LARGE SCALE CROP CLASSIFICATION USING GOOGLE EARTH ENGINE PLATFORM

AndriiShelestov1,2,MykolaLavreniuk1,2,NataliiaKussul1,2,AlexeiNovikov2,SergiiSkakun3,4
1
Space Research Institute NASU-SSAU, Kyiv, Ukraine
2
National Technical University of Ukraine “Igor Sikorsky Kiev Polytechnic Institute”, Kyiv, Ukraine
3
Department of Geographical Sciences, University of Maryland, College Park, MD, USA
4
NASA Goddard Space Flight Center Code 619, Greenbelt, MD, USA

ABSTRACT regular basis. These new opportunities allow to build high-

 resolution LULC maps on regular basis and to assess land
For many applied problems in agricultural monitoring and use/land cover changes for large territories [10]. But with
food security it is important to provide reliable crop launches of Sentinel-1, Sentinel-2, Proba-V and Landsat-8
classification maps in national or global scale. Large amount satellites, there will be generated up to petabyte of raw
RIVDWHOOLWHGDWDIRUODUJHVFDOHFURSPDSSLQJJHQHUDWHD³%LJ images per year. The increasing volume of remote sensing
'DWD´SUREOHP7KHPDLQLGHDRIWKLVSDSHUZDVFRPSDULVRn GDWD GXEEHG DV D ³%LJ 'DWD´ SUREOHP FUHDWHV QHZ
of pixel-based approaches to crop mapping in Ukraine and challenges in handling datasets that require new approaches
exploring efficiency of the Google Earth Engine (GEE) to extracting relevant information from remote sensing (RS)
FORXG SODWIRUP IRU VROYLQJ ³%LJ 'DWD´ SUREOHP DQG data from data science perspective [11±13]. Generation of
providing high resolution crop classification map for large high resolution crop maps for large areas (>10,000 sq. km)
territory. The study is carried out for the Joint Experiment of using Earth observation data from space requires processing
Crop Assessment and Monitoring (JECAM) test site in of large amount of satellite images. Images acquired at
Ukraine covering the Kyiv region (North of Ukraine) in different dates during crop growth period are usually
2013. We found that Google Earth Engine (GEE) provided required to discriminate certain crop types. The following
very good performance in enabling access to remote sensing issues should be addressed while providing classification of
products through the cloud platform, but our own approach multi-temporal satellite images for large areas: (i) non-
based on ensemble of neural networks outperformed SVM, uniformity of coverage of ground truth data and satellite
decision tree and random forest classifiers that are available scenes; (ii) seasonal differentiation of crop groups (e.g.
in GEE. winter and summer varieties) and the need for incremental
classification; (iii) the need to store, manage and seamlessly
Index Terms ² Google Earth Engine, big data, process large amount of data (big data issues).
classification, optical satellite imagery, image processing. Therefore, in this paper we propose to investigate Google
 Earth Engine (GEE) cloud platform capabilities for large
1. INTRODUCTION scale crop mapping using Landsat-8 optical imagery,
compare different classification methods from GEE to our
The information on land use/land cover distribution over the own neural network approach [14±18]. Results are presented
big areas during the long-time periods is extremely for the Joint Experiment of Crop Assessment and
important for many environmental and monitoring tasks, Monitoring (JECAM) test site in Ukraine with the area of
including climate change, ecosystem dynamics analysis, more than 28,000 km2.
food security and others. Reliable crop maps can be used for
more accurate agriculture statistics estimation [1±3], 2. STUDY AREA AND MATERIALS DESCRIPTION
stratification purposes [4], better crop yield prediction [5±6]
and drought risk assessment [7±8]. During the last decades, Ukraine is one of the most developed agricultural countries
satellite imagery became the most promising data source for in the world. According to the U.S. Department of
solving such important tasks as land use/land cover Agriculture (USDA) Foreign Agricultural Service (FAS)
mapping. Yet, there are no globally available satellite- statistics, Ukraine was the largest sunflower producer (11.6
derived crop specific maps at present. Only coarse- MT) and exporter, and the ninth largest wheat producer
resolution imagery (> 250 m spatial resolution) has been (22.2 MT) in the world in 2013. The proposed investigation
utilized to derive global cropland extent (e.g. GlobCover, is evaluated for the Joint Experiment for Crop Assessment
MODIS) [9]. Now wide range of satellites provides and Monitoring (JECAM) test site in Ukraine. The JECAM
objective, open and free high special resolution data on a test site in Ukraine was established in 2011 and covers the

‹,(((  ,*$566

administrative region of Kyiv region with the geographic 2.1 Composite products available at GEE
area of 28,100 km2 and almost 1.0 M ha of cropland (Fig.
1). Major crops include: winter wheat, maize, soybeans, Different composites derived from Landsat-8 imagery and
vegetables, sunflower, barley, winter rapeseed, and sugar available in GEE were analyzed in the study. Landsat 8 8-
beet. The crop calendar is September-July for winter crops, Day Top-of-atmosphere (TOA) Reflectance Composites
and April-October for spring and summer crops. Fields in were used from GEE. As to the time of composition, 8 day
Ukraine are quite large with size generally up to 250 ha. composites were selected over 32 day composites. The
reason for that is that 32 composites are composed based on
the latest image, and this latest image can be of not the best
quality. All the images from each 8-day period are included
in the composite, with the most recent pixel on top.

2.2 Landsat­8 data pre­processing (outside GEE)

The following pre-processing steps are applied for all

Landsat-8 images: conversion of digital numbers (DNs)
Fig. 1. Location of Ukraine and JECAM test site in Ukraine values to the top-of-atmosphere (TOA) reflectance values,
(Kyiv region, marked with bold boundaries) conversion from the TOA reflectance to the surface
reflectance (SR) using the Simplified Model for
Ground surveys to collect data on crop types and other land Atmospheric Correction (SMAC), detection of clouds and
cover classes were conducted in 2013 in Kyiv region. shadows were done using Fmask algorithm, reconstruction
European Land Use and Cover Area frame Survey (LUCAS) from missing pixel values (clouds and shadows) using self-
nomenclature was used in this study as a basis for land cover organizing Kohonen maps (SOMs) [14, 19].
/ land use types. In total, 386 polygons were collected These pre-processing steps were performed outside
covering the area of 22,700 ha. Data were collected along GEE platform. After these products were generated they
the roads using mobile devices with built-in GPS. All were uploaded in the GEE cloud platform for the further
surveyed fields are randomly divided into training set (50%) classification using classification algorithm available in
to train the classifier and testing set (50%) for testing GEE.
purposes (Fig. 2). Fields are selected in such a way so there
is no overlap between training and testing sets. All
3. METHODOLOGY
classification results, in particular overall accuracy (OA),
XVHU¶V8$DQGSURGXFHU¶V3$DFFXUDFLHVDUHUHSRUWHGIRU In the study classification was done on a per-pixel basis.
testing set. The input features are classified into one of the Multiple classification techniques were evaluated in the
13 classes. study. At first, all classification algorithms available in GEE
were analyzed and used for classifying multi-temporal 8-
days Landsat-8 TOA composites from GEE. Then the best
classification algorithms in terms of overall classification
accuracy were compared with the neural network classifier
that used multi-temporal SR values generated outside GEE.
The presented approaches were compared in terms of
classification accuracy at pixel level. GEE offers several
classification algorithms among which are decision trees and
UDQGRP IRUHVWV VXSSRUW YHFWRU PDFKLQH 690 DQG 1DwYH
Bayes classifier. It should be noted that neural network (NN)
classifiers are not available in GEE.
Fig. 2. Location of fields surveyed during ground
measurements in 2013.
3.1 Ensemble of Neural Networks
Remote sensing images acquired by the Operational Land
Our proposed neural network approach based on committee
Imager (OLI) sensor aboard Landsat-8 satellite were used
of NNs, in particular Multi-Layer Perceptron (MLPs), is
for crop mapping over the study region. Only bands 2
utilized to improve performance of individual classifiers.
through 7 were used for crop classification maps providing
The MLP classifier has a hyperbolic tangent activation
[10]. Three scenes with path/row coordinates 181/24,
function for neurons in the hidden layer and logistic
181/25 and 181/26 covered the test site region.
activation function in the output layer. A committee of
MLPs was used to increase performance of individual


classifiers. The committee is formed using four MLPs with compared to the committee of NN that was implemented
different number of hidden neurons (10, 20, 30, and 40) outside GEE, in the Matlab environment using a Netlab
trained on the same training data within 250 epochs. Outputs toolbox. CART provided overall accuracy 76.9%, RF
from different MLPs were integrated using the technique of achieved 69.9% and committee of MLPs considerably
average committee. Under this technique, the average class outperformed DT-based classifiers: by +14.8% RF and by
probability over classifiers is calculated, and the class with +7.8% DT (Fig. 3).
the highest average posterior probability for the given input
sample is selected. In order to prevent NN overfitting, we
exploited early stopping and weight decay (L2
regularization) techniques. Coefficient of regularization was
selected from 0.00005, 0.0001, 0.0003, 0.001, 0.003, 0.01,
and 0.03 using a validation set (20% of training set).
This approach is used as a benchmark for assessing
classification techniques available in GEE.

4. RESULTS

4.1 Input (product) selection

The first set of experiments was carried out to select the best
input (TOA 8-day composites or restored values) and
evaluating different classifiers available in GEE. Table 1
shows the derived OA on polygons from a testing set using
TOA 8-day composites as inputs. The best performance was Figure 3. Final map obtained by classifying multi-temporal
achieved for CART at 75%. Somewhat surprisingly, an Landsat-8 imagery using a committee of MLP classifiers.
ensemble of DTs, i.e. RF, was outperformed by CART and
yielded only 68%. Logistic regression (GMO Max Entropy) 5. DISCUSSION AND CONCLUSIONS
gave 72% accuracy. Linear classifiers, MultiClassPerceptron
and Winnow, provided up to 60% accuracy, while variants The GEE platform offers powerful capabilities in handling
of SVM achieve moderate accuracy of 57%. Unfortunately, large volumes of remote sensing imagery that can be used,
it was unable to produce stable classification results for for example, for classification purposes and crop mapping
SVM classifiers which usually resulted into the Internal for large territories. The best OA achieved on composites
Server Error on invocation from Python. from the GEE was 75%, while on atmospherically corrected
and restored images the achieved accuracy was almost 77%.
Table 1. Overall classification accuracy achieved by GEE The GEE platform provides a set of classification
classifiers for TOA 8-day composites as an input algorithms. The best results in the GEE were obtained for
Classifier OA, % the DT-based classifiers, namely CART and RF. At the same
time, committee of neural networks considerably
CART 75 outperformed DT-based classifiers: by +14.8% RF and by
GMO Max Entropy 72 +7.8% DT. Research within this paper were targeted on the
Random Forest 68 comparison of pixel-based approaches to crop mapping in
MultiClassPerceptron 60 Ukraine and exploring efficiency of the Google Earth
Engine (GEE) cloud platform for large scale crop mapping.
IKPamir 57 In general, GEE provided very good performance in
Winnow 49 enabling access to remote sensing products through the
FastNaiveBayes 32 cloud platform and powerful computational resources that
Pegasos - could help users to deal with ³%LJ 'DWD´ SUREOHP in crop
mapping for large territory.
VotingSvm -
MarginSvm - 6. ACKNOWLEDGEMENTS

4.2 Classifier selection This research was conducted in the framework of the ³Large
scale crop mapping in Ukraine using SAR and optical data
One of the best GEE classifiers (CART and RF) on fusion´Google Earth Engine Research Award funded by the
atmospherically corrected Landsat-8 imagery were Google Inc.


 7. REFERENCES terrestrial global change research´ Remote Sensing of
Environment, vol. 145, pp. 154-172, 2014.
[1] J. Gallego, E. Carfagna, and B. Baruth, ³Accuracy,
objectivity and efficiency of remote sensing for agricultural [11] Y. Ma, HW DO ³Remote sensing big data computing:
statistics´ Agriculturalsurveymethods, pp. 193-211, 2010. challenges and opportunities´ FutureGenerationComputer
Systems, vol. 51, pp. 47-60, 2015.
[2] J. Gallego, A. N. Kravchenko, N. N. Kussul, S. V.
Skakun, A. Y. Shelestov, and Y. A. Grypych, ³Efficiency [12] Y. Ma, et al. ³Towards building a data-intensive index
assessment of different approaches to crop classification for big data computing±A case study of Remote Sensing
based on satellite and ground observations´ Journal of GDWD SURFHVVLQJ´ Information Sciences, vol. 319, pp. 171-
AutomationandInformationSciences, vol. 44, no. 5, 2012. 188, 2015.

[3] F. J. Gallego, N. Kussul, S. Skakun, O. Kravchenko, A. [13] N. Kussul, A. Shelestov, R. Basarab, S. Skakun, O.
Shelestov, and O. Kussul ³Efficiency assessment of using Kussul, and M. Lavrenyuk, ³Geospatial Intelligence and
satellite data for crop area estimation in Ukraine´ Data Fusion Techniques for Sustainable Development
International Journal of Applied Earth Observation and Problems´InICTERI, pp. 196-203, 2015.
Geoinformation, vol. 29, pp. 22-30, 2014.
[14] S. Skakun, N. Kussul, A. Y. Shelestov, M. Lavreniuk,
[4] C. G. Boryan and Z. Yang, ³Deriving crop specific DQG2.XVVXO³(IILFLHQF\$VVHVVPHQWRI0XOWLWHPSRUDO&-
covariate data sets from multi-year NASS geospatial Band Radarsat-2 Intensity and Landsat-8 Surface
cropland data layers´InternationalGeoscienceandRemote Reflectance Satellite Imagery for Crop Classification in
SensingSymposium(IGARSS), pp. 4225-4228, 2013. 8NUDLQH´IEEEJ.ofSelect.TopicsinAppliedEarthObser.
andRem.Sens., vol. 9, no. 8, pp. 3712-3719, 2016.
[5] F. Kogan et al., ³Winter wheat yield forecasting in
Ukraine based on Earth observation, meteorological data [15] F. Waldner, HW DO ³Towards a set of agrosystem-
and biophysical models´ International Journal of Applied specific cropland mapping methods to address the global
Earth Observation and Geoinformation, vol. 23, pp. 192- cropland diversity´ International Journal of Remote
203, 2013. Sensing, vol. 37, no. 14, pp. 3196-3231, 2016.

[6] A. Kolotii, N. Kussul, A. Shelestov, S. Skakun, B. [16] S. V. Skakun, E. V. Nasuro, A. N. Lavrenyuk, and O.
Yailymov, R. Basarab, M. Lavreniuk, T. Oliinyk, and V. M. Kussul, ³Analysis of applicability of neural networks for
2VWDSHQNR ³&RPSDULVRQ RI ELRSK\VLFDO DQG VDWHOOLWH classification of satellite data´ Journal of Automation and
SUHGLFWRUVIRUZKHDW\LHOGIRUHFDVWLQJLQ8NUDLQH´Int.Arch. InformationSciences, vol. 39, no. 3, 2007.
Photogramm.RemoteSens.SpatialInf. Sci., XL-7/W3, pp.
39-44, 2015. DOI: 10.5194/isprsarchives-XL-7-W3-39- [17] M. S. Lavreniuk, S. V. Skakun, A. J. Shelestov, B. Y.
2015. Yalimov, S. L. Yanchevskii, D. J. Yaschuk, and A. I.
.RVWHFNL\³/DUJH-Scale Classification of Land Cover Using
[7] S. Skakun, N. Kussul, A. Shelestov, and O. Kussul, ³The RetrospectiYH 6DWHOOLWH 'DWD´ Cybernetics and Systems
use of satellite data for agriculture drought risk Analysis, vol. 52, no. 1, pp. 127±138, 2016.
quantification in Ukraine´ Geomatics,NaturalHazardsand
Risk, vol. 7, no. 3, pp. 901-917, 2016. [18] N. Kussul, G. Lemoine, F. J. Gallego, S. V. Skakun, M.
/DYUHQLXN DQG $ < 6KHOHVWRY ³3DUFHO-Based Crop
[8] N. N. Kussul, B. V. Sokolov, Y. I. Zyelyk, V. A. Classification in Ukraine Using Landsat-8 Data and
Zelentsov, S. V. Skakun, and A. Y. Shelestov, ³Disaster risk Sentinel-$'DWD´IEEEJ.ofSelect.TopicsinAppl.Earth
assessment based on heterogeneous geospatial information´ Observ.andRem.Sens., vol. 9, no. 6, pp. 2500±2508, 2016.
Journal of Automation and Information Sciences, vol. 42,
no. 12, 2010. [19] N. Kussul, S. Skakun, A. Shelestov, M. Lavreniuk, B.
<DLO\PRY DQG 2 .XVVXO ³5HJLRQDO 6FDOH Crop Mapping
[9] S. Fritz et al., ³The need for improved maps of global Using Multi-7HPSRUDO 6DWHOOLWH ,PDJHU\´ ,QW $UFK
cropland´Eos,TransactionsAmericanGeophysicalUnion, Photogramm. Remote Sens. Spatial Inf. Sci., XL-7/W3, pp.
vol. 94, no. 3, 31-32, 2010. 45±52, 2015. DOI:10.5194/isprsarchives-XL-7-W3-45-
2015.
[10] D. P. Roy, M. A. Wulder, T. R. Loveland, C. E.
Woodcock, R. G. Allen, M. C. Anderson, ... and T. A.
Scambos ³Landsat-8: Science and product vision for