link to pdf

Crop classification and field area estimation in Ukraine Nataliia Kussul*, Serhiy Skakun**, Mykola Lavreniuk*, *Space Research Institute, Ukraine **Integration-Plus Ltd 1

About us • Space Research Institute National Academy of Science & State Space Agency of Ukraine • Active participation in the Working Group on Information System and Services (WGISS) of the Committee on Earth Observation Satellites (CEOS). • Participation in international collaborative activities within GEO Working Plan • UN-SPIDER RSO, WDS The 8th GEOSS AP Symposium 9-11 Sep 2015 Beijing, China

2

Introduction • Agriculture in Ukraine – 1st world largest sunflower producer and exporter (in 2013-2014) – 8th world largest wheat producer (in 2014) [source: FAS USDA]

• Crop mapping – important input within many problems solving • crop area estimation • crop yield forecasting • environmental impact analysis

The 8th GEOSS AP Symposium 9-11 Sep 2015 Beijing, China

Main goal To develop crop classification methodology for high resolution land cover mapping in Ukraine with multi-temporal optical and multi-temporal multi-polarization SAR images using a neural network ensemble.


Information processing chain 1. No-data pixels restoration (clouds and shadows).

2. Universal machine learning time series classification for the regional level

3. Map filtration


5

Samples data for land cover № 1 2 3 4 5 6

Class Artificial Cropland Forest Grassland Bare land Water

1990 Polygons No. % 29 0,22 6278 48,17 2338 17,94 2471 18,96 315 2,42 1601 12,29

2000 Polygons No. % 14 0,12 6000 49,42 2340 19,27 2084 17,16 180 1,48 1524 12,55

2010 Polygons No. % 38 0,32 5942 50,96 2332 20,00 1801 15,44 156 1,34 1392 11,94

Total

13032

12142

11661

Training set

Test set

Luhansk Oblast


Missing Data Restoration in Time Series of Images Missing

Input X1

X2

X3

Nan

X5

Nan

Missing components are taken from neuron winner

wi1

wi2

wi3

wi4

wi5

Landsat-8 (08.07.14)

wi6

SOM: selection of neuron winner X1

X2

X3

Nan

X5

Nan

wl1

wl2

wl3

wl4

wl5

wl6

Only valid components are considered for finding a neuron winner

Landsat-8 restored (08.07.14)


Classification Method based on Machine Learning • •

Ensemble of neural networks (multilayer perceptron - MLP) MLP training: minimized the crossentropy error function N K

E (w1 ,..., w K ) = − ln p (T | w1 ,..., w K ) = −∑∑ t nk ln ynk

•

•

Samples randomly divided into training (50%) and testing (50%) sets

•

Restoration of missing data due to clouds and shadows using a selforganizing Kohonen maps (SOMs)

•

Inputs to classifiers

n =1 k =1

Ensemble

– averaging a-posteriori probability of individual networks

1 L l p = ∑ pi L l =1 e i

k * = arg max pke k =1, K

– Landsat-8 (TOA or TOC) • bands 2-7 – Proba-V (TOC) • bands 1-4 + NDVI – Sentinel-1 • bands 1-2 (VV+VH) • DN -> backscatter coeff • filter: Lee (5 x 5) • ortho-rectification


8

JECAM experiments


9

Classification maps analysis 1990

2000

2010 2000 1990 2010


Ground Surveys for Crop Classification Ground surveys, 2013-2015 № 1 2 3 4 5 6 7 8 9 10 11 12 13

Class Artificial Winter wheat Winter rapeseed Spring crops Maize Sugar beet Sunflower Soybeans Other cereals Forest Grassland Bare land Water

2013 2014 Polygons Polygons No. % No. % 6 1.6 17 2.8 51 13.2 126 20.4 12 3.1 36 5.8 9 2.3 45 7.3 87 22.5 76 12.3 8 2.1 18 2.9 30 7.8 31 5.0 60 15.5 109 17.6 32 8.3 12 1.9 17 4.4 35 5.7 48 12.4 68 11.0 10 2.6 11 1.8 16 4.1 34 5.5

Total

386

Data randomly divided into training (50%) and testing (50%) sets

2014

Kyiv oblast (area 28,131 km2)

618


11

Early season crop classification (2014) Landsat-8

Proba-V

Early stage winter crops mapping for Kyiv region (area 28,000 km2) using Landsat-7 (10 March) and Landsat-8 (09 March) in 2014


Multi mission crop classification (2015)

Kyiv oblast Satellite Landsat 8 Sentinel 1 L8+Sent1


OA 76,7% 79,6% 84.5%

Maps filtration (Kyiv region, 2013) 16557 parcels After pixel-based classification two post processing approaches were tested in order to derive parcel-based classification maps 1. A majority voting scheme; 2. A majority voting scheme incorporating information on the cloud cover.


Experiments (Kyiv region, 2013) 1. Pixel-based approach land cover classification (OA=85.32%) 2. Parcel-based approaches using vector information: –

–

Input features are averaged within the parcel and are input to the classifier which assigns the corresponding class to the parcel (OA=78.18%); vector information about parcels and raster pixel-based classification map data fusion depending on decision-making methods:  

Majority voting (OA=87.70%); weighted majority voting (take into account restoring pixels class and decrease its impact) (OA=89.40%);


Parcel-based methods Cloud cover

A majority voting scheme

Classification map

A majority voting scheme using cloud cover information


Parcel analysis (Kyiv region, 2013) • Method that divides parcel into the fields (OA=87.70%) based on geometry analysis of each object in the parcel and it’s structure similarity to the field; • Using dividing method with cloud coverage information improve OA=89.4%. 1004 of 16557 parcels were divided into the fields.


Results (Kyiv region, 2013) A majority voting scheme

Pixel-based classification map

Method that divides parcel into the fields


Results (Kyiv region, 2013)


Comparison to official statistics Comparison of crop area from official statistics with Landsat-8 derived using a pixel-based classification and parcel-based approaches (Kyiv, 2013)

Crop

Winter wheat Winter rapeseed

Pixel-based Official approach statistics (1000 ha / (1000 ha) error %)

Majority voting (1000 ha / error %)

Weighted majority voting (1000 ha / error %)

Method that divides Method that parcel into divides the fields parcel into using clouds the fields coverage (1000 ha / information error %) (1000 ha / error %)

187.3

184.5 / - 1.5 168.2 / -10.2 168.7 / -9.9 164.8 / -12.0 165.2 / -11.8

46.7

59.9 / 28.3

Maize

291.7

342.4 / 17.4 338.9 / 16.2 355.3 / 21.8 340.7 / 16.8 357.8 / 22.6

Sugar beet

15.5

11.2 / -27.7 10.8 / -30.4 11.1 / -28.2 10.5 / -31.9 10.8 / -30.1

Sunflower

108.2

117.6 / 8.7 125.2 / 15.7 125.8 / 16.3 119.3 / 10.2 119.5 / 10.5

Soybeans

145.9

168.5 / 15.5 148.3 / 1.6

54.1 / 15.8

54.0 / 15.5

53.6 / 14.8

53.5 / 14.5

136.4 / -6.5 145.4 / -0.3 134.2 / -8.0


Comparison to official statistics, cont

Crop

Winter wheat Winter rapeseed

Pixel-based approach (Accuracy UA / PA)

Majority voting (Accuracy UA / PA)

Weighted majority voting (Accuracy UA / PA)

Method that divides Method that parcel into divides the fields parcel into using clouds the fields coverage (Accuracy information UA / PA) (Accuracy UA / PA)

91.8 / 95.7

94.8 / 98.2

94.8 / 98.2

94.9 / 98,8

94.9 / 98,8

99.4 / 93.5

99.5 / 96.5

99.5 / 96.5

99.5 / 96.5

99.5 / 96.5

86.8 / 90.5

84.9 / 93.4

89.7 / 93.4

86.6 / 93.2

91.6 / 93.2

Sugar beet 89.6 / 94.9

99.9 / 97.2

99.9 / 97.2

99.6 / 99.9

99.6 / 99.9

Sunflower 85.4 / 84.1

96.4 / 82.2

96.4 / 82.2

91.2 / 82.0

91.2 / 82.0

Soybeans

77.1 / 69.7

80.3 / 74.2

82.2 / 84.6

80.2 / 73.9

82.2 / 84.3

88.2

91.5

92.9

91.4

92.8

Maize

MEAN


21

Prospects: Project H2020 ERA-PLANET ERA-PLANET THE EUROPEAN NETWORK FOR OBSERVING OUR CHANGING PLANET H2020 Topic SC5-15-2015: Strengthening the European Research Area in the domain of Earth Observation

The goal

of ERA-PLANET is to strengthen the European Research Area in the domain of Earth Observation in coherence with the European participation to Group on Earth Observation (GEO) and Copernicus.

The expected impact is to strengthen the European leadership within the forthcoming GEO 2015-2025 Work Plan. 41 organizations with 11 milling budget by EC (European Commision)


Prospects: Directions ERA-PLANET


Interoperability


Discussion and conclusions 1. Universal machine learning time series based methodology of land cover & crop classification

2. Multi-mission heterogeneous observations (optical and SAR) and data fusion (raster and vector information)

3. International collaboration for GEOSS benefits


25

References •

•

•

“Efficiency Assessment of Multitemporal C-Band Radarsat-2 Intensity and Landsat8 Surface Reflectance Satellite Imagery for Crop Classification in Ukraine” S.Skakun, N.Kussul, A.Y. Shelestov, M.Lavreniuk, O. Kussul IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 2015. (http://dx.doi.org/10.1109/JSTARS.2015.2454297) “Regional scale crop mapping using multi-temporal satellite imagery“ N. Kussul, S. Skakun, A. Shelestov, M. Lavreniuk, B. Yailymov, O. Kussul International Archives of the Photogrammetry, Remote Sensing & Spatial Information Sciences. – 2015. - P. 45-52. – (http://dx.doi.org/10.5194/isprsarchives-XL-7-W3-45-2015) “Comparison of biophysical and satellite predictors for wheat yield forecasting in Ukraine” A. Kolotii, N. Kussul, A. Shelestov, S. Skakun, B. Yailymov, R. Basarab, M. Lavreniuk, T. Oliinyk, V. Ostapenko International Archives of the Photogrammetry, Remote Sensing & Spatial Information Sciences. – 2015. - P. 39-44. (http://dx.doi.org/10.5194/isprsarchives-XL-7-W3-39-2015) The 8th GEOSS AP Symposium 9-11 Sep 2015 Beijing, China

26

References • “The use of satellite data for agriculture drought risk quantification in Ukraine” Sergii Skakun, Nataliia Kussul, Andrii Shelestov, Olga Kussul Geomatics, Natural Hazards and Risk. – 2015. – P. 1-18. (http://dx.doi.org/10.1080/19475705.2015.1016555) • “Efficiency assessment of using satellite data for crop area estimation in Ukraine“ J. Gallego, N. Kussul, S. Skakun, O. Kravchenko, A. Shelestov, O. Kussul International Journal of Applied Earth Observation and Geoinformation. - 2014. - No. 29. - P. 22 – 30. (http://dx.doi.org/10.1016/j.jag.2013.12.013) • “Geospatial information system for agricultural monitoring” Shelestov A.Yu., Kravchenko A.N., Skakun S.V., Voloshin S.V., Kussul N.N. Cybernetics and Systems Analysis. - 2013. - Vol. 49, No. 1. - P. 124-132.(http://dx.doi.org/10.1007/s10559-013-9492-5) The 8th GEOSS AP Symposium 9-11 Sep 2015 Beijing, China

27

Thank you! [email protected]


28