Classiﬁcation of Hyperspectral Remote Sensing Images for Crop Type

Identiﬁcation: State of the Art

Kawtar El Karﬁ

, Sanaa El Fkihi

, Loubna El Mansouri

and Othmane Naggar

National school For Computer Science (ENSIAS), Rabat, Morocco

The Agronomic and Veterinary Institute Hassan II (IAV), Rabat, Morocco

MAScIR - Moroccan Foundation for Advanced Science, Innovation and Research, Rabat, Morocco

Keywords:

Remote sensing data, Hyperspectral image HSI, Classiﬁcation, Crop type mapping, Machine learning, Deep

learning

Abstract:

Hyperspectral imagery (HSI) is widely considered to be one of the most used technologies in different re-

mote sensing applications, such as crop mapping, which provides an essential baseline for understanding and

monitoring the Earth. Hyperspectral remote sensing, with its multiple narrow and continuous wavebands,

allow signiﬁcant improvements in the understanding of physiological processes of crops and the changes in

their phenology, which are indistinct in multi-spectral remote sensing. A generous number of features can

be derived from the hyperspectral data, although the classiﬁcation of crops using high-dimensional and high-

resolution data is a challenging task. The main objective of this paper is to list various techniques of machine

learning mostly applied for hyperspectral data classiﬁcation, besides the different hyperspectral open datasets

mainly used in various researches.

1 INTRODUCTION

Remote sensing imagery classiﬁcation gained huge

interest (Dou et al., 2018; Lu et al., 2011; Xia et al.,

2019) during the past ﬁve years,Considerable efforts

from different researches have been made to present a

variety of approaches for crop type identiﬁcation us-

ing remote sensing images. The identiﬁcation of cul-

tures as a research ﬁeld has been widely investigated

in several studies (Sitokonstantinou et al., 2018; Lira

Melo de Oliveira Santos et al., 2019), most of these

studies rely mainly on remotely sensed imaging, as it

is an efﬁcient and robust tool for collecting the infor-

mation needed to produce maps of crops. The gener-

ated maps are required in the process of making sus-

tainable decisions to assure the proper management

of agricultural areas, reduce costs, and create agrarian

policies(Orynbaikyzy et al., 2019).The current avail-

able sensors (e.g., multi/hyperspectral, synthetic aper-

ture radar, etc.)(Pandey et al., 2019) are increasingly

yielding different types of aerial or satellite images

with different resolutions (spatial resolution, spectral

resolution, and temporal resolution). For this purpose,

signiﬁcant efforts have been made to develop vari-

ous datasets (Quinn et al., 2018; El Mansouri et al.,

2019). Despite this, accurate crop mapping still a

challenging assignment, due to the small size of the

plots, the large variety of crops species, and the mix-

ture of different crops and uncultivated areas in some

cases(Siachalou et al., 2017). A wide range of stud-

ies highly recommends the adoption of HSI as it can

provide high-resolution and high-dimensional data to

produce high-scale maps (El Mansouri et al., 2018;

Khan et al., 2018). HSI is a merging technique of dig-

ital imaging and spectroscopy. The concept is as fol-

lows: Hyperspectral sensors acquire reﬂected energy

for a wide range of spectral bands, from visible to re-

ﬂected infrared range (Signoroni et al., 2019). They

allow the object in the image to be identiﬁed with high

precision (Sahoo et al., 2015), using various types of

variables, such as spectral signatures, vegetation in-

dices, and textural information. Due to the rich spec-

tral information, HSI can classify objects according

to their spectral features (Reshma and Veni, 2017).

Despite its ability and high classiﬁcation accuracy,

HSI faces one major challenge that impacts the qual-

ity of the hyperspectral data classiﬁcation, which is

handling the enormous amount of features. The main

objective of this paper is to highlight current classiﬁ-

cation approaches applied in the ﬁeld of remote sens-

ing. Mainly as the literature is very dense with differ-

ent approaches and methods, each with strengths and

weaknesses depending on the case studied, so we aim

to develop synthesis and see the impact of HSI with

deep learning as an axis of research. The rest of this

paper is organized as follows. Section 2 gives a brief

overview of HSI. Section 3 examines approaches in

the literature that are widely used for crop identiﬁca-

tion. Our conclusion and summary are drawn in the

ﬁnal section.

2 AN OVERVIEW OF

HYPERSPECTRAL IMAGES

2.1 Hyperspectral Imaging

The spectral imaging ﬁeld can be divided into three

domains: multispectral imaging (MSI), hyperspectral

imaging (HSI), and ultra-spectral imaging (USI). MSI

is a system where the used image has few separated

wavelengths. In HSI, the image is obtained with an

abundance of continuous wavelengths. USI is a sys-

tem that uses only one image with a low spatial res-

olution of several pixels(Khan et al., 2018). Hyper-

spectral images are hypercube (a three-dimensional

shape) containing light intensity measurements where

the two ﬁrst dimensions (X and Y) represent spatial

positions, and the third dimension represents spectral

variation. The images can be interpreted, typically, as

stacks of hundreds of two-dimensional spatial images

at different wavelengths, or tens of thousands of spec-

tra, aligned in rows and columns. Hyperspectral nar-

row bands, typically, contain 100–1500 wavebands

and collect data in the near-continuous spectrum from

several regions of the electromagnetic spectrum (ul-

traviolet, visible, near-, mid-, and far-infrared), which

offers many opportunities to study speciﬁc vegetation

variables (Elmasry et al., 2012).

2.2 Public Available Hyperspectral

Datasets

The use of different remotely sensed data is accord-

ing to user requirements and the need for high spa-

tial, spectral, temporal resolution, or a combination

of one or more of these and the area of coverage.in

this section; we illustrate some datasets that are free

and available for use (Table I).

• Indian Pines dataset:Indian Pines data set was

gathered by the AVIRIS sensor, over agricultural

areas in northwestern Indiana, with 145 pixel x

145 pixel images and 224 spectral bands... Six-

teen classes (Table I) are labeled (e.g., corn, grass,

soybean, woods, and so on).

• Pavia dataset:Pavia Centre and University are two

scenes acquired by the ROSIS sensor over the city

of Pavia, northern Italy. It is divided into two

parts: Pavia University (103 bands, 610 pixels x

340 pixels) and Pavia Center (102 bands, 1,096

pixels x 715 pixels). Nine labeled classes . They

contain various urban materials (such as bricks,

asphalt, and metals), water, and vegetation. This

dataset has been gained a popularity for a long

time and mostly used because it is one of the

largest sets of labeled HIS data.

• Botswana dataset: Botswana dataset is collected

by The NASA EO-1 satellite over the Okavango

Delta, Botswana in 2001-2004. The Hyperion

sensor on EO-1 obtain data at 30 m pixel resolu-

tion in 242 bands. Preprocessing of the data was

led by the UT Center for Space Research to ﬁx the

anomalies.

• Salinas dataset: This dataset was collected by the

AVIRIS sensor with 224-band over Salinas Val-

ley, California, and is characterized by high spa-

tial resolution (3.7-meter pixels). It includes veg-

etables, bare soils, and vineyard ﬁelds.

• KSC dataset:Kennedy Space Center (KSC) was

collected by the NASA AVIRIS (Airborne Vis-

ible/Infrared Imaging Spectrometer) over the

Kennedy Space Center (KSC) on March 23, 1996,

in Florida, the KSC data was gathered with 224

bands, and spatial resolution of 18m.This dataset

represents 13 classes of the various land cover

types.

3 A CONCEPTUAL OVERVIEW

OF MACHINE LEARNING

CLASSIFIERS

Remote sensing classiﬁcation results dependent on

many factors such as suitable classiﬁer, selection of

training data, image preprocessing, feature extraction,

accuracy assessment, the user’s need, size of the study

area, and analyst’s skills. Our contribution is to stress

the impact of machine learning methods used to clas-

sify HSI. Many studies have been carried out to in-

vestigate the performance of classiﬁers for different

remotely sensed data sets and their results in terms of

accuracy.

Table 1: public available Hyperspectral Remote Sensing datasets

Dataset Classes Labels Bands Pixels Mode

Indian Pines 16 10249 224 21025 Aerial

Pavia 9 50232 103 991040 Aerial

Botswana 14 3248 145 377856 Satellite

Salinas 16 54129 227 111104 Aerial

KSC 13 5211 176 314368 Aerial

3.1 Supervised Algorithms

SVM classiﬁcation: SVMs were initially designed to

identify a linear class boundary (i.e., a hyperplane).

The SVM classiﬁer is binary, used to identify a sin-

gle boundary between two classes. However, this

problem is handled by applying the classiﬁer to each

possible cluster combination, which means that pro-

cessing time is expected to increase exponentially as

the number of clusters rises (Usha and Vasuki, 2018;

Yang et al., 2019). Decision Tree classiﬁcation: DTs

are one of the simplest classiﬁers. A DT is a recur-

sive division of input data (Salehi et al., 2017). In a

classiﬁcation tree, leaf values constitute classes, leaf

values represent a continuous variable. One of DTs

advantages,is that the logic of the model can be eas-

ily visualized at the end of the classiﬁcation process.

DTs can use categorical data, and once the model is

completed, classiﬁcation is extremely fast because no

other complex mathematics is required (Salehi et al.,

2017; Maxwell et al., 2018). Random Forest classiﬁ-

cation: RF is an ensemble classiﬁer because it uses a

large number of DTs to overcome the weaknesses of a

single DT (Jeon and Kim, 2018). The majority ”vote”

of all trees is used to assign a ﬁnal class to each un-

known. This directly overcomes the problem that a

single tree is not optimal, but by incorporating sev-

eral trees, an overall optimum should be obtained. A

particular advantage of RF is that due to the presence

of several trees, it is not necessary to prune individ-

ual trees. A disadvantage is that by having several

trees, the possibility of viewing the trees is effectively

reduced (Jeon and Kim, 2018). k-NN classiﬁcation:

The k-NN classiﬁer is different from other classiﬁers.

It is an instance-based classiﬁer based on compar-

ing the similarity between each unknown sample with

the original training data(Tan et al., 2019). The un-

known sample is attributed to one of the k predeter-

mine classes of the training samples containing sim-

ilar features to the unknown sample. A low number

of classes (k) will, therefore, provide a very compli-

cated decision boundary, while a higher k values will

Table 2: Classiﬁcation results using Indian pines dataset

(Maxwell et al., 2018).

Algorithm Overall Kappa

accuracy(%) accuracy(%)

SVM 89,1 0,844

DT 78,3 0,687

RF 87,1 0,814

ANN 85,1 0,787

K-NN 78,6 0,686

lead to greater generalization. Since no trained model

is produced, the k-NN classiﬁcation is expected to re-

quire more resources as the number of training sam-

ples expands. ANNs classiﬁcation: ANNs are gener-

ally conceptualized as a mathematical analog of the

axons of an animal brain and their many intercon-

nections through synapses (Chlingaryan et al., 2018).

The elements of an ANN are layered neurons (equiva-

lent to biological axons). An ANN has minimal input

and output layers, with one neuron for each input vari-

able and one neuron for each output class. In addition,

ANNs usually have hidden nodes arranged in one or

more additional layers. One of the main challenges of

applying ANN classiﬁers is the training process that

can be time and resources consuming and can pro-

duce non-optimal or overﬁtted models. In this sec-

tion, we have summarized the algorithms of machine

learning that are often used for the classiﬁcation of

remote sensing images. (Table II) gives an overview

of the results obtained using the Indian pines dataset,

and it can be seen that the SVM has performed well.

3.2 Unsupervised Algorithms

In recent years, unsupervised feature learning algo-

rithms has become an interesting alternative to the tra-

ditional methods used for feature extraction and has

made signiﬁcant progress in the classiﬁcation of re-

mote sensing images (Ragettli et al., 2018). By learn-

ing features from images rather than relying on man-

ually engineered features, we can obtain more dis-

criminating features and better ﬁtted to the problem

at hand.

The principal component analysis (PCA) (Kang

et al., 2019) and k-means clustering (Ratnakumar and

Nanda, 2019) are considered to be two of the most

popular unsupervised feature learning methods.

PCA could be considered as the ﬁrst unsupervised

feature extraction algorithm that attempts to ﬁnd an

optimal representative projection matrix (need refer-

ence), and it is widely used to reduce the size of

satellite images (multispec-tral/hyperspectral) (Zhang

et al., 2019). Some extensions of PCA have also been

introduced in the literature, such as PCANet (Zhang

et al., 2019) and sparse PCA.

K-Means Clustering: The k-means clustering is a

method of vector quantization that aims to divide a

collection of data items into k clusters (Ratnakumar

and Nanda, 2019).

3.3 Selection of a Machine-learning

Classiﬁer

Selecting the appropriate classiﬁer for Multispectral

or hyperspectral data is a challenging assignment be-

cause of the large variety of machine-learning tech-

niques. In addition, the existing literature seems to

be unstructured and inconsistent concerning the de-

gree of effectiveness of existing algorithms. For in-

stance, the work presented in (Kang et al., 2019) has

concluded that SVM and Random forest (RF) algo-

rithms have similar performance in terms of accuracy

using the RapidEye satellite imagery data set, while

the authors of (Chlingaryan et al., 2018) found that

(SVM) outstripped (RF) and (KNN) using the same

data set. This contradiction in classiﬁcation results

could be explained by the different procedures used

in both studies, according to (Zhao and Du, 2016).

In (Elmasry et al., 2012), authors compared the per-

formance of a wide range of machine-learning clas-

siﬁcation algorithms, using standardized procedures

and 30 different datasets from Landsat, Ikonos, and

Probe-1 satellites on different dates. They found that

(RF) had the highest average classiﬁcation accuracy

of 73.19% , which was signiﬁcantly better than that

obtained by SVM (62.28%). Even though RF was

the most efﬁcient classiﬁer for 18 of the 30 datasets,

it was not always the most accurate classiﬁer. In the

same context, the authors of (Usha and Vasuki, 2018)

compared (SVM) and (RF) with 1-D Convolutional

Neural Network (CNN) architecture using hyperspec-

tral imagery of the San Francisco Bay Area, Cali-

fornia, for the year 2015. The models were trained

to classify data under three different seasons of the

year (spring, summer, fall). All analyses were com-

pleted using simulated hyperspectral infrared imaging

(HyspIRI) for the aim of land cover mapping. The ob-

tained results showed that the overall classiﬁcation ac-

curacy of the CNN architecture reached 89.9%, which

is similar to that obtained by SVM 89.5%. The re-

sults also showed that the SVM exceeded the RF by

an overall accuracy of more than 7%. In the previous

case studies, the authors showed the most appropri-

ate classiﬁcation algorithms that were used with sev-

eral datasets and yielded different results depending

on user requirements.

3.4 Other Useful Applications of

Machine Learning Methods in

Remote-sensing

Machine learning methods are not limited to clas-

siﬁcation processes. Many algorithms are also em-

ployed for regression. For example, Tree canopy

density data were generated using a regression tree

method, based on the DT algorithm . The SVM

can also be used for regression, known as support

vector regression (SVR). For instance, (Wang et al.,

2011) used this method for predicting water qual-

ity, chemical variables from the SPOT-5 data set and

stated that SVR achieved a result better than multiple

linear regression. For the prediction of biophysical

parameters,(Camps-Valls et al., 2006) found that the

implementation of an SVR outperformed regression

using ANNs.(Mountrakis et al., 2011) also noted the

value of SVR for predicting chlorophyll content, leaf

area, and vegetation cover from hyperspectral data.

Machine learning has also been used for probabilistic

predictions in remote sensing. For example,(Maxwell

et al., 2016) used RF to forecast the topographic prob-

ability of wetlands based on terrain features (Moun-

trakis et al., 2011).

3.5 Deep Learning Methods

Most of the current state-of-the-art approaches gener-

ally rely on supervised learning to obtain good feature

representations. The past decade has witnessed an im-

portant growth in using and developing deep learning

algorithms. It has become a trend in big data analysis

and many computer vision tasks, e.g., image classi-

ﬁcation, object detection, and natural language pro-

cessing. It gave birth to a new perspective in the re-

mote sensing ﬁeld when it has been introduced as a

promising method to classify HSI data that has been

used in various studies and provided accurate results.

Therefore, we can anticipate that the topic will

be further explored, and more and more research

works will be published in the next several years.

Since then, several attempts have been made to re-

place hand-engineered features with trainable multi-

layer networks, and some deep learning models have

shown impressive feature representation capability

for a wide range of applications, including remote

sensing images classiﬁcation (Paoletti et al., 2019)

(Kussul et al., 2017) (Elnagar et al., 2020).

In comparison with traditional features extraction

methods that require a considerable amount of en-

gineering skill and domain expertise, Deep learn-

ing features are automatically learned from data us-

ing a general-purpose learning procedure via deep-

architecture neural networks, which represents the

key advantage of deep learning methods. On the other

hand, compared with aforementioned unsupervised

feature learning methods that are generally weakly

structured models (e.g., sparse coding), deep learning

models that are composed of multiple processing lay-

ers can learn more powerful feature representations of

data with multiple levels of abstraction (Kang et al.,

2019). In addition, deep feature learning methods can

automatically extract features from complex hyper-

spectral data and can effectively deal with the problem

of the large variability of spectral signature. However,

the types of features extracted from deep networks are

various, e.g., spectral, spatial, and spectral-spatial fea-

tures, which makes deep learning more suitable for

the varieties of situations.

In the existing literature, a number of deep learn-

ing models have been proposed, such as Recurrent

Neural Network (RNN), Stacked Autoencoder (SAE),

and Convolutional Neural Networks (CNNs), (Pao-

letti et al., 2019) (Kussul et al., 2017) (Liu et al.,

2019). Several authors have attempted to give a gen-

eral review of current advances deep learning tech-

niques for hyperspectral images. For instance, (Liu

et al., 2019) listed the most popular deep learning-

based algorithms used in the hyperspectral data clas-

siﬁcation and underlined their ability to deal with

a restricted number of training samples and high-

dimensional data. Among those approaches, we can

cite the widely used CNN-based classiﬁers. Here we

mainly focus on CNN since it is the most used al-

gorithm in HSI classiﬁcation. CNNs are designed to

process data that come in the form of multiple ar-

rays, for example, a multispectral image composed

of multiple 2-D arrays containing pixel intensities in

the multiple band channels. The learning process of

CNNs is computationally efﬁcient and insensitive to

data changes. In remote sensing studies, 2D CNNs

(Figure 1) have been widely used to extract spatial

features from the dimensions of width and height for

object detection and semantic segmentation of high-

resolution images (He and Chen, 2019). the Hyper-

spectral image classiﬁcation is another application of

CNNs, in which CNN’s were used to extract spatial-

spectral features, through either 1D-convolution with

spectral features, 2D-convolution with spatial fea-

tures, or 3D-convolution with a combination of spec-

tral and spatial features (Bhosle and Musande, 2019).

(Zhu et al., 2019) found that 2D-convolution pro-

vides an accurate result in crop classiﬁcation than

1D-convolution. (Goodfellow et al., 2016) combined

hyperspectral images from three seasons and applied

1D-convolution for land cover classiﬁcation. In these

studies, convolutional layers in CNNs used mostly to

extract spatial or spectral features.

4 SUMMARY

The classiﬁcation of remotely sensed data has made

signiﬁcant progress over the past years. Many fac-

tors, such as the spatial resolution of the remotely

sensed data, different data sources, and the classiﬁ-

cation system must be considered when choosing a

classiﬁcation method to use. Various classiﬁcation

methods have their own merits. It is not easy to an-

swer the question of which classiﬁcation method is

suitable for a speciﬁc study. For a particular study, it

is often difﬁcult to identify the most suitable classiﬁer

due to the lack of guidelines in the literature for the se-

lection of appropriate classiﬁcation algorithms. Also,

the combination of different classiﬁcation approaches

has proven to be useful in improving the accuracy of

classiﬁcation (Maxwell et al., 2018). Our objective

was to propose a short review of the academic litera-

ture to provide some practical considerations for the

implementation of machine learning classiﬁcation in

remote sensing data. We suggest that the following

points should be highlighted .

• SVM, RF, and enhanced DTs have proven to be

very effective methods for classifying remotely

sensed data and, in general, these methods seem

to produce high overall accuracies compared to

other machine classiﬁers such as simple DTs and

k-NN. However, the best algorithm for a speciﬁc

task may be case-speciﬁc and may depend on the

classes mapped, the nature of the training data,

and the predictive variables provided.

• the quality of training data, generally, have a sig-

niﬁcant impact on the accuracy of classiﬁcation.

The training data may even have more impact than

the algorithm used. Therefore, it is better to obtain

a large number of high-quality training samples

that fully characterize the class signatures. How-

ever, there are practical limitations to collect large

training samples. If the training sample is small

Figure 1: The general 2D CNN architechture for classiﬁcation of hyperspectral data (Makantasis et al., 2015). Conv: convo-

lution; FC: fully connected.

in number, or if the data quality is ambiguous, an

algorithm that is robust to these problems must be

used, such as the DT methods.

• The accuracy of the classiﬁcation may be affected

by the imbalance in training data. In general,

overall accuracy may not decrease signiﬁcantly

due to imbalance.

• The three input dimensions (spatial, spectral, and

temporal) have an impact on the accuracy of crop

mapping, individually or in combination. An in-

crease in one of the dimensions increases the ac-

curacy of mapping. Each dimension plays an im-

portant role and contributes signiﬁcantly to the

output. From the spatial dimension to the spec-

tral dimension, from the spectral dimension to the

temporal dimension.

• Deep learning techniques, including deep neural

networks, have already revealed a great promise in

the ﬁeld of remote sensing, and the implementa-

tions have been made available in software pack-

ages such as MATLAB, R, and scikit-learn.

5 CONCLUSIONS

Earth Observation (EO) sensors are a source of infor-

mative data covering the whole globe in spatial and

spectral resolution for better and easier land cover

classiﬁcation. They are becoming increasingly attrac-

tive as an effective alternative to traditional or conven-

tional methods, and the last years have witnessed a re-

markable increase in the use of these technologies in

the ﬁeld of crop mapping or identiﬁcation. Due to the

variable nature of the landscape and multiple sensors,

classiﬁcation techniques also play an essential role in

the accuracy of crop mapping for multispectral to hy-

perspectral data. Indeed, in recent years, many clas-

siﬁcation machine learning algorithms (supervised or

unsupervised) have been used in crop mapping. To

implement appropriate methods for HSI classiﬁca-

tion, several other factors need to be considered to

overcome the problems of pixel size and various char-

acteristics to outperform one technique over another.

Therefore, it is necessary to know the input data di-

mensions, types of remotely sensed data, and appro-

priate classiﬁers implemented.

ACKNOWLEDGEMENTS

This work is part of the Multispectral satellite

imagery, data mining and agricultural applications

project, funded by the academy Hassan II of Science

and Technology.

REFERENCES

Bhosle, K. and Musande, V. (2019). Evaluation of deep

learning cnn model for land use land cover classiﬁ-

cation and crop identiﬁcation using hyperspectral re-

mote sensing images. Journal of the Indian Society of

Remote Sensing, 47(11):1949–1958.

Camps-Valls, G., Bruzzone, L., Rojo-Alvarez, J. L., and

Melgani, F. (2006). Robust support vector regres-

sion for biophysical variable estimation from remotely

sensed images. IEEE Geoscience and remote sensing

letters, 3(3):339–343.

Chlingaryan, A., Sukkarieh, S., and Whelan, B. (2018). Ma-

chine learning approaches for crop yield prediction

and nitrogen status estimation in precision agriculture:

A review. Computers and electronics in agriculture,

151:61–69.

Dou, P., Chen, Y., and Yue, H. (2018). Remote-sensing

imagery classiﬁcation using multiple classiﬁcation

algorithm-based adaboost. International Journal of

Remote Sensing, 39(3):619–639.

El Mansouri, L., Hadria, R., Lahmer, I., Moutaib, O., Ou-

jemaa, A., and ElGorch, A. (2018). Technologies

géo-spatiales pour renforcer les dispositifs de gestion

des terres agricoles: Appui à la gestion des surfaces

agrumicoles par télédétection dans la plaine de triffa-

berkane (maroc). African Journal on Land Policy and

Geospatial Sciences, 1(3):164–177.

El Mansouri, L., Lahssini, S., Hadria, R., Eddaif, N., Ben-

abdelouahab, T., and Dakir, A. (2019). Time series

multispectral images processing for crops and forest

mapping: two moroccan cases. In Geospatial Tech-

nologies for Effective Land Governance, pages 83–

106. IGI Global.

Elmasry, G., Kamruzzaman, M., Sun, D.-W., and Allen, P.

(2012). Principles and applications of hyperspectral

imaging in quality evaluation of agro-food products: a

review. Critical reviews in food science and nutrition,

52(11):999–1023.

Elnagar, A., Al-Debsi, R., and Einea, O. (2020). Arabic text

classiﬁcation using deep learning models. Information

Processing & Management, 57(1):102121.

Goodfellow, I., Bengio, Y., Courville, A., and Bengio, Y.

(2016). Deep learning, volume 1. MIT press Cam-

bridge.

He, X. and Chen, Y. (2019). Optimized input for cnn-based

hyperspectral image classiﬁcation using spatial trans-

former network. IEEE Geoscience and Remote Sens-

ing Letters, 16(12):1884–1888.

Jeon, W. and Kim, Y. (2018). An assessment of a ran-

dom forest classiﬁer for a crop classiﬁcation using air-

borne hyperspectral imagery. 대원사지,

34(1):141–150.

Kang, X., Zhuo, B., and Duan, P. (2019). Semi-supervised

deep learning for hyperspectral image classiﬁcation.

Remote Sensing Letters, 10(4):353–362.

Khan, M. J., Khan, H. S., Yousaf, A., Khurshid, K., and Ab-

bas, A. (2018). Modern trends in hyperspectral image

analysis: a review. IEEE Access, 6:14118–14129.

Kussul, N., Lavreniuk, M., Skakun, S., and Shelestov, A.

(2017). Deep learning classiﬁcation of land cover and

crop types using remote sensing data. IEEE Geo-

science and Remote Sensing Letters, 14(5):778–782.

Lira Melo de Oliveira Santos, C., Augusto Camargo Lam-

parelli, R., Kelly Dantas Araújo Figueiredo, G.,

Dupuy, S., Boury, J., Luciano, A. C. d. S., Torres,

R. d. S., and Le Maire, G. (2019). Classiﬁcation of

crops, pastures, and tree plantations along the season

with multi-sensor image time series in a subtropical

agricultural region. Remote Sensing, 11(3):334.

Liu, Y., Zhou, S., Han, W., Liu, W., Qiu, Z., and Li, C.

(2019). Convolutional neural network for hyperspec-

tral data analysis and effective wavelengths selection.

Analytica Chimica Acta, 1086:46–54.

Lu, D., Weng, Q., Moran, E., Li, G., and Hetrick, S.

(2011). Remote sensing image classiﬁcation. CRC

Press/Taylor and Francis: Boca Raton, FL, USA.

Makantasis, K., Karantzalos, K., Doulamis, A., and

Doulamis, N. (2015). Deep supervised learning

for hyperspectral data classiﬁcation through convo-

lutional neural networks. In 2015 IEEE Interna-

tional Geoscience and Remote Sensing Symposium

(IGARSS), pages 4959–4962. IEEE.

Maxwell, A. E., Warner, T. A., and Fang, F. (2018). Im-

plementation of machine-learning classiﬁcation in re-

mote sensing: An applied review. International Jour-

nal of Remote Sensing, 39(9):2784–2817.

Maxwell, A. E., Warner, T. A., and Strager, M. P. (2016).

Predicting palustrine wetland probability using ran-

dom forest machine learning and digital elevation

data-derived terrain variables. Photogrammetric En-

gineering & Remote Sensing, 82(6):437–447.

Mountrakis, G., Im, J., and Ogole, C. (2011). Support

vector machines in remote sensing: A review. IS-

PRS Journal of Photogrammetry and Remote Sensing,

66(3):247–259.

Orynbaikyzy, A., Gessner, U., and Conrad, C. (2019). Crop

type classiﬁcation using a combination of optical and

radar remote sensing data: a review. international

journal of remote sensing, 40(17):6553–6595.

Pandey, P. C., Koutsias, N., Petropoulos, G. P., Srivastava,

P. K., and Ben Dor, E. (2019). Land use/land cover in

view of earth observation: data sources, input dimen-

sions, and classiﬁers—a review of the state of the art.

Geocarto International, pages 1–32.

Paoletti, M., Haut, J., Plaza, J., and Plaza, A. (2019). Deep

learning classiﬁers for hyperspectral imaging: A re-

view. ISPRS Journal of Photogrammetry and Remote

Sensing, 158:279–317.

Quinn, J. A., Nyhan, M. M., Navarro, C., Coluccia, D.,

Bromley, L., and Luengo-Oroz, M. (2018). Human-

itarian applications of machine learning with remote-

sensing data: review and case study in refugee set-

tlement mapping. Philosophical Transactions of the

Royal Society A: Mathematical, Physical and Engi-

neering Sciences, 376(2128):20170363.

Ragettli, S., Herberz, T., and Siegfried, T. (2018). An un-

supervised classiﬁcation algorithm for multi-temporal

irrigated area mapping in central asia. Remote Sens-

ing, 10(11):1823.

Ratnakumar, R. and Nanda, S. J. (2019). A low complexity

hardware architecture of k-means algorithm for real-

time satellite image segmentation. Multimedia Tools

and Applications, 78(9):11949–11981.

Reshma, S. and Veni, S. (2017). Comparative analysis

of classiﬁcation techniques for crop classiﬁcation us-

ing airborne hyperspectral data. In 2017 Interna-

tional Conference on Wireless Communications, Sig-

nal Processing and Networking (WiSPNET), pages

2272–2276. IEEE.

Sahoo, R. N., Ray, S., and Manjunath, K. (2015). Hyper-

spectral remote sensing of agriculture. Current Sci-

ence, pages 848–859.

Salehi, B., Daneshfar, B., and Davidson, A. M. (2017). Ac-

curate crop-type classiﬁcation using multi-temporal

optical and multi-polarization sar data in an object-

based image analysis framework. International Jour-

nal of Remote Sensing, 38(14):4130–4155.

Siachalou, S., Mallinis, G., and Tsakiri-Strati, M. (2017).

Analysis of time-series spectral index data to enhance

crop identiﬁcation over a mediterranean rural land-

scape. IEEE Geoscience and Remote Sensing Letters,

14(9):1508–1512.

Signoroni, A., Savardi, M., Baronio, A., and Benini, S.

(2019). Deep learning meets hyperspectral image

analysis: a multidisciplinary review. Journal of Imag-

ing, 5(5):52.

Sitokonstantinou, V., Papoutsis, I., Kontoes, C., Lafarga Ar-

nal, A., Armesto Andrés, A. P., and Garraza Zurbano,

J. A. (2018). Scalable parcel-based crop identiﬁca-

tion scheme using sentinel-2 data time-series for the

monitoring of the common agricultural policy. Remote

Sensing, 10(6):911.

Tan, K., Zhang, Y., Wang, X., and Chen, Y. (2019). Object-

based change detection using multiple classiﬁers and

multi-scale uncertainty analysis. Remote Sensing,

11(3):359.

Usha, S. G. A. and Vasuki, S. (2018). Improved segmen-

tation and change detection of multi-spectral satel-

lite imagery using graph cut based clustering and

multiclass svm. Multimedia Tools and Applications,

77(12):15353–15383.

Wang, X., Fu, L., and He, C. (2011). Applying support vec-

tor regression to water quality modelling by remote

sensing data. International journal of remote sensing,

32(23):8615–8627.

Xia, W., Ma, C., Liu, J., Liu, S., Chen, F., Yang, Z., and

Duan, J. (2019). High-resolution remote sensing im-

agery classiﬁcation of imbalanced data using multi-

stage sampling method and deep neural networks. Re-

mote Sensing, 11(21):2523.

Yang, L., Mansaray, L. R., Huang, J., and Wang, L. (2019).

Optimal segmentation scale parameter, feature sub-

set and classiﬁcation algorithm for geographic object-

based crop recognition using multisource satellite im-

agery. Remote Sensing, 11(5):514.

Zhang, L., Su, H., and Shen, J. (2019). Hyperspectral di-

mensionality reduction based on multiscale superpix-

elwise kernel principal component analysis. Remote

Sensing, 11(10):1219.

Zhao, W. and Du, S. (2016). Spectral–spatial feature ex-

traction for hyperspectral image classiﬁcation: A di-

mension reduction and deep learning approach. IEEE

Transactions on Geoscience and Remote Sensing,

54(8):4544–4554.

Zhu, K., Chen, Y., Ghamisi, P., Jia, X., and Benediktsson, J.

(2019). Deep Convolutional Capsule Network for Hy-

perspectral Image Spectral and Spectral-Spatial Clas-

siﬁcation. Remote Sensing, 11:223.