See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/315939269

Deep Learning Classification of Land Cover and Crop Types Using Remote
Sensing Data

Article  in  IEEE Geoscience and Remote Sensing Letters · March 2017

DOI: 10.1109/LGRS.2017.2681128

CITATIONS READS
1,323 23,269

4 authors:

Nataliia Kussul Mykola Lavreniuk

National Technical University of Ukraine Kyiv Polytechnic Institute National Academy of Sciences of Ukraine
255 PUBLICATIONS   6,163 CITATIONS    103 PUBLICATIONS   3,409 CITATIONS   

SEE PROFILE SEE PROFILE

Sergii V. Skakun Andrey Yu. Shelestov

University of Maryland, College Park National Technical University of Ukraine Kyiv Polytechnic Institute
199 PUBLICATIONS   6,743 CITATIONS    179 PUBLICATIONS   4,494 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Geospatial models and information technologies of satellite monitoring of smart city problems View project

Amazon and GEO Project "Methodology for SDGs indicators assessment" View project

All content following this page was uploaded by Andrey Yu. Shelestov on 07 November 2017.

The user has requested enhancement of the downloaded file.

 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1

Deep Learning Classification of Land Cover and

Crop Types Using Remote Sensing Data
Nataliia Kussul, Mykola Lavreniuk, Sergii Skakun, and Andrii Shelestov

Abstract— Deep learning (DL) is a powerful state-of-the-art multisources data fusion techniques [4], [5]. Land cover and
technique for image processing including remote sensing (RS) crop type maps are one of the most essential inputs when
images. This letter describes a multilevel DL architecture that dealing with environmental and agriculture monitoring tasks
targets land cover and crop type classification from multitempo-
ral multisource satellite imagery. The pillars of the architecture [6]–[8]. Multitemporal multisource satellite imagery is usually
are unsupervised neural network (NN) that is used for optical required in order to capture specific crop growth stages and
imagery segmentation and missing data restoration due to clouds thus being able to discriminate different crop types. For exam-
and shadows, and an ensemble of supervised NNs. As basic ple, multispectral optical imagery only might not be enough
supervised NN architecture, we use a traditional fully connected to discriminate summer crops in a complex and heterogeneous
multilayer perceptron (MLP) and the most commonly used
approach in RS community random forest, and compare them environment. For this, SAR-derived information adds an added
with convolutional NNs (CNNs). Experiments are carried out value that allows discrimination of particular crop types [9],
for the joint experiment of crop assessment and monitoring test [10].
site in Ukraine for classification of crops in a heterogeneous A comprehensive study on the state-of-the-art supervised
environment using nineteen multitemporal scenes acquired by pixel-based methods for land cover mapping was performed
Landsat-8 and Sentinel-1A RS satellites. The architecture with
an ensemble of CNNs outperforms the one with MLPs allowing by Khatami et al. [11]. They found that support vector
us to better discriminate certain summer crop types, in particular machine (SVM) was the most efficient for most applications
maize and soybeans, and yielding the target accuracies more than with an overall accuracy (OA) of about 75%. The second
85% for all major crops (wheat, maize, sunflower, soybeans, and method with approximately the same efficiency (74% of OA)
sugar beet). was a neural network (NN)-based classifier. In that study,
Index Terms— Agriculture, convolutional neural classification was done only for a single date image. At the
networks (CNNs), crop classification, deep learning (DL), same time, SVM is too much resource consuming to be
joint experiment of crop assessment and monitoring (JECAM), used for big data applications and large area classification
Landsat-8, remote sensing (RS), Sentinel-1, TensorFlow, Ukraine.
problems. Another popular approach in the RS domain is the
random forest (RF)-based approach [12]. However, multiple
I. I NTRODUCTION features should be engineered to feed the RF classifier for the
efficient use.
T HE last several years and onward could be called the
years of Big Free Data in remote sensing (RS). During the
2013–2016 period, several optical and synthetic aperture radar
Over the past few years, the most popular and efficient
approaches for multisensor and multitemporal land cover
(SAR) RS satellites were launched with high spatial resolution classification are ensemble-based [13]–[16] and deep learn-
(10–30 m), in particular Sentinel-1A/B and Sentinel-2A within ing (DL) [17]–[20]. These techniques are found to outperform
the European Copernicus program [1], [2], and Landsat-8 the SVM [21]–[23]. DL is a powerful machine learning
within the Landsat Project, a joint initiative between the U.S. methodology for solving a wide range of tasks arising in image
Geological Survey (USGS) and the National Aeronautics and processing, computer vision, signal processing, and natural
Space Administration [3]. These data sets are freely available language processing [24]. The main idea is to simulate the
on operational basis. This opens unprecedented opportunities human vision to deal with big data problem, use all the data
for a wide range of preoperational and operational applications available and provide the semantic information at the output.
in the environment and agricultural domains taking advantage Plenty of models, frameworks and benchmark databases of ref-
of high temporal resolution data sets and advances in the erence imagery are available for image classification domain.
Over past years, more and more studies have been using DL
Manuscript received February 17, 2017; accepted March 6, 2017. for processing of RS imagery [25], [26]. DL proved to be
N. Kussul and M. Lavreniuk are with the Department of Space Information efficient for processing both optical (hyperspectral and mul-
Technologies and Systems, Space Research Institute, National Academy
of Sciences of Ukraine and SSA Ukraine, 03680 Kyiv, Ukraine (e-mail: tispectral imagery) and radar images, in extracting different
inform@ikd.kiev.ua; nataliia.kussul@gmail.com; nick_93@ukr.net). land cover types such as road extraction, buildings extrac-
S. Skakun is with the Department of Geographical Sciences, University of tion [17], [27], [28]. In terms of particular DL architectures,
Maryland, College Park, MD 20742 USA (e-mail: skakun@umd.edu).
A. Shelestov is with the Department of Information Security, National convolutional NNs (CNNs), deep autoencoders, deep belief
Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute,” networks, and recurrent NN with long short-term memory
03056 Kyiv, Ukraine (e-mail: andrii.shelestov@gmail.com). model have already been explored for RS tasks [17], [28]–[31].
Color versions of one or more of the figures in this letter are available
online at http://ieeexplore.ieee.org. It should be noted that most studies with DL for RS utilize
Digital Object Identifier 10.1109/LGRS.2017.2681128 a single date image for classification purposes, e.g., land
1545-598X © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

TABLE I TABLE II
D ATES OF A CQUISITION OF L ANDSAT-8 AND S ENTINEL -1A N UMBER AND A REA OF P OLYGONS C OLLECTED D URING
FOR THE K YIV R EGION IN 2015 G ROUND S URVEYS FOR THE K YIV R EGION IN 2015

cover or object detection. However, multitemporal images are

usually required to reliably identify specific land cover classes
such as crop types.
When providing large scale crop mapping using multi-
temporal satellite imagery, the following challenges should
be addressed while using DL. First, pixels of a satellite
image contain physical values. In particular, each pixel of
the optical imagery contains spectral reflectance values in ered calibration, multilooking (with 2 × 2 window), speckle
multiple spectral bands, and can be contaminated with clouds filtering (3 × 3 window with Refined Lee algorithm), and
and shadows, while each pixel of the spaceborne SAR imagery terrain correction using The Shuttle Radar Topography Mis-
is characterized by backscatter intensity and phase in multiple sion (SRTM) digital elevation model (DEM). A time-series
polarizations. Both of the data sources have multitemporal of six spectral bands from each Landsat-8 scene and two
nature and different spatial resolutions. That is why, DL imple- bands with combinations of vertical transmit and vertical
mentation for land cover and crops classification based on receive and combinations of vertical transmit and horizontal
data fusion of multitemporal multisensor satellite data is a receive polarizations from each Sentinel-1 scene is used as an
challenge. input to the classification model. In the study for the 2015
In this letter, we propose a multilayer DL architecture vegetation season, we used four images acquired by Landsat-
that is targeted for classification of multisource multitemporal 8 and 15 images acquired by Sentinel-1A. Ground truth data
RS images, both optical and SAR, at a pixel level. The core were collected during ground surveys in May–July of 2015
of the architecture is an ensemble of CNNs. The proposed to generate training and testing sets to train and validate the
architecture is applied for crop classification using Landsat-8 proposed classifier, respectively. Ground surveys were con-
and Sentinel-1A time-series and provides accuracy high ducted along the road as an adopted sampling design for the
enough (>85%) to be considered for operational context at joint experiment of crop assessment and monitoring (JECAM)
the national level [9]. experiments [34]. In total, 547 polygons of different classes
were collected (Table II). These polygons were randomly
II. S TUDY A REA AND M ATERIALS
divided into training (calibration) set (50%) and validation
We address the problem of land cover and crop clas- set (50%).
sification for Kyiv region of Ukraine using multitemporal
III. M ETHODOLOGY
multisource images acquired by Landsat-8 and Sentinel-1A
satellites. The study area is classified into eleven classes A. General Architecture Overview
including major agricultural crops (water, forest, grassland, A four-level architecture is proposed for classification of
bare land, winter wheat, winter rapeseed, spring cereals, crop types from multitemporal satellite imagery. These levels
soybeans, maize, sunflowers, and sugar beet). It is rather large are preprocessing, supervised classification, postprocessing,
area (28 000 square km) with big diversity of different land and geospatial analysis (Fig. 1).
cover types and agricultural crops. The territory is big enough Since optical satellite imagery can be contaminated with
to be considered as a representative one for the extension clouds and shadows, one have to deal with missing values in
of the technology to the entire country. Such technology the imagery. Most classifiers accept only valid pixel values
is particularly important taking into account national level as an input, and therefore a preprocessing step should be
demonstration within the ESA “Sentinel 2 for Agriculture” performed to impute (or fill gaps) missing values. This pro-
project (Sen2Agri), started in 2015. cedure is performed within level I of the architecture. The
For the 2015 vegetation season (since October 2014 till next step is supervised classification (level II) which is the
September 2015) four Landsat-8 and fifteen Sentinel-1 images core of this letter. We propose different CNNs architectures,
were acquired for the study area (Table I). Atmospherically namely, 1-D and 2-D, to explore spectral and spatial features,
corrected Landsat-8 images downloaded from the USGS earth- respectively. To our best knowledge, this is the first attempt
Explorer system were used in this letter [32]. The Landsat-8 to apply CNNs to multisource multitemporal satellite imagery
product is provided with cloud and shadow masks [33]. for crop classification. The CNNs architecture is compared
These regions were masked as areas without data. Sentinel-1A to the existing methods such as an ensemble of multilayer
images went through a preprocessing procedure that cov- perceptrons (MLPs) (ENN) and RF classifier. Levels III and IV
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KUSSUL et al.: DL CLASSIFICATION OF LAND COVER AND CROP TYPES 3

Fig. 2. Deep CNN architecture. (Top) 1-D CNN architecture.

(Bottom) 2-D CNN architecture.

classes with fully connected output layer. The detailed descrip-

Fig. 1. Four-level hierarchical DL model for satellite data classification
and land cover/land use changes analysis (I—preprocessing for dealing tion of the ENN approach is given in [34] and [36], while the
with missing data on optical images due to clouds/shadows; II—supervised CNN architecture is described in the next section. We used
classification; III—postprocessing using additional geospatial data to improve the Orfeo Toolbox implementation for the RF classifier.
classification maps; IV—geospatial analysis for a high-level product, e.g., crop
area estimation). 2) Supervised Classification With Convolutional Neural
Networks: The two bands from each of the fifteen
are aimed at improving the resulting classification map with Sentinel-1A scenes and the six bands from each of the four
available geospatial layers and building high-level products. Landsat-8 scenes form a CNN input feature vector with
The latter can be crop area estimation and crop rotation area dimension size 54 (15×2+4×6). Traditional CNNs (2-D) take
estimation. All these levels of the architecture are described into account a spatial context of an image and provide higher
in more detail in the following sections. accuracy comparing to a per pixel-based approach. However,
in this case, CNN smooths not only some misclassified pixel
but also small objects like roads, and forest “stripes” and
B. Level I: Preprocessing clear cuts within the forest (with linear dimensions of several
For preprocessing, we utilize self-organizing Kohonen pixels) are missed. In this letter, we compare two different
maps (SOMs) for optical images segmentation and subse- CNN architectures: 1-D CNN with convolutions in the spectral
quent restoration of missing data in a time-series of satellite domain [37], and 2-D CNN with convolutions in the spatial
imagery [35]. SOMs are trained for each spectral band sepa- domain. Each CNN in the corresponding ensemble consists
rately using nonmissing values. Missing values are restored of two convolutional layers, each of them followed by max
through a special procedure that substitutes input sample’s pooling and two fully connected layers in the end (Fig. 2).
missing components with neuron’s weight coefficients. Pixels We used a rectified linear unit (ReLU) function that is one
that have been restored are masked, the number of cloud-free of the most popular and efficient activation functions for deep
scenes available for each pixel from optical imagery is calcu- NNs. There are advantages of using ReLU such as biological
lated, and these two layers are used for further postprocessing plausibility, efficient computation, and gradient propagation.
procedure (at level III) to improve the resulting classification Therefore, ReLU function is faster and more effective for
map [16]. The detailed description of the restoration algorithm training CNNs comparing to a sigmoid function. In both
is given in [35] and [36]. architectures, there are five CNNs in the ensemble. Each of
the CNNs has the same convolution and max-pooling structure
but differs in the trained filters and number of neurons in the
C. Level II: Supervised Classification hidden layer being 60, 70, 80, 90, and 100 for five CNNs,
1) General Overview: The core element of the model is respectively.
the supervised classification, which is performed at the second Tikhonov’s L2 regularization, dropout with probability of
stage (level II). We explore two different paradigms: state-of- 0.5 and learning rate exponential decay techniques are used
the-art methods (RF an ENN) and compare those classifiers to prevent the overfitting problem and to generalize the loss
with proposed ensemble of CNNs. Each MLP represents a function. For loss function optimization, we used advanced
classical fully connected NN with a single hidden layer. The adaptive moment estimation method that is a combination
MLP transforms an input into a feature space by a hidden of AdaGrad and RMSProp methods and has faster con-
layer, and features are subsequently used to discriminate vergence comparing to the well-known methods such as
classes by the output layer. The CNN, in turn, builds a gradient descent, stochastic gradient descent, AdaGrad and
hierarchal set of features through local convolution and down- RMSProp [38]. Batch learning technique with sample size 32
sampling. The resulting feature maps are used to discriminate is used to speed-up the NN training phase. Multiclass cross
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS

entropy function is used as a loss function and softmax

function is used to provide the a posteriori probability for
each class. Ensembles of 1-D and 2-D CNNs are implemented
using the Google’s library TensorFlow [39].
The proposed 1-D CNN architecture is able to provide clas-
sification for each pixel of the input image. In turn, a classical
2-D CNN provides a class for a window with size of 7 × 7
pixels. In this case, there is a popular approach with a direct
up-scaling to match the input size that will actually reduce
the spatial resolution of the classification map (at the order of
four times for two down-scaling layers with filter size 2 × 2).
In our experiment, we utilized a sliding window technology
with a 1 pixel step and the resulting class was assigned to
the central pixel of the sliding window. Therefore, the output
classification map had the same spatial resolution, but even in
this case some small objects were smoothed and misclassified.
D. Level III and IV: Postprocessing and Geospatial Analysis
To improve the quality of the resulting map, we developed
several filtering algorithms, based on the available information
on quality of input data and fields boundaries, for example
parcels [16]. Those filters take a pixel-based classification
map and specifically designed rules to account for several
plots (fields) within the parcel. In the result, we obtained a
clear parcel-based classification map. The final level of data Fig. 3. Example of classification result for the Kyiv region for 2015 based
on all Landsat-8 and Sentinel-1A images.
processing provides data fusion with multisourced heteroge-
neous information, in particular, statistical data, vector geospa-
TABLE III
tial data, socio-economic information, and so on. It allows
C OMPARISON OF PA, UA, AND OA FOR THE RF, ENN,
interpreting the classification results, solving applied problems 1-D CNNs, AND 2-D CNNs FOR K YIV REGION IN 2015
for different domains, and providing the support information
for decision makers. For example, classification map coupled
with area frame sampling approach can be used to estimate
crop areas [40].
IV. R ESULTS
Overall classification accuracies for RF, ENN, ensemble of
1-D and 2-D CNNs were 88.7%, 92.7%, 93.5%, and 94.6%,
respectively (Table III). User’s and producer’s accuracies
(UA and PA) provided by an ensemble of 2-D CNNs were the
highest for all classes. The RF classifier provided the lowest
accuracies comparing to NN-based approaches. Accuracy for
winter rapeseed, spring crops, sunflower, forest, and water
did not vary significantly with different approaches. At the
same time, major improvements of using CNNs comparing to
RF were achieved for maize, sugar beet, soybeans, grassland,
and bare land (Fig. 3). Usually, the main confusion in crop multisource satellite imagery. The architecture uses both unsu-
classification map for Ukraine territory is confusion between pervised and supervised NNs for segmentation and subsequent
maize and soybeans. Using the ensemble of 2-D CNNs, we classification of satellite imagery, respectively. In this letter,
were able to discriminate these classes more reliably: maize we used Landsat-8 and Sentinel-1A images over the JECAM
(PA = 94.6%, UA = 93.6%) and soybeans (PA = 86.9%, test site in Ukraine. Ensemble of 1-D and 2-D CNNs outper-
UA = 89.1%). formed the RF classifier and an ensemble of MLPs allowing
All these experiments were executed on a computer with us to better discriminate summer crops, in particular maize
Intel Core i7-4770 processor and RAM 32 Gb. Training and soybeans. In general, the use of CNN allowed us to reach
of ensemble of MLPs took up to 10 min at the same the target accuracy of 85% for major crops (wheat, maize,
time ensemble of 1-D CNNs trained approximately 4 h and sunflower, soybeans, and sugar beet) thus making a foundation
2-D CNNs training takes about 12 h. for further operational use of RS data for the whole territory
of Ukraine within the Sentinel-2 for Agriculture project. The
V. C ONCLUSION main advantage of using CNNs over MLP and RF is that
In this letter, we proposed a multilevel DL approach for it enables to build a hierarchy of local and sparse features
land cover and crop types classification using multitemporal derived from spectral and temporal profiles while MLP and
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KUSSUL et al.: DL CLASSIFICATION OF LAND COVER AND CROP TYPES 5

RF build a global transformation of features. The 2-D CNNs [20] N. Kussul, A. Shelestov, R. Basarab, S. Skakun, O. Kussul, and
outperformed the 1-D CNNs, but some small objects in the M. Lavrenyuk, “Geospatial intelligence and data fusion techniques
for sustainable development problems,” in Proc. ICTERI, 2015,
final classification map provided by 2-D CNNs were smoothed pp. 196–203.
and misclassified. [21] J. Ding, B. Chen, H. Liu, and M. Huang, “Convolutional neural
network with data augmentation for SAR target recognition,”
R EFERENCES IEEE Geosci. Remote Sens. Lett., vol. 13, no. 3, pp. 364–368,
Mar. 2016.
[1] M. Drusch et al., “Sentinel-2: ESA’s optical high-resolution mission [22] F. J. Huang and Y. LeCun, “Large-scale learning with SVM and
for GMES operational services,” Remote Sens. Environ., vol. 120, convolutional for generic object categorization,” in Proc. IEEE
pp. 25–36, May 2012. Comput. Soc. Conf. Comput. Vis. Pattern Recognit., Jun. 2006,
[2] R. Torres et al., “GMES Sentinel-1 mission,” Remote Sens. Environ., pp. 284–291.
vol. 120, pp. 9–24, May 2012. [23] T. Ishii, R. Nakamura, H. Nakada, Y. Mochizuki, and H. Ishikawa,
[3] D. P. Roy et al., “Landsat-8: Science and product vision for terrestrial “Surface object recognition with CNN and SVM in Landsat 8 images,”
global change research,” Remote Sens. Environ., vol. 145, pp. 154–172, in Proc. 14th IAPR Int. Conf. Mach. Vis. Appl. (MVA), May 2015,
Apr. 2014. pp. 341–344.
[4] J. Zhang, “Multi-source remote sensing data fusion: Status and trends,” [24] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
Int. J. Image Data Fusion, vol. 1, no. 1, pp. 5–24, Nov. 2010. pp. 436–444, May 2015.
[5] M. D. Mura, S. Prasad, F. Pacifici, P. Gamba, J. Chanussot, and [25] F. Zhang, B. Du, and L. Zhang, “Saliency-guided unsupervised feature
J. A. Benediktsson, “Challenges and opportunities of multimodality learning for scene classification,” IEEE Trans. Geosci. Remote Sens.,
and data fusion in remote sensing,” Proc. IEEE, vol. 103, no. 9, vol. 53, no. 4, pp. 2175–2184, Apr. 2015.
pp. 1585–1601, Sep. 2015. [26] F. Zhang, B. Du, and L. Zhang, “Scene classification via a gradi-
[6] A. Kolotii et al., “Comparison of biophysical and satellite pre- ent boosting random convolutional network framework,” IEEE Trans.
dictors for wheat yield forecasting in Ukraine,” Int. Arch. Pho- Geosci. Remote Sens., vol. 54, no. 3, pp. 1793–1802, Mar. 2016.
togramm., Remote Sens. Spatial Inf. Sci., vol. 40, no. 7, p. 35, 2015, [27] V. Mnih and G. E. Hinton, “Learning to detect roads in high-
doi: 10.5194/isprsarchives-XL-7-W3-39-2015. resolution aerial images,” in Proc. Eur. Conf. Comput. Vis., 2010,
[7] F. Kogan et al., “Winter wheat yield forecasting: A comparative analysis pp. 210–223.
of results of regression and biophysical models,” J. Autom. Inf. Sci., [28] J. Geng, J. Fan, H. Wang, X. Ma, B. Li, and F. Chen, “High-resolution
vol. 45, no. 6, pp. 68–81, 2013. SAR image classification via deep convolutional autoencoders,” IEEE
[8] F. Kogan et al., “Winter wheat yield forecasting in Ukraine based on Trans. Geosci. Remote Sens., vol. 12, no. 11, pp. 2351–2355,
Earth observation, meteorological data and biophysical models,” Int. Nov. 2015.
J. Appl. Earth Observat. Geoinf., vol. 23, pp. 192–203, Aug. 2013. [29] Y. Chen, X. Zhao, and X. Jia, “Spectral–spatial classification of hyper-
[9] H. McNairn, A. Kross, D. Lapen, R. Caves, and J. Shang, “Early season spectral data based on deep belief network,” IEEE J. Sel. Topics Appl.
monitoring of corn and soybeans with TerraSAR-X and RADARSAT-2,” Earth Observ. Remote Sens., vol. 8, no. 6, pp. 2381–2392, Jun. 2015.
Int. J. Appl. Earth Observat. Geoinf., vol. 28, pp. 252–259, May 2014. [30] H. Liang and Q. Li, “Hyperspectral imagery classification using sparse
[10] S. Skakun, N. Kussul, A. Y. Shelestov, M. Lavreniuk, and O. Kussul, representations of convolutional neural network features,” Remote Sens.,
“Efficiency assessment of multitemporal C-band Radarsat-2 intensity vol. 8, no. 2, p. 99, 2016.
and Landsat-8 surface reflectance satellite imagery for crop classification [31] H. Lyu, H. Lu, and L. Mou, “Learning a transferable change rule from
in Ukraine,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., a recurrent neural network for land cover change detection,” Remote
vol. 9, no. 8, pp. 3712–3719, Aug. 2016. Sens., vol. 8, no. 6, p. 506, 2016.
[11] R. Khatami, G. Mountrakis, and S. V. Stehman, “A meta-analysis of [32] E. Vermote, C. Justice, M. Claverie, and B. Franch, “Preliminary
remote sensing research on supervised pixel-based land-cover image analysis of the performance of the Landsat 8/OLI land surface
classification processes: General guidelines for practitioners and future reflectance product,” Remote Sens. Environ., vol. 185, pp. 46–56, 2016,
research,” Remote Sens. Environ., vol. 177, pp. 89–100, May 2016. doi: 10.1016/j.rse.2016.04.008.
[12] P. O. Gislason, J. A. Benediktsson, and J. R. Sveinsson, “Random Forests [33] Z. Zhu, S. Wang, and C. E. Woodcock, “Improvement and expansion
for land cover classification,” Pattern Recognit. Lett., vol. 27, no. 4, of the Fmask algorithm: Cloud, cloud shadow, and snow detection
pp. 294–300, 2006. for Landsats 4–7, 8, and Sentinel 2 images,” Remote Sens. Environ.,
[13] M. Han, X. Zhu, and W. Yao, “Remote sensing image classification vol. 159, pp. 269–277, Mar. 2015.
based on neural network ensemble algorithm,” Neurocomputing, vol. 78, [34] F. Waldner et al., “Towards a set of agrosystem-specific cropland map-
no. 1, pp. 133–138, 2012. ping methods to address the global cropland diversity,” Int. J. Remote
[14] X. Huang and L. Zhang, “An SVM ensemble approach combining Sens., vol. 37, no. 14, pp. 3196–3231, 2016.
spectral, structural, and semantic features for the classification of high- [35] S. V. Skakun and R. M. Basarab, “Reconstruction of missing data in
resolution remotely sensed imagery,” IEEE Trans. Geosci. Remote Sens., time-series of optical satellite images using self-organizing Kohonen
vol. 51, no. 1, pp. 257–272, Jan. 2013. maps,” J. Autom. Inform. Sci., vol. 46, no. 12, pp. 19–26, 2014.
[15] M. S. Lavreniuk et al., “Large-scale classification of land cover [36] N. Kussul, S. Skakun, A. Shelestov, M. Lavreniuk, B. Yailymov, and
using retrospective satellite data,” Cybern. Syst. Anal., vol. 52, no. 1, O. Kussul, “Regional scale crop mapping using multi-temporal satellite
pp. 127–138, 2016. imagery,” Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., vol. 40,
[16] N. Kussul, G. Lemoine, F. J. Gallego, S. V. Skakun, M. Lavreniuk, and no. 7, pp. 45–52, 2015.
A. Y. Shelestov, “Parcel-based crop classification in Ukraine using [37] W. Hu, Y. Huang, L. Wei, F. Zhang, and H. Li, “Deep convolutional
Landsat-8 data and Sentinel-1A data,” IEEE J. Sel. Topics Appl. Earth neural networks for hyperspectral image classification,” J. Sens., to be
Observ. Remote Sens., vol. 9, no. 6, pp. 2500–2508, Jan. 2016. published. [Online]. Available: http://dx.doi.org/10.1155/2015/258619
[17] Y. Chen, Z. Lin, X. Zhao, G. Wang, and Y. Gu, “Deep learning-based [38] D. P. Kingma and J. Ba. (Dec. 2014). “Adam: A method for stochastic
classification of hyperspectral data,” IEEE J. Sel. Topics Appl. Earth optimization.” [Online]. Available: https://arxiv.org/abs/1412.6980
Observ. Remote Sens., vol. 7, no. 6, pp. 2094–2107, Jun. 2014. [39] M. Abadi et al. (Mar. 2016). “TensorFlow: Large-scale machine
[18] W. Zhao and S. Du, “Learning multiscale and deep representations for learning on heterogeneous distributed systems.” [Online]. Available:
classifying remotely sensed imagery,” ISPRS J. Photogramm. Remote https://arxiv.org/abs/1603.04467
Sens., vol. 113, pp. 155–165, Mar. 2016. [40] F. J. Gallego, N. Kussul, S. Skakun, O. Kravchenko, A. Shelestov, and
[19] N. Kussul, N. Lavreniuk, A. Shelestov, B. Yailymov, and I. Butko, “Land O. Kussul, “Efficiency assessment of using satellite data for crop area
cover changes analysis based on deep machine learning technique,” estimation in Ukraine,” Int. J. Appl. Earth Observat. Geoinf., vol. 29,
J. Autom. Inf. Sci., vol. 48, no. 5, pp. 42–54, 2016. pp. 22–30, Jun. 2014.

View publication stats