A multi-objective neural network based method for ... - Semantic Scholar

Expert Systems with Applications 39 (2012) 10038–10048

Contents lists available at SciVerse ScienceDirect

Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa

A multi-objective neural network based method for cover crop identification from remote sensed data M. Cruz-Ramírez a,⇑, C. Hervás-Martínez a, M. Jurado-Expósito b, F. López-Granados b a b

Department of Computer Science and Numerical Analysis, University of Córdoba, Rabanales Campus, Albert Einstein Building 3rd Floor, 14071 Córdoba, Spain Institute for Sustainable Agriculture, IAS-CSIC, Apdo. 4084, 14080 Córdoba, Spain

a r t i c l e

i n f o

Keywords: Cruciferous Grass weed Multi-classification Multi-objective evolutionary Multispectral Neural networks Olive orchard Remote sensing data

a b s t r a c t One of the objectives of conservation agriculture to reduce soil erosion in olive orchards is to protect the soil with cover crops between rows. Andalusian and European administrations have developed regulations to subsidise the establishment of cover crops between rows in olive orchards. Current methods to follow-up the cover crops systems by administrations consist of sampling and on ground visits of around 1% of the total olive orchards surface at any time from March to late June. This paper outlines a multi-objective neural network based method for the classification of olive trees (OT), bare soil (BS) and different cover crops (CC), using remote sensing data taken in spring and summer. The main findings of this paper are: (1) the proposed models performed well in all seasons (particularly during the summer, where only 48 pixels of CC are confused with BS and 10 of BS with CC with the best model obtained. This model obtained a 97.80% of global classification, 95.20% in the class with the worst classification rate and 0.9710 in the KAPPA statistics), and (2) the best-performing models could potentially decrease the number of complaints made to the Andalusian and European administrations. The complaints in question concern the poor performance of current on-ground methods to address the presence or absence of cover crops in olive orchards. Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction Olive (Olea europaea L.) is an essential crop in the Mediterranean basin, where it covers around 9.2 M ha of which 2.4 M ha are in Spain and 1.5 M ha in Andalusia, southern Spain (MAPA, 2007). Traditionally, olive trees are separated between 10 and 12 m from each other and soil management is mainly based on intensive tillage operations between rows. Such operations have serious agro-environmental implications and consequences for land erosion and degradation, desertification, sediment transport and increase of atmospheric CO2 (Hill, Mégier, Mehl, & degradation, 1995; Schlesinger, 2000). To avoid these negative effects, the European Union (EU) and Andalusian administrations only subsidise the implementation of certain conservation agro-environmental techniques, which mainly consist of altering the natural soil as little as possible and protecting it with cover crops. The EU and Andalusian administrations have developed regulations to subsidise the establishment of cover crops between rows in olive orchards (EU Council Regulation 1257, 1999; EU Council Regulation 1259, 1999; Spanish Royal Decree 4, 2001). The Cover Crop Program in perennial crops is being developed as a part of these agro-environ⇑ Corresponding author. Tel.: +34 957 218 349; fax: +34 957 218 630. E-mail address: [email protected] (M. Cruz-Ramírez). 0957-4174/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2012.02.046

mental measures. The use of cover crops is one of the most effective alternative strategies to manage soil while preventing erosion. Cover crops (2–6 m wide) are sown (grass and broadleaved species [Alcántara, Sánchez, Pujadas, & Saavedra, 2009]; [Castro et al., 2008]; [Hernández, Lacasta, & Pastor, 2005]) in autumn each year (mid-November in Mediterranean conditions). At the beginning of spring (the end of March in Mediterranean conditions), the cover crops are mowed, grazed or desiccated by the application of post-emergence, non-residual herbicides just before cover crops begin to compete with olive trees for water and nutrients. Furthermore, dead cover crops consisting of the remains of the corresponding olive spring pruning are also considered in some olive orchards. A complete description of soil management practices can be found in Castro et al. (2008). Currently, to control these agrarian policy actions, the EU and local administrations require precise records of the presence or absence of cover crops. Current methods to examine the cover crop system and to map cover crop soil coverage by the Andalusian administration consist of sampling and ground visits of around 1% of the total olive orchards’ surface at any time between March and late June (Annex I, CE no. 1782/2003). This procedure is time-consuming, very expensive, and infeasible in field areas with difficult access, and sometimes delivers inconsistent results due to its coverage of relatively small areas or only very few target fields.

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048

The acquisition of data through remote sensing techniques is of great importance in agriculture (Liaghat, 2010). Remotely sensed data may offer the ability to efficiently identify and map crops and cropping methods over large areas (South, Qi, & Lusch, 2004; Yu et al., 2006). These techniques may be cheaper, faster, and more reliable than ground visits. Many studies have used remote sensing tools to map different land uses with high accuracy: for example, weeds and crops (Felton et al., 2002; Hervás-Martínez, Gutiérrez, Peña-Barragán, Jurado-Expósito, & López-Granados, 2010; LópezGranados, Jurado-Expósito, Peña-Barragán, & García-Torres, 2006; Peña-Barragán, López-Granados, Jurado-Expósito, & García-Torres, 2007; Schmidt & Skidmore, 2003) or olive oil growers (GonzálezAndújar, 2009). The accuracy of the thematic map is extremely important, because this map could be used as a tool to help the administration decide whether to continue the subsidy. Furthermore, remotely sensed images provide a tremendous amount of environmental information that can be used for diverse administrative purposes, such as crop inventories and subsidy inspection or follow up, among others. The remote sensing data have also been used in other fields, as in the detection of Ridolfia segetum infestations in sunflower crop (Gutiérrez et al., 2008; Gutiérrez, López-Granados, Peña-Barragán, Jurado-Expósito, & Hervás-Martínez, 2008) or for the delineation and characterization of vineyards (Delenne, Durrieu, Rabatel, & Deshayes, 2010). To detect and map different land uses, the discernment of subtle differences in spectral reflectance between land uses is required, and the spectral and spatial resolution of the remote sensing equipment must be sufficient to discern these differences (Thorp & Tian, 2005; Alchanatis, Ridel, Hetzroni, & Yaroslavsky, 2005). Plant species can often be identified by exploiting differences in their canopy structure (Brown, Stecklert, & Anderson, 1994; Jurado-Expósito, López-Granados, Atenciano, García-Torres, & González-Andújar, 2003) or distinctive phenological stages (PeñaBarragán, López-Granados, Jurado-Expósito, & García-Torres, 2006; Brown & Noble, 2005; Girma et al., 2005; KavdIr, 2004). Previous work suggests that the differentiation between cover crops and all other land uses in olive orchards (i.e., olive trees and bare soil grouped) could be feasible by the analysis of visible band and visible-NIR band imageries aerial photographs taken in summer by means of a two-class classification through traditional discrimination methods such as vegetation indices (Panda, Hoogenboom, & Paz, 2009; Peña-Barragán et al., 2004). However, these cropping systems would need a multi-classification with a high accuracy level as well as an acceptable level for each class (cover crop, bare soil and olive trees), not just discrimination between cover crops and the other two land uses grouped (i.e., bare soil and olive trees). Furthermore, the confusion matrix must elucidate whether misclassification of cover crops is with bare soil or olive trees, because this distinction is crucial to estimate the real field surface occupied by cover crops. The administration, in turn, requires this estimation in order to decide whether to concede or not the subsidy. The problem of assigning pixels of a specific land use to the different pixels analysed is addressed in this paper using a pattern recognition technique. Multi-classification is the problem of building a system that accurately maps an input feature space to an output space with more than two pattern classes. We test a method for discriminating cover crops from bare soil and olive trees in olive orchards as affected by their phenological stage using visible band (400–700 nm) and visible-NIR band (500–900 nm) aerial imagery. To resolve the multi-classification problem that arises, we use a Multi-Objective Evolutionary Algorithm (MOEA) (Deb, 2004) with a population of Multilayer Perceptron (MLP) neural networks (Teixeira, Braga, Takahashi, & Saldanha, 2000; Fernández-Navarro, Hervás-Martínez, García-Alonso, & Torres-Jimenez, 2011; Zhang,

10039

2000). A MOEA, also known as multi-objective optimization, is the process of simultaneously optimizing two or more conflicting objectives subject to certain constraints. As the problem with which we are working is a multi-classification problem, we not only need to have models that maximise the total number of well-classified patterns, but also to maximise the percentage of patterns that are correctly classified in the lowest classified class by the model. The conflicting objectives to be considered for training the models are Accuracy (percentage of patterns correctly classified) and Minimum Sensitivity (minimum of the percentage of patterns correctly classified in each class). Using the Minimum Sensitivity as the objective ensures that none of the three classes will be marginalised by the models obtained. The paper is organised as follows: Section 2 describes the materials and methods (the study sites, aerial photographs, the description of neural networks and the explanations of Accuracy and Minimum Sensitivity). Section 3 describes the learning method used, followed by the experimental design in Section 4. Section 5 shows the results obtained, and finally, conclusions are drawn in Section 6. 2. Materials and methods 2.1. Study sites, materials and experimental design The study was conducted in early spring and early summer 2008 in two olive orchards of 70–100 ha each, located in Andalusia (southern Spain) named Espejo and Montilla (Universal Transverse Mercator, UTM, coordinates: x = 364,226; y = 4,166,454; and x = 351,401, y = 4,157,004, respectively, zone 30 North, European 1950 Spain and Portugal datum). The two models are trained using information obtained in Espejo in the spring and summer and generalised using information obtained in Montilla, in spring and summer. The map of these orchards can be seen in Fig. 1. According to US Department of Agriculture Soil Conservation Service (1975), the soil at every farm was classified as Typic Chromoxeret, and was representative of the entire region. The cover crops considered in this study were representative of the different classes of cover crops used in olive orchards in the Mediterranean region: grass species (Fig. 2(a)); broadleaved species (Fig. 2(b)), and dead cover crops (Fig. 2(c)). Dead cover crops consisted of the remains of the corresponding olive spring pruning. Cover crop species were predominantly grass species (Hordeum murinum L. and Avena sativa L.) in Espejo, and grass species (Avena sativa L.) together with broadleaved species (cruciferous such as Sinapis spp. and Diplotaxis spp.) in Montilla. Furthermore, dead cover crops were also considered in both fields. In Espejo, the cover crops considered in these studies included only two classes of cover crops: grass species and dead cover crops; meanwhile, in Montilla the cover crops represented all possible types of cover crops in olive orchards in the Mediterranean area (grass species, broadleaved species and dead cover crops). Thus, we have considered to train our models using the information from Espejo to analyse their ability to generalise in Montilla, where are represented all types of cover crops. The cover crops were sown in autumn each year (mid-November in Mediterranean conditions) and desiccated with a postemergence herbicide (such as glyphosate) in early spring (the end of March in our conditions). Before the herbicide treatment, grass cover crops in spring had the typical green colour of the vegetative growing phase (Fig. 2(a)), while cruciferous cover crops (broadleaved) were yellowing because they were in the flowering phase (Fig. 2(b)). Just before cover crops begin to compete with olive trees for water and nutrients, the cover crops are desiccated and therefore, in early summer both cover crops, grass and broadleaved, are completely desiccated and have the typical

10040

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048

Fig. 1. Map of Espejo and Montilla orchards.

yellow–brown colour of the senescent phase (Fig. 2(d)). Olive trees are a non-deciduous crop and they keep their green colour during the entire year (Fig. 2). Taking into account the above information, we have developed two classification models, one in early spring and early summer, based only on information obtained in Espejo. Visible band (VB) (400–700 nm) and visible-NIR band (VNB) (500–900 nm) aerial imagery of the studied olive orchards were taken in early spring and early summer 2008. The fields were visited close to when the photographs were taken in order to determine the cover crop vegetative development and the other land uses that could be clearly distinguished in the images (that is, olive trees and bare soil). In addition, 40 ground control points (GCP) in each farm were georeferenced using the submeter differential GPS TRIMBLE PRO-XRS provided with the TDC-1 unit. The photographs were taken on a cloudless day between 12.00 and 14.00 standard time by a turboprop twin-engine CESSNA 310 R plane, provided with an automatic pilot to handle the photographic equipment, and the average height was 1525 m to obtain photographs at scale 1:10,000. The camera was a WILD RC-10, with an objective AF/15 UAG, a focal distance of 153.66 mm, and the film was AGFA AVIPHOT COLOUR H-100 PE-1. The images were taken with a longitudinal recovering device between photographs of 60%. Then, the photographs were digitised with an AGFA Horizon A3 scanner, considering a resolution of 635 dots per inch (dpi), without adjusting brightness and contrast on the digitised images. The next step was to orthorectify the digitised images, using the fiducial marks of the aerial calibration certificate and 40 GCPs taken with the DGPS. These GCPs were road crossings, big stones, corners of buildings (farms and other farm buildings), farm

entrance and other hard-edge points present in the images. Finally, images were resampled to a pixel size representing a 40 cm  40 cm ground area. This pixel size was based on the inherent spatial properties of the input data, i.e., the width of the cover crops (2–6 m wide). In Hengl (2006), the author discussed the rules of thumb to find the right pixel size to select the appropriate spatial resolution. He concluded that at least four pixels are required to detect smallest objects and at least two pixels to represent the narrowest objects, being objects the smallest size area that we map (cover crops in our case). In other words, if the smallest object is a cover crop of at least 2 m wide, we should use imagery with resolution of 0.5 m and finer. Input variables included the digital values of all bands in each available image, that is: VB images responded to the blue (B, 400–500 nm), green (G, 500–600 nm), and red (R, 600–700 nm) broad bands of the electromagnetic spectrum, and VNB images to the G, R and near-infrared (NIR, 700–900 nm) bands. The scanner produced an RGB digital image with an 8-bit true colour, so pixels of the image showed digital counts within the range of 0–255 values. All spatial and spectral data (the information extracted from the different pixels of the images, for each class considered, that is, olive trees, cover crops and bare soil: red, green, blue, and near-infrared digital values) were grouped and saved in a unique multi-band file taking into account the two previous requirements: (1) the georeferenced errors between images were less than one pixel, so that similar pixels had the same coordinate, and (2) the NIR digital values of VNB images were corrected to the digital values of VB images, using the differences between the green and red bands of both original images.

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048

10041

Fig. 2. Images of the different elements to discriminate.

To train and validate the classification models, a random ground sampling procedure was carried out at the time when aerial images were taken, ensuring that all the variability within classes was taken into account and all parts of the olive orchards’ area had equal chance of being sampled with no operator bias (McCoy, 2005). We georeferenced at each sampling date (early spring and early summer) at each location a total of 3000 sampling points, i.e., 1000 sampling points for every soil use: grass (Hordeum and Avena) and dead cover crops, olive trees and bare soil at Espejo farm, and broadleaved (Sinapis and Diplotaxis) and dead cover crops, olive trees and bare soil at Montilla Farm. Therefore, the total number of training patterns was 3000 sampling points in Espejo and another 3000 sampling points for generalisation in Montilla in early spring. Other 3000 sampling points in Espejo was used for training and another 3000 sampling points for generalisation in Montilla in early summer.

2.2. Multi-objective evolutionary neural networks

Fig. 3. Structure of standard feedforward Multilayer Perceptron for a multiclassification in J classes.

The use of evolutionary approaches for ANN training (known as Evolutionary Artificial Neural Networks, EANNs) have been a key research area for the last years (Yao & Liu, 1997). Methods and techniques have been developed to find better approaches to evolve ANNs, trying to design networks with good generalisation capability. On the other hand, the issue of finding a good ANN

architecture has also been debated in the field of ANNs. The main advantages of evolutionary approaches to ANN training are its ability to escape a local optimum, its robustness and its ability to adapt itself to a changing environment. The ANN architecture used in this work can be seen in Fig. 3 (where J is the number of classes).

10042

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048

Fig. 4. Unfeasible region in the two-dimensional (MS, C) space for a given classification problem.

In other matters, Pareto-based techniques (Jin & Sendhoff, 2008), specifically Multi-objective Evolutionary Algorithms (MOEAs), should provide a homogeneous distribution of population along the Pareto frontier together with an improvement of the solutions along successive generations. Also, these techniques present an uncountable set of solutions that, when they are evaluated, produce vectors whose components represent a trade-off in the objective space. A decision maker then implicitly chooses an acceptable solution (or solutions) by selecting one or more of these vectors. For example, Abbass has presented different studies on the design and training of ANNs using the accuracy and the complexity of the network (considering the number of nodes in the hidden layer and the number of connections) as objectives (Abbass, 2003; Abbass, Sarker, & Newton, 2001). Abbass, along with Ilonen, Kamarainen, and Lampinen (2003), Yau, Teo, and Anthony (2007) and other authors, use Differential Evolution (DE) as the basis of their multi-objective algorithms (Price, Storn, & Lampinen, 2005). The training of ANN by Evolutionary Pareto-based algorithms is known as Multiobjective Evolutionary Artificial Neural Networks (MOEANNs), and has been used to solve classification tasks (Ou & Murphey, 2007), some of its main exponents being Abbass (2002) and Jin and Sendhoff (2008).

2.3. Accuracy and minimum sensitivity in classification problems Statistics and machine learning communities have traditionally used the Correct Classification Rate or Accuracy (C), to measure the performance of a classifier generally avoiding the classification level of each class in the results. However, the pitfalls of using Accuracy have been pointed out by Provost et al. (1997, 1998). Actually, it is enough to realise that Accuracy cannot capture all the different behavioral aspects found in two different classifiers. Assuming that all misclassifications are equally costly and there is no profit for a correct classification, we assume that a good classifier should obtain a high Accuracy level as well as an acceptable level for each class. In real problems these objectives are usually in competition. Achieving an high Accuracy classification level usually means sacrificing the classification in some class. This problem is especially significant when we deal with classification problems that differ in their prior class probabilities (class imbalances) or where there are a great number of classes (Ho & Basu, 2002).

For these problems, two performance measures are considered; P traditionally-used Accuracy, C ¼ N1 Qj¼1 njj (where Q is the number of classes, N is the number of patterns in training or testing and njj is the number of patterns from class jth that are correctly classified), and the Minimum Sensitivity (MS), that is, the lowest percentage of examples correctly predicted as belonging to each class, Si, with respect to the total number of examples in the corresponding class, MS = min{Si}. The (MS, C) pair expresses two features associated with a classifier: global performance (C) and the rate of the worst classified class (MS). One point in (MS, C) space dominates another if it is above and to the right, i.e., it has greater C and the best MS. Let C and MS be associated with a classifier g, then:

MS 6 C 6 1  ð1  MSÞp where p⁄ is the minimum for estimated prior probabilities, value that has an important role in the relationship between the two measures.

p ¼ #Minority class patterns=#Total patterns Each classifier will be represented as a point in the white region in Fig. 4, hence the area outside of the triangle is marked as unfeasible. The area inside the triangle in Fig. 4 may be feasible or not (attainable), depending upon the classifier and the difficulty of the problem. A priori, it could seem that MS and C objectives could be positively correlated, but while this may be true for small values of MS and C, it is not so for values close to 1 on both MS and C, where the objectives are competitive and conflicting (for a more detailed description of these measures and this figure, see Martínez-Estudillo et al. (2008) and Fernández-Caballero, Martínez-Estudillo, Hervás-Martínez, & Gutiérrez (2010)). 3. Learning method 3.1. Base classifier and fitness functions For multi-classification problems, we considered feedforward Multilayer Perceptron (MLP) neural networks with one input layer with independent variables or features, one hidden layer with sigmoidal hidden nodes and one output layer with linear nodes. A scheme of these models is given in Fig. 3, where J is the number

10043

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048 Table 1 Characteristics for datasets. Dataset

#Patterns

#Training patterns

#Test patterns

#Input variables

#Classes

#Patterns per class

p⁄

Spring Summer

6000 6000

3000 3000

3000 3000

3 4

3 3

(2000, 2000, 2000) (2000, 2000, 2000)

0.3333 0.3333

of classes (in our case, J = 3, since we had three classes: Cover Crop (CC), Olive Tree (OT) and Bare Soil (BS)). MLP models have been described in Bishop (1996) and Haykin (1998). The fitness functions used in the evolutionary process are a strictly decreasing transformation of the Entropy (A(g)) and the MS of the classifier. These fitness measure are explained in more detail in Cruz-Ramírez, Sánchez-Monedero, Fernández-Navarro, Fernández, and Hervás-Martínez (2010). 3.2. Algorithm training In this paper, we use one of the most prominent Multi-Objective Evolutionary Algorithms1 in the bibliography and a memetic algorithm developed based on them. Evolutionary techniques have been widely used in the late years for training and automatically designing neural networks (adjusting the weights and architecture) (García-Pedrajas, Hervás-Martínez, & Muoz-Pérez, 2003). These techniques perform a global exploration of the search space avoiding to become trapped in local minimum as usually happens with local search procedures. Moreover, because evolutionary techniques adjust the network structure, prevents the obtained models produce overtraining or undertraining. First we use the PDE (Pareto Differential Evolution) algorithm developed by Storn and Price (1997), modified by Abbass et al., 2001 and Abbass, 2001 and adapted for C and MS by J.C. Fernández, Hervás, Martínez, Gutiérrez, and Cruz (2009). The fundamental bases of this algorithm are Differential Evolution (DE) and the concept of Pareto dominance. The main feature of the PDE algorithm is the inclusion of a crossover operator together with the mutation operator. The crossover operator is based on a random choice of three parents, where one of them (main parent) is modified using the weighted difference of the two other parents (secondary parents). The child generated by the crossover and mutation operator is included in the population if it dominates the main parent, if it has no relationship with him or if it is the best child of the rejected children. At the beginning of each generation, individuals dominated are eliminated from the population. A generation of the evolutionary process ends when the population has been completed. The following algorithm we use is a memetic of the PDE, which is described in Fernández et al. (2009). In our case, we will make the local search process 10% of the population in three generations of the evolution (the first initially, the second in the middle and the third at the end), rather than applying them all. This is to reduce the computational cost of the algorithm. The local search algorithm used is iRprop+ (Igel & Hüsken, 2003) due to it is one of the most accurate, robust, and quickly converging techniques. This algorithm will be called the Memetic Pareto Differential Evolutionary (MPDE). To compare the obtained result with PDE and MPDE, we used the well-known algorithm NSGA2 (Deb, Pratab, Agarwal, & Meyarivan, 2002) implemented by our research group, and in particular one of its hybrids presented in Fernández-Caballero et al. (2010), Fernández-Caballero, Hervás-Martínez, Martínez-Estudillo, and Gutiérrez (2011), which we will call HNSGA2, and the Multilayer Perceptron (MLP) neural networks implemented in Weka (Witten & Frank, 2005). 1 The source code in JAVA for PDE, MPDE, NSGA2 and HNSGA2 is freely available upon request to the authors.

4. Experiments In the experimental design, we consider two datasets, one with data from all locations in spring and the other with data in summer. In both cases, Espejo’s patterns are used as the training set and Montilla’s patterns as the test set. No validation set was used because in the literature on neural networks there is no agreement as to the efficiency of the models trained with this procedure (García-Pedrajas et al., 2003; García-Pedrajas, Hervás-Martínez, & Ortiz-Boyer, 2005). In Table 1, we can see the features for each dataset. We show the total number of instances by each dataset, the number of instances in training and testing, the number of input variables, the number of classes (outputs), the total number of instances per class and the p⁄ value. During the experiment, models are trained using the fitness function A(g) and MS as objective functions, but when validating, we use C and MS. A(g) is used instead of C in training because C is a discontinuous function, which makes convergence more difficult in optimization. Once the Pareto front is built, two methods are considered in order to build a neural network model that then includes the information about the models within it. These are called MethodNameE and MethodName-MS (where MethodName takes the values PDE, MPDE, NSGA2 and HNSGA2). These methods provide single models that can be compared to other classification methods found in the literature. The process followed in these methods is the following: once the first Pareto front is calculated using training set patterns, the best individual belonging to the Pareto front on E(EI) is chosen for MethodName-E, and the best individual in terms of MS(MSI) is selected for MethodName-MS. Therefore   EI we obtain an individual EIG ¼ C EI and an individual G ; MSG  MSI MSI MSIG ¼ C G ; MSG Þ. This is repeated 30 times because our learning algorithm is stochastic and then estimations are carried out of the average and standard deviation obtained from the individu    EI MSI and MSIG ¼ C MSI . The first expression als EIG ¼ C EI G ; MSG G ; MSG is the average obtained taking E into account as the primary objective, and the second one is obtained by taking MS into account as the primary objective. So, the opposite extremes of the Pareto front are taken in each of the executions. In Fig. 5, the process is shown graphically. In all experiments, for PDE and MPDE the population size is established to M = 50. The probability of crossover is set at 0.8 and the mutation probability of 0.1. Moreover, in MPDE the optimisation will process three times during execution (each 33.33% of generations). For NSGA2 and HNSGA2, the population size is established at M = 150 and the mutation probability for each operator is 1/5. 5. Results and discussion The results indicated that it is possible to make a multi-objective classification of the different land uses considered in olive orchards (cover crops, olive trees and bare soil in spring and summer) using remote sensing data. In Tables 2 and 3 we present the values of the mean and the standard deviation for C, MS and KAPPA obtained in training and generalisation (that is, in Espejo

10044

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048

Fig. 5. Scheme to obtain statistical results.

Table 2 Statistical results for PDE, MPDE, NSGA2, HNSGA2 and MLP method in training (Espejo). The best result is in bold face and the second best result in italic. Dataset

Method

C (%) Mean ± SD

MS (%) Mean ± SD

KAPPA Mean ± SD

Spring

PDE-E PDE-MS MPDE-E MPDE-MS NSGA2-E NSGA2-MS HNSGA2-E HNSGA2-MS MLP

88.01 ± 1.48 88.52 ± 1.67 87.69 ± 1.15 87.98 ± 1.72 87.33 ± 1.09 89.34 ± 1.56 88.85 ± 1.02 90.36 ± 1.22 92.69 ± 0.34

82.83 ± 3.34 87.84 ± 1.73 82.87 ± 3.02 87.25 ± 1.73 80.91 ± 4.23 88.97 ± 1.62 84.56 ± 1.73 89.95 ± 1.29 90.10 ± 2.03

0.8201 ± 0.0223 0.8278 ± 0.0251 0.8153 ± 0.0173 0.8197 ± 0.0259 0.8101 ± 0.0165 0.8401 ± 0.0235 0.8328 ± 0.0154 0.8554 ± 0.0183 0.8903 ± 0.0051

Summer

PDE-E PDE-MS MPDE-E MPDE-MS NSGA2-E NSGA2-MS HNSGA2-E HNSGA2-MS MLP

81.32 ± 0.95 81.73 ± 1.04 81.81 ± 0.90 81.75 ± 2.10 77.35 ± 2.23 79.12 ± 2.19 80.03 ± 1.30 80.22 ± 1.38 83.43 ± 0.44

69.91 ± 2.40 73.04 ± 1.18 70.85 ± 2.00 73.49 ± 1.27 57.78 ± 10.06 70.44 ± 1.66 65.12 ± 5.70 71.58 ± 0.92 70.31 ± 4.42

0.7198 ± 0.0143 0.7260 ± 0.0156 0.7272 ± 0.0135 0.7263 ± 0.0315 0.6603 ± 0.0335 0.6869 ± 0.0329 0.7004 ± 0.0195 0.7033 ± 0.0208 0.7514 ± 0.0067

and Montilla, respectively) in 30 runs for all the experiments performed. We note that the best results in training were obtained with MLP method, showing values of C and MS higher than 90 and 70% in spring and summer, respectively, and higher values in KAPPA, although all methods attained similar results (Table 2). Thus, in training in spring, the number of pixels of the worst class that was correctly classified, considering all the models, was higher than 800 (of a total of 1000), with MLP method as the most accurate with 901 pixels correctly classified. In contrast, the best-performing models in generalisation were MPDE-MS and MPDE-E with C values of 81.28% and 86.78% and KAPPA values of 0.7193 and 0.8017 in spring and summer, respectively. These models could correctly classify 537 and 710 pixels of the worst classified class in spring and summer, respectively (Table 3).

In Fig. 6 are represented individuals belonging to the best-performing iteration. The first type of graphic (Figs. 6(a) and 6(c)) shows the Pareto fronts formed by the individuals trained during the training stage (emphasising the first front). This graphic compares the MS with A(g). The other graphic (Figs. 6(b) and 6(d)) shows individuals obtained during the generalisation phase, comparing the MS with C. These individuals are within the feasible region, as explained in Section 2.3. In order to determine whether there are significant differences in mean Accuracy (C) and in the Minimum Sensitivity (MS) of the nine methods applied (PDE-E, PDE-MS, MPDE-E, MPDE-MS, NSGA2-E, NSGA2-MS, HNSGA2-E, HNSGA2-MS and MLP method) over the generalisation datasets (in spring and summer), different statistical tests have been applied. These tests have not applied using the KAPPA measure as the variable test, because this metric is equivalent to the Accuracy measure. First of all, the Kolmogorov–Smirnov (K-S) statistical test concluded, that the C and MS distributions of the best models in training and in the generalisation sets, follow a normal distribution for a standard asymptotic significance level, a = 0.05. A Levene’s (1960) test has been applied in order to assess whether there are significant differences in variance, concluding that the differences are significant, except for C in Spring (second column in Table 4, p-value > 0.05). On the basis of these hypotheses, an ANOVA I analysis (Snedecor & Cochran, 1989) has been performed where the factor to consider is the method applied and the control variables were C and MS for generalisation datasets (in Spring and Summer). In Table 4, the significance values of Snedecor’s F test are shown in the third column and according to them, we can conclude that for generalisation sets, there are significant differences in the means of C and MS (p-values > 0.05). For spring and summer datasets, we propose using parametric multiple comparison contrasts for the nine applied methods with the aim of ranking the averages of the C and MS obtained with each method. The Tamhane or Tukey test (Miller, 1986) has been applied and the results from all these tests are shown as a method ranking in the fourth column of the table. This column first shows the ranking of the averages for the nine applied methods, and then shows some pairwise comparisons with

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048 Table 3 Statistical results for PDE, MPDE, NSGA2, HNSGA2 and MLP method in generalisation (Montilla). The best result is in bold face and the second best result in italic. Dataset

Method

C (%) Mean ± SD

MS (%) Mean ± SD

KAPPA Mean ± SD

Spring

PDE-E PDE-MS MPDE-E MPDE-MS NSGA2-E NSGA2-MS HNSGA2-E HNSGA2-MS MLP

78.75 ± 3.17 80.69 ± 2.37 78.30 ± 3.05 81.28 ± 4.43 79.31 ± 2.51 81.01 ± 2.47 80.21 ± 1.34 81.17 ± 3.75 76.05 ± 6.36

46.89 ± 8.31 50.46 ± 9.11 49.65 ± 7.93 53.71 ± 13.20 49.91 ± 8.50 50.18 ± 8.21 49.18 ± 4.90 50.09 ± 7.72 33.26 ± 5.23

0.6812 ± 0.0476 0.7104 ± 0.0356 0.6745 ± 0.0458 0.7193 ± 0.0665 0.6897 ± 0.0378 0.7152 ± 0.0371 0.7032 ± 0.0206 0.7176 ± 0.0562 0.6408 ± 0.0955

Summer

PDE-E PDE-MS MPDE-E MPDE-MS NSGA2-E NSGA2-MS HNSGA2-E HNSGA2-MS MLP

82.07 ± 10.30 79.10 ± 12.53 86.78 ± 9.14 83.03 ± 12.52 75.36 ± 12.80 70.78 ± 11.13 79.94 ± 10.06 77.55 ± 10.41 76.05 ± 2.80

55.11 ± 29.73 55.75 ± 28.86 70.18 ± 23.16 63.80 ± 24.74 35.99 ± 34.62 34.20 ± 27.67 46.68 ± 31.98 48.62 ± 27.25 56.67 ± 17.70

0.7311 ± 0.1546 0.6865 ± 0.1879 0.8017 ± 0.1371 0.7455 ± 0.1878 0.6304 ± 0.1920 0.5618 ± 0.1669 0.6991 ± 0.1509 0.6633 ± 0.1561 0.6408 ± 0.0420

significant differences. Finally, the table includes the p-values of a t-test only comparing the MPDE-MS versus HNSGA2-MS and MPDE-MS versus PDE-MS methods in Spring and MPDE-E versus MPDE-MS in Summer (for MPDE-E versus MPDE-MS applying a t-test with dependencies, ‘‘Dep. t-test’’), because these are the best pairs of methods for each dataset. The MLP network is the best method in both training set (Spring: CT = 92.69%, MST = 90.10, KAPPAT = 0.8903; Summer: CT = 83.43%, MST = 70.31, KAPPAT = 0.7514), but it makes the worst classification in both generalisation set (Spring: CG = 76.05%, MSG = 33.26, KAPPAG = 0.64.08; Summer: CG = 76.05%, MSG = 56.67,

10045

KAPPAG = 0.6408). This is due to overtraining; the MLP network adapts to the training set and cannot correctly classify the generalisation set. Thus, the best method of the generalisation set in spring is the memetic multi-objective algorithm based in PDE, considering that the best model that maximises the Minimum Sensitivity (located on the right end of the first Pareto front) is MPDE-MS. This indicates that in spring, trying to improve the classification of the worst classified class further improves the performance of others, because this method gets the best values in C and MS as shown in the rank test. Besides, the t-test indicates that there are no differences in C with the next best method (HNSGA2-MS) (p-value = 0.92), or in MS with PDE-MS (p-value = 0.27). In summer, the best method is the memetic multi-objective algorithm based on PDE, considering that the best model that maximises the Accuracy (located on the left end of the first Pareto front), is MPDE-E. This is because it gets the best values in C and MS, and the means of these differences are also significant (as seen in Table 4, p-values = 0.01 in C and 0.03 in MS) when compared by the Student’s t test with dependent populations (the second best method is the same method except for choosing instead the best individual of the Pareto front that maximises MS). In summer, it is best to choose methods that attempt to improve the overall percentage of correctly classified patterns as well as improve the classification of the worst classification class. The C, the MS, KAPPA and confusion matrices of the best model for spring are included in Table 5. In spring, the best model (MPDEMS) performs a better classification in the training set (i.e., Espejo, CT = 91.66%, MST = 91.50%, KAPPAT = 0.8750) better than the generalisation set (Montilla, CG = 83.33%, MSG = 52.50%, KAPPAG = 0.7500). These differences in classification could be because in Espejo in spring, the cover crops have the typical green colour of the vegetative phase of grass cover crops, while in Montilla, cover crops are mainly cruciferous and yellow in their flowering stage.

Fig. 6. Pareto front in training (MS, A(g)) and (MS, C) associated values in testing in one specific run of the 30 runs carried out.

10046

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048

Table 4 p-Values of the equality of variances Levene’s test, p-values of the Snedecor’s F test, ranking of averages of the Tukey or Tamhane statistical multiple comparison tests in the generalisation sets (Montilla) using the nine different methods and p-values of the t-test comparing MPDE-MS vs PDE-MS and MPDE-MS vs PDE-MS methods in spring and MPDEE vs MPDE-MS in summer. Dataset

All methods

MPDE-MS vs HNSGA2-MS

MPDE-MS vs PDE-MS

MPDE-E vs MPDE-MS

Lev.

F-test

Tukey or Tamhane ranking test

t-Test

t-Test

Dep. t-test

Spring C

0.18

0.00⁄

0.92

–

–

Spring MS

0.00⁄

0.00⁄

–

0.27

–

Summer C

0.00

⁄

⁄

–

–

0.01⁄

Summer MS

0.00⁄

lMPDEMS P lHNSGA2MS P lNSGA2MS P lPDEMS P lHNSGA2E P lNSGA2E P lPDEE P lMPDEE P lMLP; lMPDEMS > lPDEMS;lNSGA2MS lMPDEE lMPDEMS P lPDEMS P lNSGA2MS P lHNSGA2MS P lNSGA2E P lMPDEE P lHNSGA2E P lPDEE > lMLP lMPDEE P lMPDEMS P lPDEE P lHNSGA2E P lPDEMS P lHNSGA2MS P lMLP P lNSGA2E P lNSGA2MS; lMPDEE > lHNSGA2MS; lPDEMS > lNSGA2MS lMPDEE P lMPDEMS P lMLP P lPDEMS P lPDEE P lHNSGA2MS P lHNSGA2E P lNSGA2E P lNSGA2MS; lMPDEE > lMLP > lNSGA2MS; lMPDEMS > lPDEMS

–

–

0.03⁄

0.00

0.00⁄

⁄

The average difference is significant with p-values = 0.05; lA P lB: method A yields better results than method B, but the differences are not significant; lA > lB: method A yields better results than method B with significant differences. The binary relation P is not transitive.

Table 5 The best model for spring.

Table 6 The best model for summer.

Best model for spring in (MS, C) space considering MS (MPDE-MS) pCC ðx; hÞ ¼

ef1 ðx;hÞ P 2 1þ efi ðx;hÞ i¼1

; pOT ðx; hÞ ¼

ef2 ðx;hÞ P 2 1þ efi ðx;hÞ i¼1

; pBS ðx; hÞ ¼

1

P2

1þ

i¼1

efi ðx;hÞ

Best model for summer in (MS,C) space considering C (MPDE-E) pCC ðx; hÞ ¼

f ðx;hÞ

1þ

e1 P 2

i¼1

efi ðx;hÞ

; pOT ðx; hÞ ¼

f ðx;hÞ

e2 P 2

1þ

i¼1

efi ðx;hÞ

; pBS ðx; hÞ ¼

1þ

P21 i¼1

efi ðx;hÞ

f1(x,h) = 5.71  4.96MLP1  4.93MLP2  5MLP3  5MLP4 f2(x,h) = 1.58  5MLP1 + 5MLP2  2.58MLP3  5MLP4 f3(x,h) = 0 MLP1 = r(2.59 + 0.69Bw  0.20Gw + 5Rw) MLP2 = r(0.30 + 3.00Bw  4.98Gw  1.51Rw) MLP3 = r(3.33 + 3.01Bw + 1.50Gw + 4.98Rw) MLP4 = r(1.72  2.37Gw + 4.79Rw)

f1(x,h) = 1.39  4.97MLP1 + 3.82MLP2 + 0.21MLP3 + 4.89MLP4 f2(x,h) = 3.75 + 4.96MLP1  5MLP2  5MLP3 + 4.94MLP4 f3(x,h) = 0 MLP1 = r(2.56 + 5Bw  2.07Gw  4.99Rw + 1.26NIRw) MLP2 = r(4.22  4.86Bw + 4.95Gw + 4.94Rw  0.17NIRw) MLP3 = r(1.42 + 4.36Bw  4.72Gw + 4.78Rw + 3.15NIRw) MLP4 = r(0.11  2.72Bw  1.15Gw  4.99Rw  4.93NIRw)

Bw, Gw, Rw 2 [1, 1] CT = 91.66; MST = 91.50; KAPPAT = 0.8750 T  Espejo CG = 83.33; MSG = 52.50; KAPPAG = 0.7500 G  Montilla

Bw, Gw, Rw, NIRw 2 [  1,1] CT = 82.20; MST = 73.30; KAPPAT = 0.7330 T  Espejo CG = 97.80; MSG = 95.20; KAPPAG = 0.9710 G  Montilla

#neurons = 4; #effective connections = 25 0 1 0 1 C2 C3 C2 C3 C1 C1 B 916 63 B 525 185 290 C 21 C B C B C ; CMG ¼ @ CM T ¼ @ 78 915 7 A 17 975 8 A 80 1 919 0 0 1000 C1  cover crops (CC); C2  olive trees (OT); C3  bare soil (BS)

#neurons = 4; #effective connections = 30 0 1 0 1 C2 C3 C2 C3 C1 C1 B 733 B 952 0 267 C 0 48 C B C B C ; CM G ¼ @ CMT ¼ @ 0 1000 0 A 0 1000 0 A 265 2 733 10 0 990 C1  cover crops (CC); C2  olive trees (OT); C3  bare soil (BS)

These differences in phenological stages are highly related to the spectral signatures of cover crops and olive trees, as reported in Alcántara et al. (2009). They showed that grass cover crops have no significant differences from olive trees using field spectroradiometer data taken in spring. That means that grass cover crops at Espejo are spectrally similar to olive trees (they keep their green colour during the whole year, (Fig. 2)), while at Montilla, pixels corresponding to cruciferous cover crops are spectrally similar to bare soil. Therefore, the misclassification of cover crops in spring at Espejo (i.e., in training) is mainly on olive trees (63 pixels corresponding to grass cover crops were misclassified as olive trees). By contrast, at Montilla, the misclassification of cover crops is mainly on bare soil (290 pixels of cover crops made up of cruciferous are misclassified as bare soil), because the spectral difference between cruciferous cover crops and bare soil is lower than between cruciferous and olive trees. The C, the MS, KAPPA and confusion matrices of the best model for summer are included in Table 6. In summer, all cover crops (grass and cruciferous) turn brown and become completely desiccated due to the application of herbicide. The best model (MPDE-E) for summer performed a better classification in generalisation (i.e., Montilla, CG = 97.80%, MSG = 95.20%, KAPPAG = 0.9710) than in training (Espejo, CT = 82.20%, MST = 73.30%, KAPPAT = 0.7330). In summer, there are few spectral differences between grass or

cruciferous desiccated cover crops and bare soil (Alcántara et al., 2009). Therefore, the misclassification of cover crops is with bare soil, not with olive trees, because it keeps its green colour in both Montilla and Espejo. The number of pixels of cover crops misclassified as bare soil in Espejo (267 pixels of a total of 1000) was much higher than in Montilla (48 of 1000). These differences in the number of pixels of cover crops misclassified as bare soil are probably because the percentages of coverage and biomass (number of plants/m2) reached by grass cover crops are lower than those reached by cruciferous cover crops. Thus, the percentage of coverage of cover crops made up of broadleaved species is higher (around 20%) than when they are made up of grassy species. This is particularly noticeable in cruciferous species, and several authors have documented that cruciferous cover crops, (especially Sinapis alba L.) are a good alternative to grass cover crops or other types of broadleaved cover crops in olive orchards because they have good emergence, excellent coverage and high production of biomass (Alcántara et al., 2009). Poor coverage means that bare soil is apparent in many parts of the area occupied by grass cover crops and a higher number of pixels of the desiccated cover crops can be confused with pixels of bare soil. The method presented in this paper increased the overall percentage of correctly classified patterns obtained in previous studies as well as improving the classification of the worst class classified,

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048

that is, cover crops and olive trees in spring, and bare soil in summer. The best-performing models (MPDE-MS and MPDE-E) obtained generalisation accuracies of 83.33% in spring and 97.80% in summer. In Peña-Barragán et al. (2004) mapped cover crops in olive orchards by the analysis of aerial photographs acquired in spring and summer through traditional discrimination methods such as vegetation indices. They obtained maps of cover crops with 92% of accuracy in summer by means of a two-class classification that grouped together bare soil and olive trees, distinguishing this group from cover crops. In this pixel-based image classification, spring cover crops could not be distinguished from olive trees, whereas both land uses were discriminated from bare soil. By contrast, none of the vegetation indices analysed were adequate for bare soil discrimination in summer. Our method is better, because it classified bare soil with an accuracy of around 99–100% in spring and summer. Our results constitute a powerful tool for the discrimination of cover crops, olive trees and bare soil in spring and summer. Therefore, there would be different timeframes for image acquisition. A wide timeframe is essential for properly mapping cover crops, olive trees and bare soil with high-resolution airborne imagery, especially taking into account that cloudy days are very common in the Mediterranean region in the spring and no remote images can be taken in these circumstances. If image acquisition fails in spring, we could program the remote imagery acquisition at the beginning of summer (June) when there are plenty of sunny days. In addition, this double possibility could be used to re-monitor a dubious and specific field analysed in previous spring. This is essential for programming and implementing the control tools and for avoiding the common and annoying bottleneck of administrative follow-up about whether to continue the subsidy. Multi-classification (between cover crops, bare soil and olive trees, not only the discrimination between cover crops and the other two land uses grouped together) is needed in olive orchards under conservation agriculture, because distinguishing between bare soil and olive trees is crucial for the E.U. and local administrations to estimate real cover crop surfaces. Once we have verified the excellent performance of MPDE-MS and MPDE-E models to discriminate cover crops, olive trees and bare soil in both seasons (particularly in summer), we can state that using our approaches (multi-objective evolutionary neural networks), only 48 pixels of cover crops are misclassified as bare soil and 0 pixels are misclassified as olive trees in summer. That implies an important improvement in the decision to continue the subsidy, because every year, a number of farmers complain to the Andalusian administration about the low performance of current on-ground methods to follow-up on the presence or absence of cover crops in olive orchards. These results showed that remote sensing could be used instead of time-consuming on-ground-farm visits, reducing costs by about 75%. Furthermore, permanent quantitative records can be obtained that support tentative administrative claims from the farmers (Peña-Barragán et al., 2004). Finally, we emphasise that with a memetic MOEA based on PDE (MPDE-E and MPDE-MS) are obtained not only a very high accuracy in generalisation (Montilla, a farm different from Espejo) but also what in our opinion is more important: an accuracy of worst class that is much higher than that obtained with the other procedures, which show great homogeneity in the accuracy of the classifier on the three classes (cover crop, bare soil and olive trees). 6. Conclusions We present the application of a new method, multi-objective evolutionary neural networks. To the best of our knowledge, this method has never been previously applied in remote sensing, for

10047

olive trees, bare soil and a number of different cover crops in olive orchards. Farmers sow different grass and broadleaved cover crops or keep the remains of the corresponding olive spring pruning, which means that olive orchards’ types of cover crops vary greatly in their effects on the decision of conceding or not the subsidy continues. Moreover, cover crops, olive trees and bare soil have similar or different spectral signatures depending on the season. That indicates that identification of the three land uses is a difficult task. To address this task, our proposal involves the application of a new method that improves the overall percentage of correctly classified patterns over the percentage obtained in previous studies, as well as improving the classification of the worst classified class (that is, cover crops and olive trees in spring and bare soil in summer) over traditional remote sensing classification, such as vegetation indices. The best model obtained with data taken in spring provided a global classification of 83.33% and a classification of the worst classified class of 52.50%. These values are improved by the best model with the data obtained in summer. This model achieved an global classificatio of 97.80% and a minimum sensitivity of 95.20%. These values are obtained with the generalisation set, and can serve as an example of the application of the models. To take full advantage of our results, future investigations could explore the potential of high-resolution satellite imagery, such as QuickBird for mapping cover crops in larger areas. The QuickBird imagery provides four channels (B, G, R and NIR) of multispectral wavebands with 2.4 m or 2.8 m of spatial resolution, and it would be very interesting to test whether it is possible to successfully map cover crops on surfaces of at least 8400 ha (the minimum scene that can be acquired). This greater amount of surface in a unique country dedicated to olive orchards is very common in Mediterranean conditions, and may contribute to optimising the procedure. Acknowledgements This work was supported in part by the Spanish Inter-Ministerial Commission of Science and Technology under Project TIN2011–22794, the Spanish Minister of Science and Innovation by project AGL2011–30442-CO2–01 (FEDER), the European Regional Development fund and the ‘‘Junta de Andalucía’’ (Spain), under Project P2011-TIC-7508. M. Cruz-Ramírez’s research has been subsidized by the FPU Predoctoral Program (Spanish Ministry of Education and Science), grant reference AP2009–0487. References Abbass, H. A. (2001). A memetic Pareto evolutionary approach to artificial neural networks. In M. Brooks, D. Corbet, & M. Stumptner (Eds.), AI2001, LNAI 2256 (pp. 1–12). Springer-Verlag. Abbass, H. (2002). An evolutionary artificial neural networks approach for breast cancer diagnosis. Artificial Intelligence in Medicine, 25(3), 265–281. Abbass, H. (2003). Speeding up backpropagation using multiobjective evolutionary algorithms. Neural Computation, 15, 2705–2726. Abbass, H.A., Sarker, R., Newton, C. (2001). PDE: a Pareto-frontier differential evolution approach for multi-objective optimization problems. In: Proceedings of the 2001 Congress on Evolutionary Computation, Vol. 2, Seoul, South Korea, 2001. Alcántara, C., Sánchez, S., Pujadas, A., & Saavedra, M. (2009). Brassica species as winter cover crops in sustainable agricultural systems in southern spain. Journal of Sustainable Agriculture, 33(6), 619–635. Alchanatis, V., Ridel, L., Hetzroni, A., & Yaroslavsky, L. (2005). Weed detection in multi-spectral images of cotton fields. Computers and Electronics in Agriculture, 47(3), 243–260. Bishop, C. M. (1996). Neural networks for pattern recognition. Oxford, UK: Oxford University Press. Brown, R. B., & Noble, S. D. (2005). Site-specific weed management: sensing requirements - what do we need to see? Weed Science, 53(2), 252–258. Brown, R. B., Stecklert, J.-P. G. A., & Anderson, G. W. (1994). Remote sensing for identification of weeds in no-till corn. Trans. ASAE, 37(1), 297–302. Castro, J., Fernández-Ondoo, E., Rodríguez, C., Lallena, A., Sierra, M., & Aguilar, J. (2008). Effects of different olive-grove management systems on the organic carbon and nitrogen content of the soil in jaén (spain). Soil and Tillage Research, 98(1), 56–67.

10048

M. Cruz-Ramírez et al. / Expert Systems with Applications 39 (2012) 10038–10048

Cruz-Ramírez, M., Sánchez-Monedero, J., Fernández-Navarro, F., Fernández, J., & Hervás-Martínez, C. (2010). Memetic Pareto differential evolutionary artificial neural networks to determine growth multi-classes in predictive microbiology. Evolutionary Intelligence, 3(3-4), 187–199. Deb, K. (2004). Multi-Objective Optimization using Evolutionary Algorithms. WileyInterscience Series in Systems and Optimization. John Wiley & Sons, LTD. Deb, K., Pratab, A., Agarwal, S., & Meyarivan, T. (2002). A fast and elitist multiobjective genetic algorithm: NSGA2. IEEE Transactions on Evolutionary Computation, 6(2), 182–197. Delenne, C., Durrieu, S., Rabatel, G., & Deshayes, M. (2010). From pixel to vine parcel: A complete methodology for vineyard delineation and characterization using remote-sensing data. Computers and Electronics in Agriculture, 70(1), 78–83. EU Council Regulation 1257, On support for rural development from the European Agricultural Guidance and Guarantee Fund (EAGGF) and amending in repealing certain regulations. Official J. Eur. Communities, Brussels 160 (1999). EU Council Regulation 1259, Establishing common rules for direct support schemes under the common agricultural policy. Official J. Eur. Communities, Brussels 160 (1999). Felton, W. L., Alston, C. L., Haigh, B. M., Nash, P. G., Wicks, G. A., & Hanson, G. E. (2002). Using reflectance sensors in agronomy and weed science. Weed Technology, 16(3), 520–527. Fernández-Caballero, J. C., Hervás-Martínez, C., Martínez-Estudillo, F. J., & Gutiérrez, P. A. (2011). Memetic Pareto Evolutionary Artificial Neural Networks to determine growth/no-growth in predictive microbiology. Applied Soft Computing, 11(1), 534–550. Fernández-Caballero, J. C., Martínez-Estudillo, F. J., Hervás-Martínez, C., & Gutiérrez, P. A. (2010). Sensitivity versus accuracy in multiclass problems using memetic Pareto evolutionary neural networks. IEEE Transactions on Neural Networks, 21(5), 750–770. Fernández, J.C., Hervás, C., Martínez, F.J., Gutiérrez, P.A., Cruz, M. (2009). Memetic Pareto differential evolution for designing artificial neural networks in multiclassification problems using cross-entropy versus sensitivity. In Hybrid artificial intelligence systems, vol. 5572 (pp. 433–441). Springer Berlin/ Heidelberg. Fernández-Navarro, F., Hervás-Martínez, C., García-Alonso, C., & Torres-Jimenez, M. (2011). Determination of relative agrarian technical efficiency by a dynamic over-sampling procedure guided by minimum sensitivity. Expert Systems with Applications, 38(10), 12483–12490. García-Pedrajas, N., Hervás-Martínez, C., & Muñoz-Pérez, J. (2003). COVNET: A cooperative coevolutionary model for evolving artificial neural networks. IEEE Transaction on Neural Networks, 14(3), 575–596. García-Pedrajas, N., Hervás-Martínez, C., & Ortiz-Boyer, D. (2005). Cooperative coevolution of artificial neural network ensembles for pattern classification. IEEE Transactions on Evolutionary Computation, 9(3), 271–302. Girma, K., Mosali, J., Raun, W., Freeman, K., Martin, K., Solie, J., et al. (2005). Identification of optical spectral signatures for detecting cheat and ryegrass in winter wheat. Crop Sci, 45, 477–485. González-Andújar, J. (2009). Expert system for pests, diseases and weeds identification in olive crops. Expert Systems with Applications, 36(2 PART 2), 3278–3283. Gutiérrez, P. A., López-Granados, F., Peña-Barragán, J. M., Jurado-Expósito, M., Gómez-Casero, M. T., & Hervás, C. (2008). Mapping sunflower yield as affected by Ridolfia segetum patches and elevation by applying evolutionary product unit neural networks to remote sensed data. Computers and Electronics in Agriculture, 60(2), 122–132. Gutiérrez, P. A., López-Granados, F., Peña-Barragán, J. M., Jurado-Expósito, M., & Hervás-Martínez, C. (2008). Logistic regression product-unit neural networks for mapping Ridolfia segetum infestations in sunflower crop using multitemporal remote sensed data. Computers and Electronics in Agriculture, 62(2), 293–306. Haykin, S. (1998). Neural Networks: A comprehensive Foundation (2nd edition.). Upper Saddle River, NJ, USA: Prentice Hall. Hengl, T. (2006). Finding the right pixel size. Computers and Electronics in Agriculture, 32, 1283–1298. Hernández, A., Lacasta, C., & Pastor, J. J. (2005). Effects of different management practices on soil conservation and soil water in a rainfed olive orchard. Agricultural Water management, 77, 232–248. Hervás-Martínez, C., Gutiérrez, P., Peña-Barragán, J., Jurado-Expósito, M., & LópezGranados, F. (2010). A logistic radial basis function regression method for discrimination of cover crops in olive orchards. Expert Systems with Applications, 37(12), 8432–8444. Hill, J., Mégier, J., Mehl, W., & degradation, Land (1995). soil erosion and desertification monitoring in mediterranean ecosystems. Remote Sensing Reviews, 12(1), 107–130. Ho, T. K., & Basu, M. (2002). Complexity measures of supervised classification problems. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(3), 289–300. Igel, C., & Hüsken, M. (2003). Empirical evaluation of the improved rprop learning algorithms. Neurocomputing, 50(6), 105–123. Ilonen, J., Kamarainen, J., & Lampinen, J. (2003). Differential evolution training algorithm for feed-forward neural networks. Neural Processing Letters, 17, 93–105.

Jin, Y., & Sendhoff, B. (2008). Pareto-based multiobjective machine learning: An overview and case studies. IEEE Transactions on Systems, Man and Cybernetics Part C: Applications and Reviews, 38(3), 397–415. Jurado-Expósito, M., López-Granados, F., Atenciano, S., García-Torres, L., & GonzálezAndújar, J. L. (2003). Discrimination of weed seedlings, wheat (triticum aestivum) stubble and sunflower (helianthus annuus) by near-infrared reflectance spectroscopy (nirs). Crop Protection, 22(10), 1177–1180. KavdIr, I. (2004). Discrimination of sunflower, weed and soil by artificial neural networks. Computers and Electronics in Agriculture, 44(2), 153–160. Levene, H. (1960). Contributions to Probability and Statistics. London, UK: Stanford University Press. Ch. Essays in Honor of Harold Hotelling. Liaghat, B. S. (2010). A review: The role of remote sensing in precision agriculture. American Journal of Agricultural and Biological Science, 5(1), 50–55. López-Granados, F., Jurado-Expósito, M., Peña-Barragán, J., & García-Torres, L. (2006). Using remote sensing for identification of late-season grass weed patches in wheat. Weed Science, 54, 346–353. MAPA, Spanish Ministry of Agriculture, Fishery and Food (2007). URL ww.mapa.es/ es/estadistica/pags/anuario/introduccion.htm. Martínez-Estudillo, F. J., Gutiérrez, P. A., Hervás-Martínez, C., & Fernández, J. C., (2008). Evolutionary learning by a sensitivity-accuracy approach for multi-class problems, in: Proceedings of the 2008 IEEE Congress on Evolutionary Computation (CEC’08), IEEE Press, Hong Kong, China, pp. 1581–1588. McCoy, R. M. (2005). Field Methods in Remote Sensing. The Guilford Press. Miller, R. G. (1986). Beyond ANOVA. Basics of Applied Statistics. London, UK: John Wiley & Sons. Ou, G. B., & Murphey, Y. L. (2007). Multi-class pattern classification using neural networks. Pattern Recognition, 40(1), 4–18. Panda, S., Hoogenboom, G., & Paz, J. (2009). Distinguishing blueberry bushes from mixed vegetation land use using high resolution satellite imagery and geospatial techniques. Computers and Electronics in Agriculture, 67, 51–58. Peña-Barragán, J., Jurado-Expósito, M., López-Granados, F., Atenciano, S., Sánchez de la Orden, M., García-Ferrer, A., et al. (2004). Assessing land-use in olive groves from erial photographs. Agriculture Ecosystems and Environment, 103, 117–122. Peña-Barragán, J., López-Granados, F., Jurado-Expósito, M., & García-Torres, L. (2006). Spectral discrimination of ridolfia segetum and sunflower as affected by phenological stage. Weed Research, 46(1), 10–21. Peña-Barragán, J. M., López-Granados, F., Jurado-Expósito, M., & García-Torres, L. (2007). Mapping Ridolfia segetum moris patches in sunflower (Helianthus annuus L.) crop using remote sensing. Weed Research, 47(2), 164–172. Price, K. V., Storn, R. M., & Lampinen, J. A. (2005). Differential Evolution. A Practical Approach to Global Optimization. Natural Computing Series. Springer. Provost, F., & Fawcett, T. (1997). Analysis and visualization of the classifier performance: comparison under imprecise class and cost distribution. In Proceedings of the third international conference on knowledge discovery (KDD97) and data mining (pp. 43–88). AAAI Press. Provost, F., & Fawcett, T. (1998). Robust classification system for imprecise environments, in: Proccedings of the fithteenth national conference on artificial intelligence (pp. 706–713). Schlesinger, W. H. (2000). Carbon sequestration in soils: some cautions amidst optimism. Agriculture, Ecosystems & Environment, 82(1-3), 121–127. Schmidt, K., & Skidmore, A. (2003). Spectral discrimination of vegetation types in a coastal wetland. Remote Sensing of Environment, 85(17), 92–108. Snedecor, G. W., & Cochran, W. G. (1989). Statistical Methods (6th Edition). Iowa State University Press. South, S., Qi, J., & Lusch, D. P. (2004). Optimal classification methods for mapping agricultural tillage practices. Remote Sensing of Environment, 91(1), 90–97. Spanish Royal Decree 4, Establecimiento de un régimen de ayudas a la utilización de métodos de producción agraria compatibles con el medio ambiente. Ministerio de Agricultura, Pesca y Alimentación, Madrid, BOE 12 (2001). Storn, R., & Price, K. (1997). Differential evolution. a fast and efficient heuristic for global optimization over continuous spaces. Journal of Global Optimization, 11, 341–359. Teixeira, R. A., Braga, A. P., Takahashi, R. H. C., & Saldanha, R. R. (2000). Improving generalization of mlps with multi-objective optimization. Neurocomputing, 35, 189–194. Thorp, K. R., & Tian, L. F. (2005). A review on remote sensing of weeds in agriculture. Precision Agriculture, 5(5), 477–508. US Department of Agriculture Soil Conservation Service, Soil taxonomy: a basic system of soil classification for making and interpreting soil surveys, Tech. rep., 1975. Witten, I. H., & Frank, E. (2005). Data Mining: Practical Machine Learning Tools and Techniques. Data Management Systems (2nd edition.). Morgan Kaufmann, Elsevier. Yao, X., & Liu, Y. (1997). A new evolutionary system for evolving artificial neural networks. IEEE Transactions on Neural Networks, 8(3), 694–713. Yau, Y., Teo, J., & Anthony, P. (2007). Pareto evolution and co-evolution in cognitive game ai syntesis. Lecture Notes in Computer Sciences, 4403, 227–241. Yu, Q., Gong, P., Clinton, N., Biging, G., Kelly, M., & Schirokauer, D. (2006). Objectbased detailed vegetation classification with airborne high spatial resolution remote sensing imagery. Photogrammetric Engineering and Remote Sensing, 72(7), 799–812. Zhang, G. P. (2000). Neural networks for classification: A survey. IEEE Transactions on Systems, Man and Cybernetics, Part C: Applications and Reviews, 30(4), 451–462.