Computers and Electronics in Agriculture 151 (2018) 61–69

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

Review

Machine learning approaches for crop yield prediction and nitrogen status T
estimation in precision agriculture: A review
⁎
Anna Chlingaryana, , Salah Sukkarieha, Brett Whelanb
a
Australian Centre for Field Robotics, Dept. of Aerospace, Mechanical & Mechatronic Engineering, The University of Sydney, NSW 2006, Australia
b
Centre for Carbon, Water and Food, School of Life and Environmental Sciences, The University of Sydney, NSW 2006, Australia

A R T I C LE I N FO A B S T R A C T

Keywords: Accurate yield estimation and optimised nitrogen management is essential in agriculture. Remote sensing (RS)
Vegetation indices systems are being more widely used in building decision support tools for contemporary farming systems to
Features extraction improve yield production and nitrogen management while reducing operating costs and environmental impact.
Predictive modelling However, RS based approaches require processing of enormous amounts of remotely sensed data from different
Decision making
platforms and, therefore, greater attention is currently being devoted to machine learning (ML) methods. This is
Information fusion
due to the capability of machine learning based systems to process a large number of inputs and handle non-
linear tasks. This paper discusses research developments conducted within the last 15 years on machine learning
based techniques for accurate crop yield prediction and nitrogen status estimation. The paper concludes that the
rapid advances in sensing technologies and ML techniques will provide cost-effective and comprehensive so-
lutions for better crop and environment state estimation and decision making. More targeted application of the
sensor platforms and ML techniques, the fusion of different sensor modalities and expert knowledge, and the
development of hybrid systems combining different ML and signal processing techniques are all likely to be part
of precision agriculture (PA) in the near future.

1. Introduction (Gontia and Tiwari, 2008), maize (Taghvaeian et al., 2012), trees
(Bellvert et al., 2016) and vineyards (Gutiérrez et al., 2018). At the
Improving crop yield production and quality while reducing oper- same time many ground-based platforms have been developed and
ating costs and environmental pollution is a key goal in precision aimed at different PA tasks such as mapping of soil properties (Barnes
agriculture (PA). The potential growth and yield depends on many et al., 2003), estimating evapotranspiration and drought stress (Maes
different production attributes such as the weather, soil properties, and Steppe, 2012), weed mapping (Sui et al., 2008) and assessing crop
topography, irrigation and fertilizer management. The need for timely water and nitrogen status (El-Shikha et al., 2007; Govender et al.,
and accurate sensing of these inputs for large agricultural fields has led 2009).
to increased adoption of remote and proximal sensing technologies Remote sensing at visible and near-infrared wavelengths (vis-NIR)
(Campbell and Wynne, 2011) in PA (Curran, 1987). These sensing has been used to devise many spectral indices for estimating different
techniques provide acquisition of spectral, spatial and temporal in- vegetation properties. This includes the amount of chlorophylls and
formation about the objects via ground-based vehicles, aircraft, sa- other photosynthetic/photoprotective pigments and the leaf area index
tellites and handheld radiometers. (LAI) (Barati et al., 2011; Gao, 1996; Haboudane et al., 2004; Huete,
Remote sensing, such as satellite and airborne multi-spectral scan- 1988; Qi et al., 1994; Sims and Gamon, 2002; Viña et al., 2011; Zarco-
ning, photography and video, enables precision weed management Tejada et al., 2012). More than 100 vegetation indices along with their
through the generation of timely and accurate weed maps (Lamb and applicability, representativeness, environment and implementation
Brown, 2001). Thermal remote sensing via airborne thermal imagery precision have recently been reviewed by Xue and Su (2017). They
has the potential to identify spatial variations in crop water status concluded that for real-world applications the use of any existing ve-
(Tilling et al., 2006), which can enable improvements in the water getation indices requires careful consideration of the strengths and
management in irrigated cropping systems. Nowadays, this remote shortcomings of those indices and the specific environment where they
sensing technique is widely used in different crop species such as wheat will be applied. Making crop yield predictions using remotely sensed

⁎
Corresponding author.
E-mail address: [email protected] (A. Chlingaryan).

https://doi.org/10.1016/j.compag.2018.05.012
Received 24 November 2017; Received in revised form 7 May 2018; Accepted 8 May 2018
0168-1699/ Published by Elsevier B.V.
A. Chlingaryan et al. Computers and Electronics in Agriculture 151 (2018) 61–69

vegetation indices has been attempted by Panda et al. (2010), Jaafar Dalponte et al., 2012; Pohl and Van Genderen, 1998) and providing
and Ahmad (2015) and many others. confidence intervals for the predictions. ML enables better decision
Utilization of wireless sensors and actuators in PA, as well as algo- making and informed actions in real-world scenarios without (or with
rithms for wireless sensor network data integration is now advancing minimal) human intervention. ML provides a powerful and flexible
(Zhou et al., 2012). Aqeel-ur-Rehman et al. (2014) presented a review framework for not only data-driven decision making but also for in-
on wireless sensor network technology and their applications in dif- corporation of expert knowledge into the system. These are some of the
ferent aspects of agriculture and reported on existing system frame- key characteristics of the ML techniques that make them widely used in
works in PA. many domains, and highly applicable to PA.
Nitrogen (N) is considered by growers as a major mineral nutrient The major aim of PA in cropping systems is to provide information
for plant growth and development because it is directly related to the that will enable better decisions to be made on management across
photosynthetic process (Andrews et al., 2013). At the same time N has a space and time (Whelan and Taylor, 2013). Specifically, information on
high environmental and economic impact. Hence, the optimization of N variation in plant health and physiology, nutrient status or pest/disease
fertilization for different crops has become a subject of many spectro- burden may allow different treatments or treatment intensity to be
metric studies (Cao et al., 2017; Chen et al., 2008; Goron et al., 2017; applied to specific areas of crop. The practical identification and seg-
Lukina et al., 2001; MacKerron et al., 1993; Maresma et al., 2016; regation of areas is commonly achieved by dividing a large field area
Quemada et al., 2014; Raun et al., 2005; Schepers and Raun, 2008). into smaller management zones with identified requirements for dif-
The estimation of the plant N status can be divided into two main ferent treatments. Conventionally, such delineation is based on maps of
types: destructive and non-destructive. The most common method of the crop field variability derived from soil and yield measurements.
destructive measurement is a chemical analysis which is associated Nawar et al. (2017) provided a comprehensive review on management
with the Kjeldahl technique and is laborious, lengthy and costly (Jones zone delineation approaches for PA applications. They illustrated how
and Moseley, 1993; Vigneau et al., 2011). Optical remote sensing of the recent developments in sensing technologies, geostatistical analysis,
plant N status is a non-destructive method based on canopy reflectance data fusion and interpolation techniques have improved precision and
in the visible–NIR wavelengths (400–900 nm). This measurement is reliability of management zone delineation, making it a viable strategy
completed in-situ, lowering the number of field samples required and in commercial agriculture. They also compared traditional with ad-
thus reducing the time and financial cost of field sample collection, vanced sensing technologies for delineating management zones.
preparation and laboratory analysis. Many studies have been dedicated Recently, Pantazi et al. (2015) demonstrated that the combination
to non-destructive measurement of the N status inferring in plants via of data fusion modelling and clustering methods was able to improve
remote sensing technology (Apostol et al., 2007; Lamb et al., 2002; the quality of management zone delineation. Specifically, they com-
Reyniers and Vrindts, 2006; Scharf et al., 2002; Tremblay et al., 2012) pared k-means clustering with the Self-Organizing Map (SOM) for de-
and spectral indices indicative of the plant N status have been derived lineating management zones maps for variable-rate N application.
from hyperspectral data (Chen et al., 2010; Tian et al., 2014; Yao et al., Furthermore, a hybrid SOM algorithm in combination with k-means
2010). was compared with k-means in terms of cluster separation and man-
Adoption of remote sensing in geology (Gupta, 2003), forestry agement zone formation based on data fusion of NDVI and soil para-
(Holmgren and Thuresson, 1998; Hultquist et al., 2014), hydrology meters.
(Engman and Gurney, 1991), agriculture (Seelan et al., 2003) and other As reviewed by Behmann et al. (2015), ML techniques have been
domains has led to the collection of significant volumes of data. The widely used for the early and accurate detection of biotic stress in crop,
volume is continuously growing and it is beyond human ability to specifically, for detection of weeds, plant diseases and insect pests.
personally integrate, analyse and make the best informed decisions Mehra et al. (2016) used ML techniques such as Artificial Neural Net-
from the information. This is particularly the case when the data is not works (ANNs), categorical and regression trees and Random Forests
homogeneous, i.e. is sensed by sensors with different spatial, temporal (RFs) to approach the problem of predicting the pre-planting risk of
and spectral modalities. Machine Learning (ML) is an emerging tech- Stagonospora nodorum blotch (SNB) in winter wheat. They developed
nology that can aid in the discovery of rules and patterns in large sets of risk assessment models that could be useful in making disease man-
data (Du and Jeffrey, 2007). agement decisions prior to planting of the wheat crop. Also Tellaeche
Crop yield prediction and N status estimation are considered to- et al. (2008) showed that cost savings and reduced pollution could be
gether here because of the direct linkage in fertiliser management de- achieved by a Bayesian framework based automatic decision making
cisions. Crop yield goals are routinely utilised for calculating N re- process for detecting weeds in corn crops.
quirements, both pre- and in-season. For devising potential site-specific Machine learning techniques applied to hyperspectral imaging data
management plans for N fertiliser, especially in-season, an estimation of can be used to reveal physiological and structural characteristics in
both would be ideal. The aim of this review is to show the capability of plants and enable tracking physiological dynamics due to environ-
different ML techniques to effectively handle these different but closely mental effects (Wahabzada et al., 2016). Goldstein et al. (2017) de-
related tasks. A review is presented of recent studies in the area of crop monstrated that field data, such as soil moisture, weather, irrigation
yield prediction and N status estimation, which incorporate different characteristics, and resulting yield could be fused via ML techniques to
ML techniques. It also covers comparative studies of different ML provide automated recommendations for irrigation. Gutiérrez et al.
techniques as they are applied to the same task in PA. Some technical (2018) have used thermal imaging and a combination of two ML
aspects of the ML techniques used in the reviewed studies are discussed. techniques (Rotation Forests and Decision Trees) to develop a new
methodology for the on-the-go assessment of vineyard water status with
2. Machine learning techniques potential for irrigation decision making.
Machine learning techniques can be used in field spectroscopy for
One of the main advantages of ML techniques is that they are cap- offline and online prediction of the soil parameters studied in the field
able of autonomously solving large non-linear problems using datasets (Morellos et al., 2016). They can work not only with variables such as
from multiple (potentially interconnected) sources. Some ML techni- derived spectral indices, but also with the entire spectral response trace
ques, such as Gaussian Processes (GPs) (Bishop, 2006; Rasmussen and (Wittenberghe et al., 2014). Spectral indices depend on a small number
Williams, 2005), Dirichlet Processes (DP) (Ferguson, 1973) and Indian of available spectral bands and therefore don’t use the entire informa-
Buffet Process (IBP) (Griffiths and Ghahramani, 2011) are probabilistic tion conveyed by the spectral trace. Thus, there is always the question:
and enable consideration of sensor noise while conducting probabilistic which vegetation index or suite of vegetation indices is better for the
fusion of information from different sensors (Castaldi et al., 2016; given task? (Panda et al., 2010). ML techniques, such as Neural

62
A. Chlingaryan et al. Computers and Electronics in Agriculture 151 (2018) 61–69

Networks, can be employed for automated extraction of relevant fea- Recently, Gonzalez-Sanchez et al. (2014) presented a comparative
tures from the data (You et al., 2017). Continuum-removal (CR) can be study of ANN, SVR, M5-Prime, kNN ML techniques and Multiple Linear
used to effectively isolate individual absorption features of interest and, Regression for crop yield prediction in ten crop datasets. To validate the
for example, estimate the chemical concentration in dry leaves (Clark models they used four accuracy metrics: Root Mean Square Error
and Roush, 1984; Huang et al., 2004; Kokaly, 2001; Vigneau et al., (RMS), Root Relative Square Error (RRSE), Normalized Mean Absolute
2011). After the automated extraction of features is performed, the ML Error (MAE) and Correlation Factor (R). Results showed that M5-Prime
techniques (e.g. BPNNs) can be used for automated crop yield predic- achieved the lowest errors across the produced crop yield models. The
tion and assessing which feature is more informative (Panda et al., results of that study ranked the techniques from the best to the worst,
2010). according to RMSE, RRSE, R, and MAE results, in the following order:
Due to the large number of bands in the hyperspectral data, some M5-Prime, kNN, SVR, ANN and MLR.
methods (e.g. Stepwise Multiple Linear Regression (SMLR)) are likely to In another study Nari and Yang-Won (2016) applied four ML tech-
suffer from multicollinearity when applied to canopy hyperspectral niques, SVM, Random Forest (RF), Extremely Randomized Trees (ERT)
data (Bolster et al., 1996; Grossman et al., 1996). Multicollinearity and Deep Learning (DL) to estimate corn yield in Iowa State. Compar-
means that one predicted variable can be linearly predicted from the isons of the validation statistics showed that DL provided more stable
others with a substantial degree of accuracy. A Partial Least Squares results by overcoming the overfitting problem.
Regression (PLSR) approach can be adopted to reduce the large number As soil and climatic conditions play a significant role in crop growth
of measured collinear spectral variables to a few non-correlated latent and yield, online proximal soil sensing for estimation of relevant soil
variables. It avoids the potential overfitting problems that are typically properties remains a missing component in the management system.
associated with SMLR (Grossman et al., 1996; Hansen and Schjoerring, Pantazi et al. (2016) predicted the within field variation in wheat yield
2003). using online multi-layer soil data and satellite imagery crop growth
Although Artificial Neural Networks (ANNs) are widely used in re- characteristics. They used supervised Self-Organizing Maps (SOM) in
mote sensing to predict vegetation parameters and crop yield (Farifteh this work. The data for a single cropping season was used and the
et al., 2007; Kaul et al., 2005; Kuwata and Shibasaki, 2015) and are performance of counter-propagation artificial neural networks (CP-
generally used to deal with non-linear models, their practical applica- ANN), XY-fused Networks (XY-F) and Supervised Kohonen Networks
tion has some difficulties, such as selection of the number and size of (SKN) for predicting wheat yield was compared. The average overall
hidden layers, learning rate, the need for a large training dataset and accuracy for SKN was 81.65%, for CP-ANN was 78.3% and for XY-F was
the problem of overfitting. Support Vector Machines (SVMs) are also 80.92%, showing that the SKN model had the best overall performance.
being used in PA and have the potential to resolve the problem of Spectral vegetation indices (VIs) are mathematical combinations
overfitting when analysing high-dimensional data such as hyperspectral (often ratios) of mainly red, green and infrared spectral bands. They are
imaging data. Applications of SVMs include crop type classification designed to find functional relationships between crop characteristics
(Peña et al., 2014; Ustuner et al., 2014; Yang et al., 2016) and crop- and remote sensing observations (Wiegand et al., 1979). Since the de-
weeds classification (Tellaeche et al., 2007). There are other ML tech- velopment of the Simple Ratio Index (SR) (Birth and McVey, 1968;
niques which may be used for different applications in PA. Advantages Jordan, 1969; Knipling, 1970) and the Normalized Difference Vegeta-
and disadvantages of some of them have been listed by Ali et al. (2015). tion Index (NDVI) (Johnson, 2014; Rouse et al., 1973; Tucker, 1979) a
Despite the significant recent developments in ML and the suc- large number of vegetation indices have been developed, such as the
cessful application in many areas, ML techniques have some funda- two-band Enhanced Vegetation Index (EVI2) (Bolton and Friedl, 2013)
mental limitations when used naively in a purely data-driven fashion. and Normalized Difference Water Index (NDWI) (Satir and Berberoglu,
The accuracy of the predictions and their uncertainties produced by the 2016) to name a few. The availability of a large number of indices leads
ML algorithms strongly depend on the data quality, model representa- to the need to optimally choose and combine indices for maximally
tiveness and the dependencies between the input and target variables in accurate crop yield estimation.
the collected datasets. Data with a high level of noise, erroneous data, Panda et al. (2010) implemented Back-propagation Neural Network
presence of outliers and biases in the data, and incomplete datasets may (BPNN) modelling to test the efficiency of the following four spectral
significantly reduce the predictive power of the models. The appro- vegetation indices: NDVI, green vegetation index (GVI), soil adjusted
priate definition of the ML model, such as the GP covariance function, vegetation index (SAVI) and perpendicular vegetation index (PVI) in
ANN architecture and SVR parametrisations is also critical for optimal corn crop yield prediction. The results showed that the corn yield was
performance. Many strategies, such as incorporation of expert knowl- best predicted using BPNN models that used the means and standard
edge into the covariance function, outlier detection, transfer learning, deviations of PVI grid images.
and model selection through automated cross-validation can be em- Although spectral vegetation indices are widely used, they depend
ployed to overcome these limitations. only on a small number (usually two) of the available image bands and
the full spectrum information in hyperspectral data is not exploited. In
3. Yield estimation their recent publication You et al. (2017) used Convolutional Neural
Network (CNN) and Long-Short Term Memory (LSTM) to automatically
Achieving maximum crop yield at minimum cost with a healthy discover relevant features from raw data. Deep Gaussian Process was
ecosystem is one of the main goals of agricultural production. Early employed to integrate the spatio-temporal information from the data.
detection and management of problems associated with crop yield re- They evaluated the proposed approach on the task of predicting county-
strictions can help increase yield and subsequent profit, and estimating level soybean production in the United States. The results of this study
yield is important to numerous crop management and business deci- showed that the proposed approach outperformed competing techni-
sions. ques by a large margin.
In recent years different ML techniques have been implemented to Yuan et al. (2015) addressed the problem of selection of the most
achieve accurate yield prediction for different crops (Subhadra et al., representative bands to reduce the dimensionality of the data while
2016). The most successful ML techniques have been Artificial Neural maximally preserving its original information. They proposed a group-
Networks (Drummond et al., 2003; Fortin et al., 2011; Liu et al., 2001; wise band selection framework, with a joint sparsity constraint, which
Safa et al., 2004), Support Vector Regression (Ruß, 2009), M5-Prime is fully unsupervised and computationally feasible. They conducted
Regression Trees (Frausto-Solis et al., 2009; Marinković et al., 2009; experiments on hyperspectral classification and colour visualisation.
Ruß and Kruse, 2010; Wang and Witten, 1997) and k-nearest neighbour The results of those experiments demonstrated that, in these two ap-
(Zhang et al., 2010). plications, the proposed framework is more robust and reliable

63
A. Chlingaryan et al. Computers and Electronics in Agriculture 151 (2018) 61–69

compared to the other state of the art alternatives (e.g. traditional validation (RV2 ) sets, the root mean square errors of prediction (RMSEP)
pointwise-selection-based methods). for the calibration and validation sets, the ratio of prediction to de-
To make a timely prediction of crop yield, the Spiking Neural viation (RPD), the computational efficiency (CE) and the complexity
Networks (SNN) model has been presented by Bose et al. (2016). It level (CL). The results of the comparison showed that the SVM method
introduces for the first time SNN as a promising technique for spatio- was more robust in coping with potential confounding factors for most
temporal data modelling, analysis, and land use/crop prediction. As varieties, ecological site and growth stage. However, the VI method
deep learning has the capability to extract key features from the data utilising the Soil-Adjusted Vegetation Index (1200 and 705 bands) was
for estimation, it can be expected to have less dependency on the input most accurate for the estimation of the LNC in wheat.
data. Because of this, even in areas where data acquisition is limited, Gaussian Processes (GPs) machine learning regression algorithms
deep learning can be expected to provide good quality estimation of have been applied to estimate chlorophyll content, N content, leaf
crop yield (Kuwata and Shibasaki, 2015). water content and specific leaf area from a field-based multi-species
Many other studies have been conducted on the application of ML dataset (for trees) (Wittenberghe et al., 2014). As an input the GP used
techniques to crop yield estimation from remotely sensed and in-situ the entire spectral data and then several distinct wavebands were au-
data. Table 1. presents a review of the studies and provides a summary, tomatically chosen for estimation of different leaf parameters. Results
methodology and discussion for each publication. This discussion is illustrate that the information to predict a leaf parameter is not re-
concentrated on some key technical aspects of the used ML techniques. stricted to one or few spectral bands, but that six or more separate
bands can be (equally) involved. Taking more bands into account helps
4. Precision nitrogen management better discriminate the source of change when some bands are also
influenced by other variables.
As nitrogen (N) plays a significant role in the process of photo- Three methods (PLS, ANN, LS-SVM) have been used to estimate the
synthesis, it is important for crop health and development. At the same N status non-destructively in rice using canopy spectral reflectance with
time, environmental factors and cost require a prudent application of N. visible and near-infrared reflectance spectroscopy (Shao et al., 2012).
Because of these factors the problem of optimal N management has The comparative analysis showed that the LS-SVM outperformed the
attracted the attention of numerous researchers (Cao et al., 2015, 2017; other methods and it was concluded that LS-SVM is a promising alter-
Dai et al., 2013; Magney et al., 2017; Zhao et al., 2013). native for the regression analysis to quantify N status in rice.
One of the approaches to optimal N management in PA is to use The Random Forest (RF) algorithm is a data mining method de-
management zones, that is, identify subfield regions with homogeneous veloped by Breiman (2001). It can be employed to reduce the re-
characteristics that require similar treatment. The most widely used dundancy in complex high-dimensional hyperspectral datasets. The
methods for delineation of site-specific management zones are the fuzzy results by Elfatih et al. (2013) show that RF regression applied to hy-
C-means and k-means algorithms (Schuster et al., 2011; Vrindts et al., perspectral data has the potential to accurately predict sugarcane leaf N
2005). These are popular clustering methods used extensively for un- concentration thus assisting in making informed decisions regarding
supervised learning and identification of structure in datasets. How- site-specific application of N fertilizers.
ever, determining subfield areas is a difficult task because of the com- There are many other studies dedicated to precision N management
plex correlations and spatial variability of soil properties and nutrient using not only ML but also other techniques such as kriging, multi-
concentrations, which are responsible for variations in crop yield within variate methods and inverse distance weighting. Table 2. presents those
the field. studies and provides a summary, methodology and discussion for each
The non-destructive methods for making N fertilizer recommenda- publication. This discussion is concentrated on some key technical as-
tions for crops are commonly based on plant N status testing, using pects of the used ML and other techniques.
remote sensing and in-situ data (Cilia et al., 2014; Maresma et al., 2016;
Tremblay et al., 2011). Cao et al. (2015) evaluated two sensor systems, 5. Conclusions
based on a three band user-configurable Crop Circle ACS-470 sensor
and a two fixed-band GreenSeeker sensor, in order to estimate winter Sensing technologies and ML techniques have rapidly advanced
wheat N status. The results of the comparison demonstrated that the during the last decade. These developments are likely to continue
winter wheat N status can be better predicted via the three band Crop providing cost-effective and more comprehensive datasets combined
Circle ACS-470 sensor. with more sophisticated algorithmic solutions enabling better crop and
A comprehensive review of the advantages and disadvantages of environment state estimation and decision making. We are at the be-
different methods for sensing the N status in plants has been conducted ginning of a promising path that has the potential to significantly alter
by Muñoz-Huerta et al. (2013). Diacono et al. (2013) produced a review crop yield management.
of precision N management of a wheat crop. They investigated ap- A number of ML techniques have already been successfully applied
proaches/results of site-specific N management of wheat to analyse to different PA tasks. This review particularly demonstrates that
both performance and sustainability of this agricultural practice.
However, to the best of the authors’ knowledge, there are no published (a) Back-propagation Neural Networks allow identification of the im-
review papers focused on ML techniques for precision N management. portance of different Vegetation Indices (VI) for more accurate crop
A number of spectrometric studies have been undertaken devoted to yield estimation.
the estimation of the N status in plants using CR, vegetation indices (b) The combination of Convolutional Neural Networks or Long-short
(VIs) (Li et al., 2010; Maresma et al., 2016; Ruß and Kruse, 2011), Term Memory with Gaussian Processes enables feature extraction
SMLR and PLSR (Grossman et al., 1996; Haboudane et al., 2008; from the data and efficient reduction of the error maps.
Hansen and Schjoerring, 2003; Huang et al., 2004; Tian et al., 2011; Yu (c) Gaussian Processes are useful for automatic selection of wavebands
et al., 2013; Zhu et al., 2006). from the entire spectrum in order to predict different characteristics
Yao et al. (2015) applied different linear (CR, VI, SMLR and PLSR) of plant leaves.
and nonlinear (ANN and SVM) regression methods in order to de- (d) M5-Prime Regression Trees are a suitable tool for multi-class crop
termine which method, input variable and model could estimate the prediction.
Leaf Nitrogen Concentration (LNC) in winter wheat with higher accu- (e) Least Squares Support Vector Machine is a promising tool for re-
racy, more robustness, less time and lower complexity. A comparative gression analysis to quantify Nitrogen status.
assessment of those six methods was conducted using the following six (f) Fuzzy cognitive Map (FCM) can be used to model and represent
metrics: coefficients of determination for the calibration (RC2 ) and expert knowledge for yield prediction and crop management.

64
 Table 1
Publications that use machine learning techniques for crop yield estimation with a focus on their technical aspects.
Paper Summary Methodology Discussion

Pantazi et al. (2016) This paper developed and evaluated a yield prediction model for wheat. For the yield prediction the Self-Organizing Map Models (SOMs): The presented approach incorporates the yield limiting factors in a multi-layer fusion model.
A. Chlingaryan et al.

fusion vectors have been used as input for the three ANNs. The fusion vectors consist of the values of the
eight soil parameters collected with the on-line soil sensor, the satellite imagery calculated NDVI values Counter-Propagation Artificial Neural Networks (CPANN)
and historic yield data from the previous two years. XY-Fused Networks (XY-Fs)
••
Supervised Kohonen Networks (SKNs)
•
Stas et al. (2016) The paper presented a comparison of two machine learning techniques (BRT and SVM) for prediction of Boosted Regression Trees (BRT) When a limited number of training samples is available, ML techniques used in this paper are better
winter wheat yield in Henan province of China. Three types of NDVI-related predictors have been used: Support Vector Machines (SVM) able to cope with large set of predictors (compared to MLR)
••
Single NDVI, Incremental NDVI and Targeted NDVI.
The results of comparison, which are based on a cross-validation error (RMSE), showed that BRT model
consistently outperforms SVM.

Heremans et al. (2015) In this paper two regression tree methods (BRT and RF) were used in order to evaluate the accuracy of Boosted Regression Trees (BRT) BRT is sensitive to noise, prone to overfitting and much slower than bagging. At the same time,
winter wheat yield, using NDVI data from the SPOT-VEGETATION sensor together with meteorological Random Forest (RF) boosting has been shown to be more accurate than bagging.
••
variables and fertilization levels, in the North China. The aim was not only to compare the performance RF can be used to improve the performance of bagging. In terms of accuracy, RFs are comparable to
of the methods but also to assess the potential for early-season predictions of winter wheat yield at the boosting but don’t have the mentioned limitations. RF has much lower computational cost than
prefecture level (five prefectures were involved). boosting
The comparison of methods was based on cross-validation R2 and RMSE. The results showed that BRT
outperforms RF for four out of the five prefectures.

Liang et al. (2015) The paper presented a non-destructive method - the hybrid inversion method, for estimation of leaf area Curve fitting In contrast to full-spectrum approaches, using VIs to estimate LAI requires a reduction in the number
index (LAI) values of crops. The method used different regression algorithms and allowed determining Artificial Neural Network (ANN) of model input parameters and therefore may result in lower inversion accuracy. However, RFR can
••
the relationships between optimal simulated VIs and simulated LAI values. To establish hybrid inversion Random Forest Regression (RFR) enable good performance with several or even a single parameter if that input parameter is highly
model ANN and RFR have been used.The comparison of the used algorithms showed that RFR was a
• correlated and representative.
better method for modelling with the higher R2 and lower RMSE values for different datasets and
various VIs.

Wu et al. (2015) This paper developed and compared two inversion models, using Statistical Regression model and BPNN Statistical Regression Model BPNN refers to a broad family of ANNs where the error is calculated at the output layer (using the
model, to estimate the LAI of a temperate meadow steppe in China. BPNN observations) and is propagated back through the layers of the ANN. The optimisation process adjusts
••
The results of comparison showed that BPNN method (accuracy: 82.2%) outperforms Statistical the weights in each layer by minimising the pre-defined loss function.
Regression model (accuracy: 78.8%).

Jin et al. (2016) In the paper the particle swarm optimization (PSO) algorithm was used to assimilate field spectroscopic Particle Swarm Optimization (PSO) algorithm Particle Swarm Optimization (PSO) algorithm can be used to minimise the difference between the

65
data into the AquaCrop model to improve the estimation accuracy of winter wheat yield under different regression based and AquaCrop model based estimates. This helps to improve the predictions.
planting dates and irrigation management strategies.
The results showed that the PSO algorithm is an effective method for improving the estimates of
biomass and yield of winter wheat.

Li et al. (2016) The paper aimed to produce accurate and timely predictions of grassland LAI for the meadow steppes of Partial Least Squares Regression (PLSR) RFs can provide better resistance to the over-fitting problem and to noise in the data compared with
northern China, using different regression approaches and hybrid geostatistical methods. ANNs other regression methods. However, RF method ignores spatial autocorrelation information.
••
The comparison of predictions via hybrid geostatistical methods, followed by different regression RFs RFRK is an extension of RF and is very similar to RK. It helps to include the spatial autocorrelation
models was presented. The results showed that the RF model provides the most accurate predictions Regression Kriging (RK) into the RF
••
among the regression models. Random Forests Residuals Kriging (RFRK)
•
Papageorgiou et al. (2013) The paper mainly aimed to present a method, based on FCM learning technique, to develop a Fuzzy Cognitive Mapping (FCM) The FCM method has many advantages such as simplicity, adaptability and capability of
computational intelligent tool for categorizing apple fruit yield.
• approximating abstractive structures. FCMs are knowledge-based and, therefore, provide meaningful
The results showed that the knowledge-based FCM learning approach predicts properly the results. As FCMs are mainly constructed manually, data-driven learning algorithms are required when
phenomenon, gives a front-end decision about the class of apple fruit yield, and provides similar results dealing with a large number of variables.
to those obtained from horticulturist experts.

Papageorgiou et al. (2011) The main aim of the paper was to connect yield defining parameters with yield in cotton crop Fuzzy Cognitive Mapping (FCM) Fuzzy cognitive Map (FCM) represents a combination of neural networks and fuzzy logic, and can be
production in Central Greece. The simulation approach based on the soft computing technique of Fuzzy ANNs used for information representation and decision making in complex processing environments. In
••
Cognitive Maps was investigated (FCM tool). Decision Trees (DTs) particular, FCMs can be used to model and represent expert knowledge for cotton yield prediction and
The data from six subsequent years were used to estimate the average classification accuracy of the yield Bayesian Networks (BNs) crop management
••
production, using the FCM tool. The results of estimation were compared with results of some ML
techniques obtained from the same data.
The results of comparison based on the overall accuracy of each method showed that the FCM technique
performed better in most of the cases.

Kaul et al. (2005) The paper described the development of ANN models as an accurate technique for corn and soybean Artificial Neural Network (ANN) ANN and MLR are among the techniques that can be used for agricultural modelling and prediction.
yield prediction in Maryland nutrient management planning. Multiple Linear Regression (MLR) The MLR is a simple methodology which is also easy to apply. ANN is a much more sophisticated
••
The results showed that ANN yield prediction is more accurate than the MLR-based yield model. technique the difficulties of the practical application of which are described in Section 2.
Computers and Electronics in Agriculture 151 (2018) 61–69
 Table 2
Publications that use machine learning and other techniques for precision nitrogen management.
Paper Summary Methodology Discussion

Song et al. (2017) The paper showed that based on the data collected from some georeferenced locations the Ordinary Ordinary Kriging Analysis (OKA) Ordinary Kriging Analysis is a widely used approach in PA. However, ML via Gaussian
A. Chlingaryan et al.

Kriging Analysis allows the interpolation of maps for Processes based modelling will provide
• automation
• wheat grain protein content (GPC) • quick update of the models as new data becomes available
• GPC yield • optimised data collection (based on maximising the information gain)
• wheat canopy fluorescence parameters, including Simple Florescence Ratio and Nitrogen Balance • probabilistic outcomes with representative uncertainties
Indices (NBI) • possibility for further fusion of data from different sensors
• soil Nitrate-Nitrogen (NO3-N) content and soil Time Domain Reflectometry (TDR)
The comparison of the fluorescence parameter maps, soil NO3-N and soil TDR maps with the wheat
GPC and the GPC yield maps demonstrated that the NBI spatial variability map in the late stage of
wheat growth can be used to distinguish areas that produce higher GPC

Magney et al. (2017) The paper aimed Seventeen commonly used spectral VIs Although VIs have been used for a number of different applications in satellite remote sensing
for PA, this paper reports that VIs from RapidEye imagery can be used for estimating wheat N
• to evaluate the usefulness of RapidEye spectral VIs to predict cumulative N uptake in wheat uptake.
• to examine the usefulness of remotely sensed N uptake maps for PA applications
It was concluded that

• the top performing VI was the Normalized Difference Red-Edge index (NDRE)
• N uptake maps from RapidEye imagery could have important implications for PA

Morellos et al. (2016) To predict total nitrogen (TN), organic carbon (OC) and moisture content (MC) in fresh (wet and Multivariate methods: The advantage of ML methods is that they are capable of tackling non-linear problems in the
unprocessed) soil samples two multivariate and two machine learning methods have been compared. dataset.
The results indicated that machine learning methods outperformed the multivariate methods for the • Principal Component Regression (PCR) The ML techniques can be used in field spectroscopy for off-line and online prediction of the
prediction of all three soil properties. • Partial Least Squares Regression (PLSR), soil parameters studied in the fields (if the soil type and variability is similar to the one studied
Machine learning methods: in this paper)

• Least Squares Support Vector Machines (LS-

66
SVM)
• Cubist

Castaldi et al. (2016) The paper proposed data fusion process in order to improve the choice of satellite bands for grain N Combination of LMEM can be a very efficient technique to estimate the spatial variability of the soil and crop
uptake prediction. variables accurately across the field with limited data, thus saving time and reducing the
The results showed that the best spectral regions vary over the growing season of the wheat crop. • Stepwise Regression with Backward costs.
Selection
• Stepwise Variance Inflation Factors (VIFs)
analysis
• Linear Mixed Effect Model (LMEM)

Wang et al. (2017) The paper investigated the modelling performances of four different chemometric techniques and Four techniques: PCR, PLSR, and BPNN use all available wavelengths simultaneously, while SMLR selects
three vegetation indices. Results showed that the best modelling and prediction accuracy were found useful wavelengths from the available spectrum and ignores the remaining wavebands.
in the model established by PLSR and spectra measured with a black background. A higher coefficient • Principal Components Regression (PCR) To improve the performance of the methods normalization can be used on the raw spectra
of determination between the leaf N concentration and fruit yield was found at 50 days after full • Partial Least Squares Regression (PLSR) collected by the probe, and wavelengths with very large atmospheric influence can be
bloom. • Stepwise Multiple Linear Regression (SMLR) removed.
• BPNN
Three indices:
• Difference Spectral Index
• Normalized Difference Spectral Index
• Ratio Spectral Index

Guo et al. (2015) The paper compared two different approaches (SLR and RFRK) to predict and map the spatial • Stepwise Linear Regression (SLR) RFRK model required no assumptions about the relationships between the target variable and
distribution of soil organic matter for the rubber plantation. • RFRK the predictor variables. Those relationships could be nonlinear and hierarchical. This can be
Results showed that RFRK outperforms SLR, by providing lower prediction errors (ME, MAE, and • Generalized Additive Mixed Model (GAMM) revealed by using GAMM and CART.
RMSE) and higher R2 • Classification And Regression Tree (CART)

Dai et al. (2014) The paper presented ANN-kriging methodology in order to predict accurate Soil Organic Matter (SOM) • ANN-kriging It is suggested that the proposed ANN-kriging methodology can be used to improve the
content maps. • ANN accuracy of SOM content mapping at large scale.
A comparison of proposed method with the other interpolation methods was performed to assess the • Inverse Distance Weighting (IDW)
prediction accuracy.
The results indicated that ANN-kriging provides the lower RMSE.
Computers and Electronics in Agriculture 151 (2018) 61–69
A. Chlingaryan et al. Computers and Electronics in Agriculture 151 (2018) 61–69

Based on the current dynamics in algorithmic developments and for remote sensing applications. J. Geophys. Res.: Solid Earth 89, 6329–6340.
sensor technologies the following future trends can be expected: Curran, P.J., 1987. Remote sensing in agriculture: an introductory review. J. Geogr. 86,
147–156.
Dai, X., Zhou, X., Jia, D., Xiao, L., Kong, H., He, M., 2013. Managing the seeding rate to
(a) More optimised, targeted application of the currently available improve nitrogen-use efficiency of winter wheat. Field Crops Res. 154, 100–109.
Dai, F., Zhou, Q., Lv, Z., Wang, X., Liu, G., 2014. Spatial prediction of soil organic matter
sensors and established ML techniques to specific PA tasks content integrating artificial neural network and ordinary kriging in Tibetan Plateau.
(b) Interconnected treatment of spatial, spectral and temporal domains Ecol. Indicat. 45, 184–194.
and incorporation of expert knowledge into the ML techniques Dalponte, M., Bruzzone, L., Gianelle, D., 2012. Tree species classification in the Southern
Alps based on the fusion of very high geometrical resolution multispectral/hyper-
aimed at modelling and decision making in different aspects of PA. spectral images and LiDAR data. Remote Sens. Environ. 123, 258–270.
(c) The combination of multiple ML as well as signal processing tech- Diacono, M., Rubino, P., Montemurro, F., 2013. Precision nitrogen management of wheat
A review. Agron. Sustain. Dev. 33, 219–241.
niques into hybrid systems to benefit from the strengths of those
Drummond, S.T., Sudduth, K.A., Joshi, A., Birrel, S.J., Kitchen, N.R., 2003. Statistical and
techniques and compensate for their individual shortcomings. neural methods for site-specific yield prediction. Trans. Am. Soc. Agric. Eng. 46,
(d) Spatial and spectral fusion of information from sensors with dif- 5–14.
Du, Z., Jeffrey, J.P.T., 2007. Advances in Machine Learning Applications in Software
ferent spatial resolution and spectral characteristics. Engineering. IGI Global, Hershey, PA, USA.
(e) The dynamic combination of stationary (e.g. in-ground probes, Elfatih, M.A., Fethi, B.A., Riyad, I., 2013. Random forest regression and spectral band
weather station) and mobile (e.g. ground and aerial vehicles, sa- selection for estimating sugarcane leaf nitrogen concentration using EO-1 Hyperion
hyperspectral data. Int. J. Remote Sens. 34, 712–728.
tellites) equipment to enable active optimal data collection, in- El-Shikha, D.M., Waller, P., Hunsaker, D., Clarke, T., Barnes, E., 2007. Ground-based
formation fusion and model update for high value areas. remote sensing for assessing water and nitrogen status of broccoli. Agric. Water
Manage. 92, 183–193.
Engman, E.T., Gurney, R.J., 1991. Remote sensing in hydrology. Chapman and Hall,
Acknowledgement London.
Farifteh, J., Van der Meer, F., Atzberger, C., Carranza, E.J.M., 2007. Quantitative analysis
of salt-affected soil reflectance spectra: A comparison of two adaptive methods (PLSR
This work has been supported by the Australian Centre for Field and ANN). Remote Sens. Environ. 110, 59–78.
Robotics. Ferguson, T.S., 1973. A Bayesian analysis of some nonparametric problems. Ann. Statist.
1, 209–230.
Fortin, J.G., Anctil, F., Parent, L., Bolinder, M.A., 2011. Site-specific early season potato
References yield forecast by neural network in Eastern Canada. Prec. Agric. 12, 905–923.
Frausto-Solis, J., Gonzalez-Sanchez, A., Larre, M., 2009. A New Method for Optimal
Cropping Pattern, In: Aguirre, A.H., Borja, R.M., Garciá, C.A.R. (Eds.), MICAI 2009:
Ali, I., Greifeneder, F., Stamenkovic, J., Neumann, M., Notarnicola, C., 2015. Review of
Advances in Artificial Intelligence: 8th Mexican International Conference on Artificial
Machine Learning Approaches for Biomass and Soil Moisture Retrievals from Remote
Intelligence, Guanajuato, México, November 9-13, 2009. Proceedings. Springer
Sensing Data. Remote Sens. 7, 15841.
Berlin Heidelberg, Berlin, Heidelberg, pp. 566–577.
Andrews, M., Raven, J.A., Lea, P.J., 2013. Do plants need nitrate? The mechanisms by
Gao, B., 1996. NDWI—A normalized difference water index for remote sensing of vege-
which nitrogen form affects plants. Ann. Appl. Biol. 163, 174–199.
tation liquid water from space. Remote Sens. Environ. 58, 257–266.
Apostol, S., Viau, A., Tremblay, N., 2007. A comparison of multiwavelength laser-induced
Goldstein, A., Fink, L., Meitin, A., Bohadana, S., Lutenberg, O., Ravid, G., 2017. Applying
fluorescence parameters for the remote sensing of nitrogen stress in field-cultivated
machine learning on sensor data for irrigation recommendations: revealing the
corn. Can. J. Remote Sens. 33, 150–161.
agronomist’s tacit knowledge. Prec. Agric.
Aqeel-ur-Rehman, Abbasi, A.Z., Islam, N., Shaikh, Z.A., 2014. A review of wireless sensors
Gontia, N.K., Tiwari, K.N., 2008. Development of crop water stress index of wheat crop
and networks' applications in agriculture. Comp. Stand. Interf. 36, 263–270.
for scheduling irrigation using infrared thermometry. Agric. Water Manage. 95,
Barati, S., Rayegani, B., Saati, M., Sharifi, A., Nasri, M., 2011. Comparison the accuracies
1144–1152.
of different spectral indices for estimation of vegetation cover fraction in sparse ve-
Gonzalez-Sanchez, A., Frausto-Solis, J., Ojeda-Bustamante, W., 2014. Predictive ability of
getated areas. Egypt. J. Remote Sens. Space Sci. 14, 49–56.
machine learning methods for massive crop yield prediction. Spanish J. Agric. Res.
Barnes, E.M., Sudduth, K.A., Hummel, J.W., Lesch, S.M., Corwin, D.L., Yang, C.,
12, 313–328.
Daughtry, C.S.T., Bausch, W.C., 2003. Remote- and ground-based sensor techniques
Goron, T., Nederend, J., Stewart, G., Deen, B., Raizada, M., 2017. Mid-season leaf glu-
to map soil properties. Photogramm. Eng. Remote Sens. 6, 619–630.
tamine predicts end-season maize grain yield and nitrogen content in response to
Behmann, J., Mahlein, A., Rumpf, T., Römer, C., Plümer, L., 2015. A review of advanced
nitrogen fertilization under field conditions. Agronomy 7, 41.
machine learning methods for the detection of biotic stress in precision crop pro-
Govender, M., Dye, P.J., Weiersbye, I.M., Witkowski, E.T.F., Ahmed, F., 2009. Review of
tection. Prec. Agric. 16, 239–260.
commonly used remote sensing and ground-based technologies to measure plant
Bellvert, J., Marsal, J., Girona, J., Gonzalez-Dugo, V., Fereres, E., Ustin, S., Zarco-Tejada,
water stress. Water SA 35.
P., 2016. Airborne thermal imagery to detect the seasonal evolution of crop water
Griffiths, T.L., Ghahramani, Z., 2011. The Indian buffet process: an introduction and
status in peach, nectarine and saturn peach orchards. Remote Sens. 8, 39.
review. J. Mach. Learn. Res. 12, 1185–1224.
Birth, G.S., McVey, G.R., 1968. Measuring the color of growing turf with a reflectance
Grossman, Y.L., Ustin, S.L., Jacquemoud, S., Sanderson, E.W., Schmuck, G., Verdebout, J.,
spectrophotometer. Agron. J. 60, 640–643.
1996. Critique of stepwise multiple linear regression for the extraction of leaf bio-
Bishop, C.M., 2006. Pattern Recognition and Machine Learning (Information Science and
chemistry information from leaf reflectance data. Remote Sens. Environ. 56,
Statistics). Springer-Verlag, New York Inc.
182–193.
Bolster, K.L., Martin, M.E., Aber, J.D., 1996. Determination of carbon fraction and ni-
Guo, P., Li, M., Luo, W., Tang, Q., Liu, Z., Lin, Z., 2015. Digital mapping of soil organic
trogen concentration in tree foliage by near infrared reflectances: a comparison of
matter for rubber plantation at regional scale: An application of random forest plus
statistical methods. Can. J. For. Res. 26, 590–600.
residuals kriging approach. Geoderma 237, 49–59.
Bolton, D.K., Friedl, M.A., 2013. Forecasting crop yield using remotely sensed vegetation
Gupta, R.P., 2003. Remote Sensing Geology. Springer-Verlag, Berlin Heidelberg,
indices and crop phenology metrics. Agric. For. Meteorol. 173, 74–84.
Germany.
Bose, P., Kasabov, N.K., Bruzzone, L., Hartono, R.N., 2016. Spiking neural networks for
Gutiérrez, S., Diago, M.P., Fernández-Novales, J., Tardaguila, J., 2018. Vineyard water
crop yield estimation based on spatiotemporal analysis of image time series. IEEE
status assessment using on-the-go thermal imaging and machine learning. PLOS One
Trans. Geosci. Remote Sens. 54, 6563–6573.
13, e0192037.
Breiman, L., 2001. Random forests. Mach. Learn. 45, 5–32.
Haboudane, D., Miller, J.R., Pattey, E., Zarco-Tejada, P.J., Strachan, I.B., 2004.
Campbell, J.B., Wynne, R.H., 2011. Introduction to Remote Sensing. fifth ed. Guildford
Hyperspectral vegetation indices and novel algorithms for predicting green LAI of
Press.
crop canopies: Modeling and validation in the context of precision agriculture.
Cao, Q., Miao, Y., Feng, G., Gao, X., Li, F., Liu, B., Yue, S., Cheng, S., Ustin, S.L., Khosla,
Remote Sens. Environ. 90, 337–352.
R., 2015. Active canopy sensing of winter wheat nitrogen status: An evaluation of two
Haboudane, D., Tremblay, N., Miller, J.R., Vigneault, P., 2008. Remote estimation of crop
sensor systems. Comput. Electr. Agric. 112, 54–67.
chlorophyll content using spectral indices derived from hyperspectral data. IEEE
Cao, Q., Miao, Y., Li, F., Gao, X., Liu, B., Lu, D., Chen, X., 2017. Developing a new crop
Trans. Geosci. Remote Sens. 46, 423–437.
circle active canopy sensor-based precision nitrogen management strategy for winter
Hansen, P.M., Schjoerring, J.K., 2003. Reflectance measurement of canopy biomass and
wheat in North China Plain. Prec. Agric. 18, 2–18.
nitrogen status in wheat crops using normalized difference vegetation indices and
Castaldi, F., Castrignanò, A., Casa, R., 2016. A data fusion and spatial data analysis ap-
partial least squares regression. Remote Sens. Environ. 86, 542–553.
proach for the estimation of wheat grain nitrogen uptake from satellite data. Int. J.
Heremans, S., Dong, Q., Zhang, B., Bydekerke, L., Van Orshoven, J., 2015. Potential of
Remote Sens. 37, 4317–4336.
ensemble tree methods for early-season prediction of winter wheat yield from short
Chen, P., Haboudane, D., Tremblay, N., Wang, J., Vigneault, P., Li, B., 2010. New spectral
time series of remotely sensed normalized difference vegetation index and in situ
indicator assessing the efficiency of crop nitrogen treatment in corn and wheat.
meteorological data. APPRES 9, 097095.
Remote Sens. Environ. 114, 1987–1997.
Holmgren, P., Thuresson, T., 1998. Satellite remote sensing for forestry planning—A re-
Chen, D., Suter, H., Islam, A., Edis, R., Freney, J.R., Walker, C.N., 2008. Prospects of
view. Scandinavian J. For. Res. 13, 90–110.
improving efficiency of fertiliser nitrogen in Australian agriculture: a review of en-
Huang, Z., Turner, B.J., Dury, S.J., Wallis, I.R., Foley, W.J., 2004. Estimating foliage
hanced efficiency fertilisers. Soil Res. 46, 289–301.
nitrogen concentration from HYMAP data using continuum removal analysis. Remote
Cilia, C., Panigada, C., Rossini, M., Meroni, M., Busetto, L., Amaducci, S., Boschetti, M.,
Sens. Environ. 93, 18–29.
Picchi, V., Colombo, R., 2014. Nitrogen status assessment for variable rate fertiliza-
Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25,
tion in maize through hyperspectral imagery. Remote Sens. 6, 6549.
295–309.
Clark, R.N., Roush, T.L., 1984. Reflectance spectroscopy: Quantitative analysis techniques

67
A. Chlingaryan et al. Computers and Electronics in Agriculture 151 (2018) 61–69

Hultquist, C., Chen, G., Zhao, K., 2014. A comparison of Gaussian process regression, (HAICTA 2015), Kavala - Greece.
random forests and support vector regression for burn severity assessment in diseased Pantazi, X.E., Moshou, D., Alexandridis, T., Whetton, R.L., Mouazen, A.M., 2016. Wheat
forests. Remote Sens. Lett. 5, 723–732. yield prediction using machine learning and advanced sensing techniques. Comput.
Jaafar, H.H., Ahmad, F.A., 2015. Crop yield prediction from remotely sensed vegetation Electr. Agric. 121, 57–65.
indices and primary productivity in arid and semi-arid lands. Int. J. Remote Sens. 36, Papageorgiou, E.I., Markinos, A.T., Gemtos, T.A., 2011. Fuzzy cognitive map based ap-
4570–4589. proach for predicting yield in cotton crop production as a basis for decision support
Jin, X., Kumar, L., Li, Z., Xu, X., Yang, G., Wang, J., 2016. Estimation of winter wheat system in precision agriculture application. Appl. Soft Comput. 11, 3643–3657.
biomass and yield by combining the AquaCrop model and field hyperspectral data. Papageorgiou, E.I., Aggelopoulou, K.D., Gemtos, T.A., Nanos, G.D., 2013. Yield prediction
Remote Sens. 8, 972. in apples using Fuzzy Cognitive Map learning approach. Comput. Electr. Agric. 91,
Johnson, D.M., 2014. An assessment of pre- and within-season remotely sensed variables 19–29.
for forecasting corn and soybean yields in the United States. Remote Sens. Environ. Peña, J., Gutiérrez, P., Hervás-Martínez, C., Six, J., Plant, R., López-Granados, F., 2014.
141, 116–128. Object-based image classification of summer crops with machine learning methods.
Jones, D.I.H., Moseley, G., 1993. Laboratory methods for estimating nutritive quality. In: Remote Sens. 6, 5019.
Davies, A., Baker, R.D., Grant, S.A., Laidlaw, A.S. (Eds.), Sward Measurement Pohl, C., Van Genderen, J.L., 1998. Review article Multisensor image fusion in remote
Handbook, second ed., pp. 265–283. sensing: Concepts, methods and applications. Int. J. Remote Sens. 19, 823–854.
Jordan, C.F., 1969. Derivation of leaf-area index from quality of light on the forest floor. Qi, J., Chehbouni, A., Huete, A.R., Kerr, Y.H., Sorooshian, S., 1994. A modified soil ad-
Ecology 50, 663–666. justed vegetation index. Remote Sens. Environ. 48, 119–126.
Kaul, M., Hill, R.L., Walthall, C., 2005. Artificial neural networks for corn and soybean Quemada, M., Gabriel, J., Zarco-Tejada, P., 2014. Airborne hyperspectral images and
yield prediction. Agric. Syst. 85, 1–18. ground-level optical sensors as assessment tools for maize nitrogen fertilization.
Knipling, E.B., 1970. Physical and physiological bases for the reference of visible and near Remote Sens. 6, 2940.
infrared radiation from vegetation. Remote Sens. Environ. 1, 155–159. Rasmussen, C.E., Williams, C.K.I., 2005. Gaussian Processes for Machine Learning
Kokaly, R.F., 2001. Investigating a physical basis for spectroscopic estimates of leaf ni- (Adaptive Computation and Machine Learning). The MIT Press.
trogen concentration. Remote Sens. Environ. 75, 153–161. Raun, W.R., Solie, J.B., Stone, M.L., Martin, K.L., Freeman, K.W., Mullen, R.W., Zhang, H.,
Kuwata, K., Shibasaki, R., 2015. Estimating crop yields with deep learning and remotely Schepers, J.S., Johnson, G.V., 2005. Optical sensor-based algorithm for crop nitrogen
sensed data, 2015 IEEE International Geoscience and Remote Sensing Symposium fertilization. Commun. Soil Sci. Plant Anal. 36, 2759–2781.
(IGARSS), pp. 858–861. Reyniers, M., Vrindts, E., 2006. Measuring wheat nitrogen status from space and ground-
Lamb, D.W., Brown, R.B., 2001. PA—Precision agriculture. J. Agric. Eng. Res. 78, based platform. Int. J. Remote Sens. 27, 549–567.
117–125. Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., 1973. Monitoring Vegetation Systems
Lamb, D.W., Steyn-Ross, M., Schaare, P., Hanna, M.M., Silvester, W., Steyn-Ross, A., in the Great plains with ERTS, NASA. Goddard Space Flight Center 3d ERTS-1
2002. Estimating leaf nitrogen concentration in ryegrass (Lolium spp.) pasture using Symposium. United States, Washington, DC, pp. 309–317.
the chlorophyll red-edge: Theoretical modelling and experimental observations. Int. Ruß, G., Kruse, R., 2010. Feature Selection for Wheat Yield Prediction. In: Bramer, M.,
J. Remote Sens. 23, 3619–3648. Ellis, R., Petridis, M. (Eds.), Research and Development in Intelligent Systems XXVI:
Li, F., Miao, Y., Hennig, S.D., Gnyp, M.L., Chen, X., Jia, L., Bareth, G., 2010. Evaluating Incorporating Applications and Innovations in Intelligent Systems XVII. Springer,
hyperspectral vegetation indices for estimating nitrogen concentration of winter London, London, pp. 465–478.
wheat at different growth stages. Prec. Agric. 11, 335–357. Ruß, G., Kruse, R., 2011. Machine learning methods for spatial clustering on precision
Li, Z., Wang, J., Tang, H., Huang, C., Yang, F., Chen, B., Wang, X., Xin, X., Ge, Y., 2016. agriculture data. In: Kofod-Petersen, A., Heintz, F., Langseth, H. (Eds.), Eleventh
Predicting grassland leaf area index in the meadow steppes of northern china: a Scandinavian Conference on Artificial Intelligence. IOS Press, pp. 40–49.
comparative study of regression approaches and hybrid geostatistical methods. Ruß, G., 2009. Data mining of agricultural yield data: a comparison of regression models.
Remote Sens. 8, 632. In: Perner, P. (Ed.), Advances in Data Mining. Applications and Theoretical Aspects:
Liang, L., Di, L., Zhang, L., Deng, M., Qin, Z., Zhao, S., Lin, H., 2015. Estimation of crop 9th Industrial Conference, ICDM 2009, Leipzig, Germany, July 20 - 22, 2009.
LAI using hyperspectral vegetation indices and a hybrid inversion method. Remote Proceedings. Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 24–37.
Sens. Environ. 165, 123–134. Safa, B., Khalili, A., Teshnehlab, M., Liaghat, A., 2004. Artificial neural networks appli-
Liu, J., Goering, C.E., Tian, L., 2001. A neural network for setting target corn yields. 44, cation to predict wheat yield using climatic data. In: Proc. 20th Int. Conf. on IIPS, pp.
pp. 705–713. 1–39.
Lukina, E.V., Freeman, K.W., Wynn, K.J., Thomason, W.E., Mullen, R.W., Stone, M.L., Satir, O., Berberoglu, S., 2016. Crop yield prediction under soil salinity using satellite
Solie, J.B., Klatt, A.R., Johnson, G.V., Elliott, R.L., Raun, W.R., 2001. Nitrogen fer- derived vegetation indices. Field Crops Res. 192, 134–143.
tilization optimization algorithm based on in-season estimates of yield and plant Scharf, P.C., Schmidt, J.P., Kitchen, N.R., Sudduth, K.A., Hong, S.Y., Lory, J.A., Davis,
nitrogen uptake. J. Plant Nutr. 24, 885–898. J.G., 2002. Remote sensing for nitrogen management. J. Soil Water Conserv. 57,
MacKerron, D.K.L., Young, M.W., Davies, H.V., 1993. A method to optimize N-application 518–524.
in relation to soil supply of N, and yield of potato. In: Fragoso, M.A.C., Van Schepers, J.S., Raun, W.R., 2008. Nitrogen in Agricultural Systems. American Society of
Beusichem, M.L., Houwers, A. (Eds.), Optimization of Plant Nutrition: Refereed pa- Agronomy, Crop Science Society of America, Soil Science Society of America,
pers from the Eighth International Colloquium for the Optimization of Plant Madison, WI.
Nutrition, 31 August – 8 September 1992, Lisbon, Portugal. Springer, Netherlands, Schuster, E.W., Kumar, S., Sarma, S.E., Willers, J.L., Milliken, G.A., 2011. Infrastructure
Dordrecht, pp. 635–640. for data-driven agriculture: identifying management zones for cotton using statistical
Maes, W.H., Steppe, K., 2012. Estimating evapotranspiration and drought stress with modeling and machine learning techniques. In: 2011 8th International Conference &
ground-based thermal remote sensing in agriculture: a review. J. Exp. Bot. 63, Expo on Emerging Technologies for a Smarter World, p. 1.
4671–4712. Seelan, S.K., Laguette, S., Casady, G.M., Seielstad, G.A., 2003. Remote sensing applica-
Magney, T.S., Eitel, J.U.H., Vierling, L.A., 2017. Mapping wheat nitrogen uptake from tions for precision agriculture: A learning community approach. Remote Sens.
RapidEye vegetation indices. Prec. Agric. 18, 429–451. Environ. 88, 157–169.
Maresma, Á., Ariza, M., Martínez, E., Lloveras, J., Martínez-Casasnovas, J., 2016. Analysis Shao, Y., Zhao, C., Bao, Y., He, Y., 2012. Quantification of nitrogen status in rice by least
of Vegetation Indices to Determine Nitrogen Application and Yield Prediction in squares support vector machines and reflectance spectroscopy. Food Bioprocess
Maize (Zea mays L.) from a Standard UAV Service. Remote Sensing 8, 973. Technol. 5, 100–107.
Marinković, B., Crnobarac, J., Brdar, S., Antić, B., Jaćimović, G., Crnojević, V., 2009. Data Sims, D.A., Gamon, J.A., 2002. Relationships between leaf pigment content and spectral
mining approach for predictive modeling of agricultural yield data. Proc. First reflectance across a wide range of species, leaf structures and developmental stages.
International Workshop on Sensing Technologies in Agriculture, Novi Sad, Serbia, Remote Sens. Environ. 81, 337–354.
pp. 1–5. Song, X., Yang, G., Yang, C., Wang, J., Cui, B., 2017. Spatial variability analysis of within-
Mehra, L.K., Cowger, C., Gross, K., Ojiambo, P.S., 2016. Predicting pre-planting risk of field winter wheat nitrogen and grain quality using canopy fluorescence sensor
stagonospora nodorum blotch in winter wheat using machine learning models. Front. measurements. Remote Sens. 9, 237.
Plant Sci. 7, 390. Stas, M., Orshoven, J.V., Dong, Q., Heremans, S., Zhang, B., 2016. A comparison of
Morellos, A., Pantazi, X.E., Moshou, D., Alexandridis, T., Whetton, R., Tziotzios, G., machine learning algorithms for regional wheat yield prediction using NDVI time
Wiebensohn, J., Bill, R., Mouazen, A.M., 2016. Machine learning based prediction of series of SPOT-VGT. In: 2016 Fifth International Conference on Agro-Geoinformatics
soil total nitrogen, organic carbon and moisture content by using VIS-NIR spectro- (Agro-Geoinformatics), pp. 1–5.
scopy. Biosyst. Eng. 152, 104–116. Subhadra, M., Debahuti, M., Gour Hari, S., 2016. Applications of Machine Learning
Muñoz-Huerta, R., Guevara-Gonzalez, R., Contreras-Medina, L., Torres-Pacheco, I., Prado- Techniques in Agricultural Crop Production: A Review Paper. Indian J. Sci.
Olivarez, J., Ocampo-Velazquez, R., 2013. A review of methods for sensing the ni- Technol. 9.
trogen status in plants: advantages disadvantages and recent advances. Sensors 13, Sui, R., Thomasson, J.A., Hanks, J., Wooten, J., 2008. Ground-based sensing system for
10823. weed mapping in cotton. Comput. Electr. Agric. 60, 31–38.
Nari, K., Yang-Won, L., 2016. Machine learning approaches to corn yield estimation using Taghvaeian, S., Chávez, J., Hansen, N., 2012. Infrared thermometry to estimate crop
satellite images and climate data: a case of iowa state. J. Korean Soc. Surv., Geodesy, water stress index and water use of irrigated maize in Northeastern Colorado. Remote
Photogramm. Cartogr. 34, 383–390. Sens. 4, 3619.
Nawar, S., Corstanje, R., Halcro, G., Mulla, D., Mouazen, A.M., 2017. Chapter Four - Tellaeche, A., Burgos-Artizzu, X.P., Pajares, G., Ribeiro, A., 2008. A vision-based method
Delineation of Soil Management Zones for Variable-Rate Fertilization: A Review. In: for weeds identification through the Bayesian decision theory. Pattern Recogn. 41,
Sparks, D.L. (Ed.), Advances in Agronomy. Academic Press, pp. 175–245. 521–530.
Panda, S.S., Ames, D.P., Panigrahi, S., 2010. Application of vegetation indices for agri- Tellaeche, A., BurgosArtizzu, X.P., Pajares, G., Ribeiro, A., 2007. A vision-based classifier
cultural crop yield prediction using neural network techniques. Remote Sens. 2, in precision agriculture combining bayes and support vector machines. In: IEEE
673–696. International Symposium on Intelligent Signal Processing, pp. 1–6.
Pantazi, X.E., Moshou, D., Mouazen, A., Alexandridis, T., Kuang, B., 2015. Data Fusion of Tian, Y.C., Yao, X., Yang, J., Cao, W.X., Hannaway, D.B., Zhu, Y., 2011. Assessing newly
Proximal Soil Sensing and Remote Crop Sensing for the Delineation of Management developed and published vegetation indices for estimating rice leaf nitrogen con-
Zones in Arable Crop Precision Farming. In: 7th International Conference on centration with ground- and space-based hyperspectral reflectance. Field Crops Res.
Information and Communication Technologies in Agriculture, Food and Environment 120, 299–310.

68
A. Chlingaryan et al. Computers and Electronics in Agriculture 151 (2018) 61–69

Tian, Y.C., Gu, K., Chu, X., Yao, X., Cao, W.X., Zhu, Y., 2014. Comparison of different Gaussian processes retrieval of leaf parameters from a multi-species reflectance, ab-
hyperspectral vegetation indices for canopy leaf nitrogen concentration estimation in sorbance and fluorescence dataset. J. Photochem. Photobiol. B: Biol. 134, 37–48.
rice. Plant Soil 376, 193–209. Wu, Q., Jin, Y., Bao, Y., Hai, Q., Yan, R., Chen, B., Zhang, H., Zhang, B., Li, Z., Li, X., Xin,
Tilling, A.K., O'Leary, G.J., Ferwerda, J.G., Jones, S.D., Fitzgerald, G.J., Rodriguez, D., X., 2015. Comparison of two inversion methods for leaf area index using HJ-1 sa-
Belford, R., 2006. Remote sensing of nitrogen and water stress in wheat. Field Crops tellite data in a temperate meadow steppe. Int. J. Remote Sens. 36, 5192–5207.
Res. 104, 77–85. Xue, J., Su, B., 2017. Significant remote sensing vegetation indices: a review of devel-
Tremblay, N., Fallon, E., Ziadi, N., 2011. Sensing of crop nitrogen status: opportunities, opments and applications. J. Sens. 2017, 17.
tools, limitations, and supporting information requirements. HortTechnology 21, Yang, J., Gong, W., Shi, S., Du, L., Sun, J., Song, S., 2016. Laser-induced fluorescence
274–281. characteristics of vegetation by a new excitation wavelength. Spectrosc. Lett. 49,
Tremblay, N., Wang, Z., Cerovic, Z.G., 2012. Sensing crop nitrogen status with fluores- 263–267.
cence indicators. A review. Agron. Sustain. Dev. 32, 451–464. Yao, X., Zhu, Y., Tian, Y.C., Feng, W., Cao, W., 2010. Exploring hyperspectral bands and
Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring estimation indices for leaf nitrogen accumulation in wheat. Int. J. Appl. Earth Observ.
vegetation. Remote Sens. Environ. 8, 127–150. Geoinform. 12, 89–100.
Ustuner, M., Sanli, F.B., Abdikan, S., Esetlili, M.T., Kurucu, Y., 2014. Crop Type Yao, X., Huang, Y., Shang, G., Zhou, C., Cheng, T., Tian, Y.C., Cao, W., Zhu, Y., 2015.
Classification Using Vegetation Indices of RapidEye Imagery. ISPRS Technical Evaluation of six algorithms to monitor wheat leaf nitrogen concentration. Remote
Commission VII Symposium, Istanbul, Turkey. Sens. 7, 14939.
Vigneau, N., Ecarnot, M., Rabatel, G., Roumet, P., 2011. Potential of field hyperspectral You, J., Li, X., Low, M., Lobell, D., Ermon, S., 2017. Deep gaussian process for crop yield
imaging as a non destructive method to assess leaf nitrogen content in Wheat. Field prediction based on remote sensing data. Association for the Advancement of
Crops Res. 122, 25–31. Artificial Intelligence.
Viña, A., Gitelson, A.A., Nguy-Robertson, A.L., Peng, Y., 2011. Comparison of different Yu, K., Li, F., Gnyp, M.L., Miao, Y., Bareth, G., Chen, X., 2013. Remotely detecting canopy
vegetation indices for the remote assessment of green leaf area index of crops. nitrogen concentration and uptake of paddy rice in the Northeast China Plain. ISPRS
Remote Sens. Environ. 115, 3468–3478. J. Photogramm. Remote Sens. 78, 102–115.
Vrindts, E., Mouazen, A.M., Reyniers, M., Maertens, K., Maleki, M.R., Ramon, H., De Yuan, Y., Zhu, G., Wang, Q., 2015. Hyperspectral band selection by multitask sparsity
Baerdemaeker, J., 2005. Management zones based on correlation between soil pursuit. IEEE Trans. Geosci. Remote Sens. 53, 631–644.
compaction, yield and crop data. Biosyst. Eng. 92, 419–428. Zarco-Tejada, P.J., González-Dugo, V., Berni, J.A.J., 2012. Fluorescence, temperature and
Wahabzada, M., Mahlein, A., Bauckhage, C., Steiner, U., Oerke, E., Kersting, K., 2016. narrow-band indices acquired from a UAV platform for water stress detection using a
Plant Phenotyping using Probabilistic Topic Models: Uncovering the Hyperspectral micro-hyperspectral imager and a thermal camera. Remote Sens. Environ. 117,
Language of Plants. 6, 22482. 322–337.
Wang, J., Shen, C., Liu, N., Jin, X., Fan, X., Dong, C., Xu, Y., 2017. Non-destructive Zhang, L., Zhang, J., Kyei-Boahen, S., Zhang, M., 2010. Simulation and prediction of
evaluation of the leaf nitrogen concentration by in-field visible/near-infrared spec- soybean growth and development under field conditions. Am.-Eurasian J. Agric.
troscopy in pear orchards. Sensors 17, 538. Environ. Sci. 7, 374–385.
Wang, Y., Witten, I., 1997. Inducing model trees for continuous classes, Proc. 9th Eur. Zhao, G., Miao, Y., Wang, H., Su, M., Fan, M., Zhang, F., Jiang, R., Zhang, Z., Liu, C., Liu,
Conf. Mach. Learn. 128–137. P., Ma, D., 2013. A preliminary precision rice management system for increasing both
Whelan, B.M., Taylor, J.A., 2013. Precision agriculture for grain production systems. grain yield and nitrogen use efficiency. Field Crops Res. 154, 23–30.
CSIRO Publishing. Zhou, Y., Song, G., Wang, M., 2012. Wireless sensor network data fusion algorithm based
Wiegand, C.L., Richardson, A.J., Kanemasu, E.T., 1979. Leaf area index estimates for on neural network in the area of agriculture. Sens. Transducers J. 16, 128–136.
wheat from LANDSAT and their implications for evapotranspiration and crop mod- Zhu, Y., Li, Y., Zhou, D., Tian, Y.C., Yao, X., Cao, W., 2006. Quantitative relationship
eling. Agron. J. 71, 336–342. between leaf nitrogen concentration and canopy reflectance spectra in rice and
Wittenberghe, V.S., Verrelst, J., Rivera, J.P., Alonso, L., Moreno, J., Samson, R., 2014. wheat. Acta Ecol. Sin. 26, 3463–3469.

69