in the Curriculum - · PDF fileGuided Imagery in the Curriculum ... for Integrative Learning, Long Beach, Cali ... their own power, recognizing latent caPublished in:

Learning about numbers with patterns - Oxford …lib.oup.com.au/.../research/Learningaboutnumberswithpatterns.pdf · Learning about numbers with patterns using structured visual imagery

Learning How to Generate Kinesthetic Motor Imagery Using a

214 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE … · assessment, remotely sensed (RS) imagery, semilabeled sam-ples, semisupervised learning, supervised learning, unsupervised learning

VISUALIZATION OF DEEP TRANSFER LEARNING IN SAR IMAGERY …

Imagery Intelligence –IMINT RAF Aerial Imagery of … RAVERSIJDE-23mai.pdf · Imagery Intelligence –IMINT RAF Aerial Imagery of the Atlantic Wall Imagery Analysis seen from a

Running Head: OPTIMAL TIMING OF MENTAL IMAGERY · Mental imagery (MI) has been found to produce improvements in motor skills (Driskell ... Timing, Learning, Emotion, and Perspective)

Pedestrian Detection in Low-resolution Imagery by … ·  · 2014-05-18Pedestrian Detection in Low-resolution Imagery by Learning Multi-scale Intrinsic Motion Structures (MIMS) Jiejie

A physical learning companion for Mental-Imagery BCI User

Applying Deep Learning in Street View Imagery for

Connotation learning sequence English Stage 3€¦  · Web view2021. 7. 4. · English Stage 3 learning sequence – connotation, imagery and symbol. Learning sequence description