Assessing fruit-tree crop classification from Landsat-8 time series for the Maipo Valley, Chile

Remote Sensing of Environment 171 (2015) 234–244

Contents lists available at ScienceDirect

Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse

Assessing fruit-tree crop classification from Landsat-8 time series for the Maipo Valley, Chile M.A. Peña a,⁎, A. Brenning b a b

Departamento de Geografía, Universidad Alberto Hurtado, Cienfuegos 41, Santiago, Chile Department of Geography, Friedrich Schiller University, Jena, Löbdergraben 32, 07737 Jena, Germany

a r t i c l e

i n f o

Article history: Received 24 April 2015 Received in revised form 28 September 2015 Accepted 25 October 2015 Available online 3 November 2015 Keywords: Satellite image time series NDVI temporal profile Landsat-8 Crop type classification

a b s t r a c t Satellite image time series (SITS) provide spectral–temporal features that describe phenological changes in vegetation over the growing season, which is expected to facilitate the classification of crop types. While most SITS-based crop type classifications were focused on NDVI (normalized difference vegetation index) temporal profiles, less attention has been paid to using the complete image spectral resolution of the time series. In this work we assessed different approaches to SITS-based classification of four major fruit-tree crops in the Maipo Valley, central Chile, during the 2013–14 growing season. We compared four feature sets from a time series comprised of eight cloud-free Landsat-8 images: the full-band SITS, the NDVI and NDWI (normalized difference water index) temporal profiles, and an image stack with all the feature sets combined. State-of-the-art classifiers (linear discriminant analysis, LDA; random forest; and support vector machine) were applied on each feature set at different training sample sizes (N = 100, 200, 400, 800 and 2291 fields), and classification results were assessed by cross-validation of the misclassification error rate (MER). For all the feature sets overall results were good (MERs ≤ 0.21) although substantially improved classification accuracies were achieved when the full-band SITS was employed (MER 0.14–0.05). Classifications applied on the NDVI temporal profile consistently had the worst performance. For a sample size of 200 fields, LDA using the full-band SITS of image dates 1, 3, 6 and 8 produced the best tradeoff between the number of images and classification accuracy (MER = 0.06), being the green, red, blue and SWIR (short-wave infrared) bands of image date 1 (acquired at the early greenup stage) the most relevant for crop type discrimination. Our results show the importance of considering the complete image spectral resolution for SITS-based crop type classifications as the commonly used NDVI temporal profile and their red and near infrared bands were not found the most significant to discriminate the crop types of interest. Furthermore, in light of the good results obtained, the methodology used here might be transferred to similar agricultural lands cultivated with the same crop types, thus providing a reliable and relatively efficient methodology for creating and updating crop inventories. © 2015 Elsevier Inc. All rights reserved.

1. Introduction Passive optical remote sensing may provide meaningful information for monitoring and managing agricultural lands. A particularly significant application is the spectral classification of crop types from satellite-based multispectral images, which makes it possible to create or update crop inventories based on limited field data. It may furthermore help to direct management practices (e.g., site-specific water and nutrient supply) and assist in yield forecasting, particularly where little historical field data is available. Nonetheless, different vegetation species may show quite similar spectral behavior (or low inter-specific spectral variability), especially during some particular phenological stages and at the typical spectral resolution and bandwidth of satellite⁎ Corresponding author. E-mail addresses: [email protected] (M.A. Peña), [email protected] (A. Brenning).

http://dx.doi.org/10.1016/j.rse.2015.10.029 0034-4257/© 2015 Elsevier Inc. All rights reserved.

based multispectral images (Esch, Metz, Marconcini, & Keil, 2014; Galvão, Epiphanio, Breunig, & Formaggio, 2012). This issue may strongly limit the accuracy of crop types classified from single image dates (Jewell, 1989; Lo, Scarpace, & Lillesand, 1986; Murakami, Ogawa, Ishitsuka, Kumagai, & Saito, 2001; Van Niel & McVicar, 2004). The use of satellite image time series (SITS) arises as a promising approach in remote sensing-based crop type classification. By using SITS, spectral–temporal profiles of the crops of interest are retrieved from a set of multispectral images acquired within a temporal window of interest (usually the entire growing season). This multi-temporal spectral behavior is related to the phenology of the crop, i.e., their seasonal dynamics or intra-annual developmental stages, whose onset and offset dates usually differ from other crop types planted on the same agricultural land. Thus, distinctive spectral–temporal features of each crop type may be extracted from SITS, increasing the chances of classifying them correctly (Jakubauskas, Legates, & Kastens, 2002; Masialeti, Egbert, & Wardlow, 2010; Odenweller & Johnson, 1984;

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

Ozdogan, 2010; Wardlow, Egbert, & Kastens, 2007). Within a growing season the main phenological stages driving the bulk spectral behavior of a crop are: (1) greenup: the date of onset of photosynthetic activity, (2) maturity: the date at which plant green leaf area is maximum, and (3) senescence: the date at which photosynthetic activity and green leaf area begin to rapidly decrease (Zhang et al., 2003). Although there is a growing body of work and practical experience in agricultural land cover classifications using multi-temporal image sets as part of land use and land cover (LULC) mapping and monitoring (see references in Giri, 2012, and Xie, Sha, & Yu, 2008), we center our attention more narrowly on SITS-based classification of individual types of crop. While LULC classifications are often focused on tracking longterm changes of major classes, crop type classification primarily deals with finer thematic granularity and subtle spectral differences. In this case, the spectral separability between crop types is pursued by tracking their phenological differences across a set of images regularly acquired within a given growing season. SITS have been widely used to classify single and multiple crops types mainly by retrieving NDVI (normalized difference vegetation index) temporal profiles as indicators of crop phenology from coarse spatial resolution images such as MODIS (Moderate Resolution Imaging Spectroradiometer) (Arvor et al., 2011; Chenab, Sonb, Changab, & Chenb, 2011; Jakubauskas et al., 2002; Mingwei et al., 2008; Masialeti et al., 2010; Sakamoto et al., 2005; Shao, Lunetta, Ediriwickrema, & Liames, 2010; Sibanda & Murwira, 2012; Sun, Xu, Lin, Zhang, & Mei, 2012; Wardlow & Egbert, 2010; Xavier, Rudorff, Shimabukuro, Berka, & Moreira, 2006; Zhong, Hawkins, Biging, & Gong, 2011) and from finer spatial resolution images such as Landsat or SPOT (Satellite Pour l'Observation de la Terre) (Badhwar, Gargantini, & Redondo, 1987; Jewell, 1989; Murakami et al., 2001; Nguyen, De Bie, Ali, Smaling, & Chu, 2011; Simonneaux et al., 2007; Turker & Arikan, 2005; Zhong, Gong, & Biging, 2014; Zheng, Myint, Thenkabail, & Aggarwal, 2015). The NDVI is constructed from a red band (R) sensitive to foliar chlorophyll content, and a near infrared (NIR) band sensitive to canopy foliage amount, as (NIR − R) / (NIR + R). This spectral index represents in a convenient and synthetic fashion the physiological and structural condition of vegetation, often referred as greenness, which from a temporal perspective relates to vegetation phenology (Odenweller & Johnson, 1984; Jones & Vaughan, 2010). The widespread reliance on MODIS for crop type classification is mainly due to the free availability of 16- and 32-day NDVI image composites, as well as 8-day surface reflectance products that enable to construct 8-day NDVI composites. These composite products strongly increase the availability of cloud-free imagery while reducing the need for radiometric and geometric processing by the end user. Furthermore, they are particularly useful in agricultural lands with relatively large and homogeneous crop fields, considering MODIS spatial resolutions ranging from 250 to 1000 m (Murakami et al., 2001; Ozdogan, 2010; Roy et al., 2014; Wardlow et al., 2007; Zhong et al., 2011). The phenology of a given crop type may change in space and time in response to variation in environmental conditions (e.g., temperature, precipitation, soil type and humidity) and management practices (e.g., fertilization, irrigation, crop rotation). As a result, spectral–temporal profiles extracted from fields corresponding to the same crop type may show some intra-annual differences (within-crop phenological variability) as well as inter-annual differences (early or delayed onset of phenological stages), even at local scales (Arvor et al., 2011; Masialeti et al., 2010; Odenweller & Johnson, 1984; Wardlow et al., 2007). Intra-annual variation in the amplitude of the NDVI temporal profiles of two fields of the same crop type (with phase unchanged) indicates different vegetation conditions. Meanwhile, intra-annual variation in the phase of temporal profiles (with amplitude unchanged) of fields of the same crop type indicates shifts in developmental stages (Jakubauskas et al., 2002). To support and enhance the crop type classification based on NDVI temporal profiles, reference field data about crop phenology have been used (Sakamoto et al., 2005; Shao et al.,

235

2010; Sibanda & Murwira, 2012; Zheng et al., 2015; Zhong et al., 2011, 2014), although the success of these approaches may strongly depend on the quality and quantity of such data (Masialeti et al., 2010). Although the use of the complete spectral resolution of SITS in the realm of agricultural land cover classifications is not a new approach (see early references in Lo et al., 1986), quantitative assessments of its potential are still lacking, and it is necessary to update earlier results considering the state of the art in data mining and statistical learning. Full-band Landsat-7 time series were previously used by Van Niel and McVicar (2004) to classify four cereal crop types in the Coleambally Irrigation Area, Australia. In their study the SITS only included image dates that previously yielded the highest classification accuracies for each single crop in an iterative multi-date classification process of the whole time series. So, even though they did not utilize all the images of the SITS, their results showed per-class accuracies mostly above 80%. The scarce attempts to classify crop types by full-band SITS may be attributed to some extent to the general assumption that using the complete spectral resolution of a multispectral image (e.g., from blue to short-wave infrared, SWIR, bands) may provide redundant or little additional information about crop condition compared with a synthetic measure such as the commonly used NDVI. However, the underlying assumption of an autocorrelation between band absorption and reflection features may not necessarily hold true for all vegetation types. For instance, throughout the phenological stages of some vegetation targets it has been found that variations in foliar chlorophyll and water content (sensitive at SWIR wavelengths) may not necessarily be proportional due to the numerous environmental factors affecting their physiological and structural properties (Ceccato, Flasse, & Gregoire, 2002; Ceccato, Flasse, Tarantola, Jacquemoud, & Grégoire, 2001). On the other hand, some image bands are usually avoided for many applications because they may contain a large amount of noise (e.g., blue and SWIR bands) even after atmospheric correction, yielding unreliable spectral signals (Jones & Vaughan, 2010; Peña & Altmann, 2009; Roberto et al., 2012). However, new satellite remote sensing technologies provide improved signal-to-noise ratios, and the capabilities of novel atmospheric correction algorithms to remove extraneous path radiance under a wide range of environmental conditions have been continuously enhanced (Liang, Li, & Wang, 2012; Roy et al., 2014). These advances have improved the quality of bands traditionally considered noisy. In this paper we assessed the SITS-based classification of four major fruit-tree crop types in a study area located in central Chile during the 2013–14 growing season. We used all the cloud-free Landsat-8 images available for the season of interest. Three state-of-the-art statisticallearning classifiers were selected and applied at different training sample sizes on the full-band SITS, as well as on NDVI and NDWI (normalized difference water index) temporal profiles, and their accuracies were then assessed. The methods to be employed for the study area must consider the relatively small field size of the crops of interest, which precludes the use of MODIS products, as well as the absence of reliable auxiliary data on the phenology of the crops of interest. We expect to find optimal combinations of feature sets, classification techniques and training sample sizes to classify the crop types of interest, which may be transferable to nearby similar agricultural lands. This could result in the construction of less costly and more frequent crop inventories, which are currently updated in our study area at least every four years mostly by field campaigns. 2. Materials and methods 2.1. Study area The study area is located in the Maipo River basin (15,157 km2) in the Metropolitan Region of Chile south of the capital Santiago. The Mediterranean climate in which the basin is located is characterized by hot, dry summers and cool, wet winters, which results in hydrological regimes controlled by snowmelt in the Andean headwaters in spring

236

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

and rain in winter. Agriculture is its main land use, although native sclerophyllic vegetation still remains as forest and shrubland formations mainly in steep terrain. Most of its agricultural land is occupied by fruit

trees (36%), vegetables (17%) and cereals (11%) (ODEPA (Oficina de Estudios y Políticas Agrarias), 2014). The study area administratively comprises nine counties situated along the middle reaches of the

Fig. 1. Distribution of the tree-fruit crops within the study area (top), located in the Maipo Valley in central Chile. Bottom: Maipo River basin (thick line) and administrative boundaries of the Metropolitan Region (thin line) highlighting the study area's counties in gray.

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

Maipo River: Padre Hurtado, Peñaflor, Calera de Tango, Talagante, Isla de Maipo, Buin, Paine and Pirque (Fig. 1). Together, these counties reach an area of 1878 km2, mostly cultivated with four tree-fruit species: walnut (Juglans regia L.), table grape (Vitis vinifera L.), almond (Prunus dulcis Mill.) and European plum (Prunus domestica L.) (Table 1), heterogeneously distributed in the form of relatively small fields (N = 3120; mean size 44,000 m2).Greenup stage of these species broadly starts during October (southern spring), while senescence stage starts in early May (southern autumn). Inter- and intra-specific onset dates of these stages show shifts of several days according to the variety or management practice applied on the crop. The mentioned species were the target crops of this study. 2.2. Satellite image time series All cloud-free Landsat-8 images acquired from early October, 2013 to late April, 2014 (every 16 days, the revisit time of the satellite) were selected and downloaded from the Global Visualization Viewer (http://glovis.usgs.gov/) of the USGS (United States Geological Survey). The list of image acquisition dates used in this study is presented in Table 2. Landsat-8 satellite was launched on February 11th, 2013, carrying the multispectral pushbroom sensor OLI (Operational Land Imager), which incorporates numerous technical improvements in regard with their predecessor instrument, ETM + (Enhanced Thematic Mapper Plus), on board Landsat-7. Particularly noteworthy are the higher signal-to-noise ratio (increased by a factor of at least eight in comparison with ETM +) and radiometric resolution (12 bits rather than 8 bits used by ETM+) as well as the improved geometric fidelity (by an on-board global positioning system, GPS) and the better combination of pre-launch, on-board and vicarious calibration procedures (Roy et al., 2014). In this study, the panchromatic band 8 (0.5–0.68 μm) was not used because of its broad spectral width, and bands 1 (0.43–0.45 μm) and 9 (1.36–1.39 μm) were omitted because they were designed for coastal water and atmospheric aerosol applications (Roy et al., 2014). Meanwhile, thermal bands 10 (10.60– 11.19 μm) and 11 (11.50–12.51 μm) acquired by the TIRS (Thermal Infrared Sensor) instrument on board this satellite were also discarded, mainly because their relatively large original image pixel size of 100 m limits their utility in the classification of the crop types targeted in this study. Table 3 lists the main technical properties of the Landsat-8 bands used in this study. 2.3. Image pre-processing Landsat-8 images are originally distributed as Level 1 terraincorrected (L1T) products. The procedures applied on these images include a systematic geometric correction, precision correction assisted by ground control chips, and the use of a digital elevation model (DEM) to correct parallax error due to terrain relief. The output coordinate system assigned is the Universal Transverse Mercator (UTM) with World Geodetic System 1984 (WGS84) datum, and the circular geolocation error is b12 m at the 90% confidence level (Roy et al., 2014). As products subject to the same geometric correction procedures, L1T images Table 1 Area occupied by the fruit-tree species of interest and the rest of the tree fruits cultivated in the study area. Tree-fruit crop

Walnut Table grape Almond European plum Others Total

Area Km2

%

55.35 42.21 21.64 18.66 47.7 182.9

29.2 24.6 10.1 10.1 26.1 100

237

Table 2 Acquisition dates of the Landsat-8 images used in this study. Image

Season in the study area

#

Acquisition date

1 2 3 4 5 6 7 8

October 5, 2013 October 21, 2013 November 22, 2013 January 9, 2014 January 25, 2014 February 10, 2014 February 26, 2014 April 15, 2014

Spring Spring Spring Summer Summer Summer Summer Autumn

acquired at different dates over the same path/row satisfy the spatial match required for time series analysis. In spite of this, image geometric coregistration was visually assessed by checking the spatial match of some randomly selected field boundaries within the study area across the time series. The spatial shifts in the target boundaries were found to be far below half the pixel size and therefore negligible for our purposes. The conversion of image digital numbers to absolute at-sensor radiances was automatically performed by applying to each band the OLI radiometric calibration parameters available in the ENVI© (Environment for Visualizing Images) 5.0.3 software (Exelis Visual Information Solutions, Inc., Boulder, USA). Afterwards, images were atmospherically corrected using its FLAASH (Fast Line-of-Sight Atmospheric Analysis of Spectral Hypercubes) module. This is a MODTRAN (Moderate Resolution Atmospheric Transmission) based algorithm that aims to reduce the extraneous path radiance affecting the pixel's at-sensor radiometry (i.e., adjacency and haze effects), while modeling the at-surface irradiance. To do that, the algorithm requires a set of user-defined scene and study area parameters (image acquisition date and time, geographic location, terrain elevation, visibility, atmosphere and aerosol type, among others). The output of this procedure is an image in apparent surface reflectance pixel units (resulting from rationing at-surface irradiance to at-surface radiance). 2.4. Training stage Fields required to train the classification of the crops of interest were taken from the Fruit Cadastre of the study area, carried out by Chile's Oficina de Estudios y Políticas Agrarias (ODEPA). This is an online digital database (http://icet.odepa.cl/) that contains crop field boundaries as polygons updated by field campaigns to the 2013–2014 season. The training sites were manually selected from within-field areas in which a continuous vegetation cover was observed. To refine this task, we first removed too small fields (≤22,500 m2 or ≤5 × 5 pixels) or excessively narrow fields from the original database because they are mainly comprised of mixed edge (Landsat-8) pixels useless as training sites and detrimental for the subsequent classification accuracy assessments. As a result, the original number of polygons per crop type (walnut: 1,121, table grape: 1,106, almond: 477, European plum: 416) was somewhat reduced (796, 889, 308 and 298 fields, respectively; Table 4). Table 3 Main technical characteristics of the Landsat-8 OLI (Operational Land Imager) image bands used in this study. Band #

Spectral region

Spectral width (μm)

Spatial resolution (m)

Radiometric resolution (bits)

2 3 4 5 6 7

Blue Green Red Near infrared Shortwave infrared Shortwave infrared

0.45–0.51 0.53–0.59 0.64–0.67 0.85–0.88 1.57–1.65 2.11–2.19

30 m

12

238

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

Table 4 Number and size of the crop fields used in this study. Crop type

Field #

Walnut Table grape Almond European plum Total

Size

Polygons

Pixels

Mean (m2)

Standard deviation (m2)

796 889 308 298 2291

8310 7374 3468 3533 22,685

63,525 43,677 63,173 57,096 59,940

50,531 21,982 41,402 35,726 39,442

2.5. Classification and validation stages Crop type classification was carried out for four feature sets with and without full-band SITS data, using three different classifiers and five different training sample sizes in order to identify efficient classification procedures using a methodology similar to Brenning, Long, and Fieguth (2012). Backward and forward image selection was furthermore carried out for the best-performing classifier and feature set in order to determine whether a reduced number of image dates can achieve competitive results. The following feature sets were derived from the Landsat-8 time series: (1) NDVI temporal profile (i.e., eight NDVI values as predictors), (2) NDWI temporal profile, (3) full-band SITS (i.e., a total of 8 × 6 predictors) and (4) an image stack with all the previously mentioned feature sets combined (64 predictors). NDVI was calculated as mentioned in the introduction chapter of this work, while NDWI was calculated using the same formula, but replacing the red band with the SWIR band comprising 1.57–1.65 μm (Gao, 1996). Since different classification techniques differ in their ability to leverage nonlinear or otherwise complex relationships between features and crop type, we used three different statistical and machine learning techniques that are representative of the state–of-the-art in image classification (Brenning et al., 2012): linear discriminant analysis (LDA), random forest (RF) and support vector machine (SVM). While LDA is less flexible than the other methods as it is only able to separate the classes linearly, it is less prone to overfitting and computationally very efficient. RF is a bootstrap-aggregated tree-based classifier that is particularly suited to detect nonlinear interactions among predictors (Breiman, 2001). SVM is also a relatively novel technique that is based on nonlinear kernel transformations of the predictors, projecting them into a higher-dimensional feature space. In this space, an optimal separating hyperplane is computed (Moguerza & Muñoz, 2006). We refer the reader to James, Witten, Hastie, and Tibshirani (2013) for further details on the classification techniques used. RF was applied with its default settings in the R implementation used in this study (500 trees, square root of no. of predictors selected randomly; Liaw & Wiener, 2002). SVM in its implementation in the ‘e1071’ package (Meyer, Dimitriadou, Hornik, Weingessel, & Leisch, 2014) was used for C-classification with radial basis function kernels. The costs parameter C and the kernel bandwidth parameter γ were tuned using an inner cross-validation grid search on each training data set, resulting in a computationally highly expensive procedure. In order to reduce noise and obtain an unambiguous class prediction for each field polygon, pixel-level predictions were filtered by picking each field's predicted majority class. All analyses were performed in the data analysis software R (version 3.1.1; R Core Team, 2014). Classification accuracy for each classifier and feature set was assessed for five training sample sizes (N = 100, 200, 400 and 800 randomly selected fields, and all 2291 fields) using cross-validation estimation of the misclassification error rate (MER). To do that, MERs were estimated by cross-validation. MER is defined as the total proportion of objects in a data set that a classifier assigns to the wrong class. Cross-validation is a resampling-based bias-reduced technique for the

estimation of predictive performance. It repeatedly generates training and test data sets from the available set of fields with known class membership, effectively using the entire data set for performance estimation. k-fold cross-validation consists of partitioning the set of fields into k equally-sized subsets; a classifier is trained on all but one of these subsets and evaluated on the remaining one. To be independent of a particular partitioning, the whole procedure is repeated r times (r-repeated k-fold cross-validation). Error rates are averaged over all test data sets. In this study, we used k = 10 and r = 100 with the exception of SVM where only r = 20 was feasible due to the high computational complexity of the additional inner cross-validation step on each of the k × r training data sets. Cross-validation was performed on the field level, not on the pixel level, as the latter would result in a contamination of test data sets with pixels that are pseudoreplications of training pixels, which would yield over-optimistic accuracy estimates and encourage over-fitting (Brenning, Kaden, & Itzerott, 2006). Field-level cross-validation was carried out using the ‘sperrorest’ package in R (Brenning, 2012). Considering a 5% level of significance and a Bonferroni correction to control the overall error rate of a family of tests, an estimate for the critical difference in cross-validated performance estimates was determined to be 0.007 (i.e., 0.7 percentage points) based on 100 crossvalidation repetitions and 0.017 based on 20 cross-validation repetitions. This critical difference was calculated for a two-sample t-test and using the largest observed standard deviation of cross-validated MERs to obtain a conservative significance threshold. In the Bonferroni correction for multiple comparisons we accounted for all possible tests between situations where only one of the factors changed, e.g. LDA versus SVM performance for the same feature set and training sample size. Using the best feature set and classifier from the performance comparison, forward and backward image selection were performed in order to explore if similar performances can be achieved with a smaller set of image dates. In forward (backward) selection, predictors corresponding to one image date were added to (removed from) the feature set, starting with an empty feature set (the full feature set). In each step, the image combination that achieved the best performance was selected. Image selection was performed for a training sample size of 200 fields only based on intermediate results from the previous steps. In order to assess which of the features within the resulting feature set contributed most to the classifier's predictive capability, a permutation-based variable accuracy importance measure was furthermore calculated (Brenning et al., 2012). In this approach, one predictor at a time is permuted or “messed up” in order to calculate how much the MER increases compared to predictions using undisturbed data. This technique was embedded in the field-level spatial cross-validation using ‘sperrorest’ (Brenning, 2012). For each variable, permutation was repeated 100 times within each cross-validation repetition and fold. 3. Results 3.1. Overall classification performance Regardless of the feature set and training sample size used, all the classifications achieved good accuracies (MER ≤ 0.21; Fig. 2 and Table 5). For all the feature sets, classification accuracies consistently increased as the training sample size increased from 100 to all 2291 fields (MER differences of 0.02–0.1, depending on the classifier; mean differences N0.007 are significant at the 5% level while accounting for multiplicity of tests). However, performance improvements with training sample sizes increasing above 200 fields were relatively small in full-band SITS classification and when using all feature sets combined (MER differences of 0.01–0.05 compared to a training sample size of 200 fields). Classification performances obtained with the full-band SITS and all the feature sets combined were almost identical (MER differences ≤0.01

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

239

Fig. 2. Cross-validated MERs for fruit-tree fields using (a) image stacks and (b) only NDVI/NDWI temporal profiles for different classifiers and training sample sizes. Mean differences of N0.007 (SVM: N0.017) should be considered significant at the 5% level while accounting for multiple comparisons.

for same classifier and training sample size), providing higher accuracies across the five training sample sizes than classifications using only NDVI or NDWI temporal profiles (MER differences of 0.05–0.1, depending on the classifier and the training sample size used). LDA using the full-band SITS was the best-performing classifier at each of the training sample sizes, providing the maximum overall accuracy value at all 2291 fields (MER = 0.05). RF using the full-band SITS started with poorer classification performances for smaller training sample sizes but caught up with LDA when using the largest training sample size. SVM using the full-band SITS yielded classification accuracies between LDA and RF but came at a high computational cost, making its application impractical for the larger training sample sizes. Classifications based on NDVI temporal profile consistently had the worst performance, being LDA the classifier that produced the lowest overall accuracies at each of the training sample sizes. By comparison, NDWI temporal profile consistently achieved a somewhat better performance (up to 3 percentage points). RF trained on all 2291 fields and using NDWI temporal profile produced the highest accuracy among index-based classifications (MER = 0.1). Based on these results, we consider that LDA using (at least) the full-band SITS provided the most satisfying classification performance overall, taking into account also its computational efficiency and relative insensitivity to training

Table 5 Cross-validated MERs provided by three classifiers applied on four feature sets at five training sample sizes. Differences of N0.007 (SVM: N0.017) are significant at the 5% level while accounting for multiplicity of comparisons. Classifier

Feature set

Misclassification error rate

sample size. With this classifier, nearly optimal results were achieved with a training sample of only 200 fields (MER = 0.06). 3.2. Classification performance using linear discriminant analysis When used with the full-band SITS or all the feature sets combined, LDA had the best overall performance (MERs 0.08–0.05; Fig. 2 and Table 5). Classification accuracies between both features sets were about the same (MER b 0.01), showing a slight increasing as the training sample size increased from 100 to all 2291 fields (MER difference of 0.03). Conversely, LDA had the worst overall performance when used with index-based feature sets, particularly with NDVI temporal profile (MERs 0.21–0.17). For both index-based feature sets, classification accuracies showed a slight increase as the training sample size increased from 100 to all 2291 fields (MER difference of 0.04). 3.3. Classification performance using support vector machine Classification accuracies obtained from SVM were somewhat lower than those achieved with LDA when used with the full-band SITS or all the features combined (Fig. 2 and Table 5). Both feature sets provided the highest classification accuracies (MERs 0.06–0.01), showing a very similar performance (MER ≤ 0.01) and a slight increasing as the training sample size increased from 100 to 400 fields (MER difference of 0.05). For comparison, index-based feature sets and especially NDVI temporal profile yielded poorer performances (MERs 0.20–0.13). Classification accuracies obtained from NDVI and NDWI temporal profiles slightly increased as the training sample size increased from 100 to 400 fields (MER difference between 0.05 and 0.04, respectively).

Training sample size

Linear discriminant analysis

Random forest

Support vector machine

NDVI NDWI Full-band SITS All feature sets NDVI NDWI Full-band SITS All feature sets NDVI NDWI Full-band SITS All feature sets

100

200

400

800

All 2291 fields

0.21 0.18 0.08 0.08 0.20 0.18 0.13 0.14 0.20 0.17 0.11 0.10

0.19 0.17 0.06 0.06 0.17 0.15 0.10 0.11 0.17 0.14 0.08 0.08

0.18 0.16 0.06 0.06 0.14 0.13 0.08 0.09 0.15 0.13 0.06 0.06

0.17 0.15 0.05 0.05 0.12 0.11 0.06 0.07 – – – –

0.17 0.14 0.05 0.05 0.10 0.10 0.05 0.06 – – – –

3.4. Classification performance using random forest RF had the worst overall performance when used in classifications that included full-band SITS; it did, however, exploit index-based feature sets more effectively than LDA especially for large sample sizes (Fig. 2 and Table 5). Classification accuracies obtained from full-band SITS and all the feature sets combined were also in this case higher than those retrieved by index-based feature sets. Full-band SITS classification accuracies at all training sample sizes were only one percentage point higher than those produced by combining all the feature sets (MERs 0.13–0.05). Both feature sets showed a more substantial performance improvement with training sample size than LDA (MER difference of 0.08).

240

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

3.5. Optimal image selection and band importance A forward image selection based on the best-performing classifier and feature set (LDA using full-band SITS feature set) at a training sample size of 200 fields showed that the combination of only four images: 1, 3, 6 and 8 (see acquisition dates in Table 1), achieved a nearly optimal MER (0.06), arising as the best tradeoff between number of images and classification accuracy (Fig. 3; backward selection not shown because of nearly identical results). The map resulting from this classification is shown in Fig. 4. Classification accuracies increased only slightly when the remaining images were progressively added to the model. Classification accuracies started to decrease more substantially when three or fewer images were considered. The image acquired on April 15th, 2014 (image no. 8, Table 1) provided the best single date-based classification accuracy (MER = 0.19, Fig. 3), which equals the result yielded by this classifier when applied on the NDVI temporal profile of all eight images at the same training sample size. A further examination of variable importance in LDA using full-band SITS of image dates 1, 3, 6 and 8 and a training sample size of 200 fields revealed that the most important bands belonged to image date 1, with the green, red, blue and SWIR bands leading the ranking (Fig. 5). Meanwhile, the less important bands were mostly concentrated in image dates 6 and 8, with a mixed combination of visible to SWIR bands from both dates contributing the least to predictive performance. The cross-validated confusion matrix of LDA using full-band SITS from image dates 1, 3, 6 and 8 and a sample size of 200 fields yielded very good overall and per-class results, with user's and producer's accuracies showing only little variations among crop types (Table 6). The highest commission and omission errors (i.e, the complement of user and producer accuracies, respectively) were obtained for European plum, while both error measures were very low for table grape. 4. Discussion 4.1. Classification performance and feature sets The three classifiers used in this work had a similar overall performance, providing classification accuracies that in the realm of remote sensing-based classifications are commonly considered as good. Pal (2005) obtained good overall accuracies (88%) and per-class accuracies (≥81%) from SVM and RF applied on a Landsat image to classify seven crop types, but did not compare these computationally intensive statistical learning methods with simpler linear discriminant techniques. Pal and Mather (2005) found that SVM had a slightly better performance than a neural network classifier when applied on an airborne hyperspectral image to classify eight land cover types, including some crop types (overall accuracies of 93.6 and 86%, respectively). The fairly consistent classification accuracies found in our work were useful for the pursued analysis as instead of comparing the performance

of a set of classifiers we rather focused on assessing the performance of different SITS-derived feature sets to classify the crop types of interest. We selected the classifiers simply based on their representativeness as state-of–the-art machine-learning techniques. During the last years, these techniques have gained increasing interest in the remote sensing community because of their potential to deal with high-dimensional and complex relationships that are often required to discriminate spectrally similar targets (Gislason et al., 2006; Mountrakis, Im, & Ogole, 2011; Szuster, Chen, & Borger, 2011). Nevertheless, in our study linear techniques outperformed nonlinear SVM and RF classifiers, which underlines the utility of linear techniques and their variants (e.g., with lasso penalties) for computationally efficient processing of remote sensing data with often competitive predictive performances (e.g., Zandler, Brenning, & Samimi, 2015). A particular strength conferred to machine learning classifiers has been their relatively low sensitivity to training sample size reduction (Huang, Davis, & Townshend, 2002; Mathur & Foody, 2008; Mountrakis et al., 2011). Sample size effects have rarely been addressed in crop or similar vegetation land cover classifications using SITS. Zheng et al. (2015) classified nine single and double crop types applying SVM on a NDVI temporal profile constructed with 22 Landsat-5 and − 7 images acquired throughout one year. They compared classification accuracies using training sites selected by the aid of auxiliary data (n = 44) and by a stratified random sampling technique (n = 174). Although in some cases producer's and user's accuracies showed significant variations, overall accuracies remained stable (90 and 86%, respectively). Similarly, Rodriguez-Galiano, Ghimire, Rogan, Chica-Olmo, and Rigol-Sanchez (2012) classified 14 land covers (including some crops and other vegetation targets) applying RF on two Landsat-5 images acquired in spring and summer. They found that the reduction of training sample size did not have a significant effect on the classifier's accuracy until reaching a 50% data reduction. Our results agree with these findings, as regardless of the feature set and classifier used, classification accuracies did not significantly diminish when the training sample size decreased even to 200 fields. Unlike our work, most of the previously published SITS-based crop type classifications have relied on the use of MODIS-derived temporal profiles of greenness indices like the NDVI. These classifications have been commonly applied at regional scales on relatively large graminaceous crop fields. Using Maximum Likelihood algorithm, Van Niel and McVicar (2004) classified four crop types with an overall accuracy of 95.8%, and Arvor et al. (2011) obtained an overall accuracy of 74% in the classification of three crops types planted in single or double cropping systems. With decision trees, Wardlow and Egbert (2008) classified four crop types with an overall accuracy of 84%, and Zhong et al. (2011) obtained overall accuracies of 74.6 and 78.6% in the classification of two areas with eight and 12 crop types, respectively. By using Fourier analysis Jakubauskas et al. (2002) classified three crop types with an overall accuracy of 52%, and Mingwei et al. (2008) obtained a

Fig. 3. MERs obtained in a forward image selection based on the best-performing classifier and feature set: LDA applied on the full-band SITS, at a training sample size of 200 fields.

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

241

Fig. 4. Crop type classification using LDA on the full-band SITS comprised by the image dates 1, 3, 6 and 8 at a training sample size of 200 fields.

RMSE (root mean square error) of 48.11 and 82 km2 in the area estimation of cotton and maize, respectively. By using temporal unmixing and independent component analysis Ozdogan (2010) obtained RMSE values between 15 and 30% in the area estimation of three crops types

subject to different management practices. In spite of the good results reported in these works, our study showed that substantial information is lost when a SITS-based crop type classification is performed by an index-based temporal profile rather than by the full-band SITS, having

Fig. 5. Permutation-based variable importance of each band in terms of the mean increase in MER for image dates 1, 3, 6 and 8 previously selected for LDA with a training sample size of 200 fields. Y axis labels identify the image number (first digit) and the band number (second number in variable name).

242

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

Table 6 Confusion matrix of crop types classified by LDA applied on the full-band SITS comprised by the image dates 1, 3, 6 and 8 at a training sample size of 200 fields. Numbers in the table are pixel counts (table margins: fractions) averaged over 100 cross-validation repetitions.

Walnut Table grape European plum Almond Producer's accuracy Overall accuracy

Walnut

Table grape

European plum

Almond

User's accuracy

7902 80 191 137 0.95

177 7130 41 27 0.97

295 13 3052 173 0.86

131 25 149 3163 0.91

0.92 0.98 0.89 0.9

Overall accuracy

0.94

a detrimental impact on the classification accuracies. As a matter of fact, we found that even the best-performing single-image classifier was able to achieve the same performance as a classifier using NDVI temporal profile based on all eight images. Our study area comprised relatively small fields of fruit-tree crops which technically preclude the use of the typical MODIS-derived NDVI composite products. Instead, we decided to use a Landsat-8 time series for a given growing season to examine if the full-band SITS or the temporal profile of a wetness spectral index (NDWI) would yield improved predictive performances. Even when the acquisition dates of the images comprising the time series did not necessarily match with critical phenological dates of the crops of interest and some images had to be missed due to cloud cover, the results obtained here are quite promising. Classification accuracies were the highest when the full-band SITS was employed, regardless of the classifier and training sample size used (MER 0.14–0.05). The best classification accuracies were obtained with the simple and less flexible LDA classifier, which is consistent with the results obtained by Brenning et al. (2006), who used Landsat time series from 1996 and 2000 to classify seven and nine crop types, respectively, in Havelland, Germany. In that work, classification accuracy obtained for LDA (mean MER = 20.1) was among the best results obtained for the set of classifiers used, which included SVM. As in our work, this classifier also stood out by achieving very good classification with a small training sample size of 200 fields. Moreover, in our work classifications performed using the NDWI temporal profile provided slightly better results than using the NDVI temporal profile. This result highlights the potential that a wetness spectral index may have to track changes in vegetation phenology, which for SITS-based crop type classifications has been overlooked. NDVI temporal profiles of graminaceous crop types have shown good separabilities due to the clear differences in some of the phenological stages of these crops, resulting in good classification accuracies. As annual crops, first interclass differences in the NDVI temporal profiles of gramineous may be found as a result of the different times at which some of them are planted (e.g., winter wheat versus summer crops). Budburst and green-up times therefore offer good opportunities to discriminate them. Moreover, many of them are cultivated in annual rotations or double cropping systems. Harvesting and regrowth therefore improve the ability to discriminate them by NDVI temporal profiles based on their distinctive bimodal patterns (Arvor et al., 2011; Masialeti et al., 2010; Mingwei et al., 2008; Murakami et al., 2001; Ozdogan, 2010; Wardlow et al., 2007; Zheng et al., 2015). Although studies addressing SITS-based classification of fruit-tree crop types are still rare (e.g., Simonneaux et al., 2007; Zhong et al., 2011), it is possible to argue that classification based on NDVI temporal profiles may be more challenging than for graminaceous crops, because as perennial summer crops they may exhibit similar interclass timings in their phenological stages. Zhong et al. (2011) found similar NDVI temporal profiles for almonds, pistachio and vineyard crops, with increasing values during spring (green-up), constant high values throughout the summer (maturity), and a gradual decrease in the autumn (senescence and defoliation). Further complications may arise if the temporal resolution of

the NDVI temporal profiles is too broad to clearly distinguish the interclass differences. 4.2. Optimal image selection and band importance Previous studies have found a decreasing tendency in classification accuracies when data dimensionality increases, a widely-known issue in remote sensing that refers the so-called Hughes phenomenon: the reduction of the predictive power as the data dimensionality increases. Pal and Mather (2003, 2005) classified eight land covers applying several parametric and non-parametric classifiers (including RF, SVM and an artificial neural network classifier) on an airborne hyperspectral image. They examined classification accuracies for a fixed training sample size by progressively increasing the band number and found that values declined or almost unchanged when the band number used in all the classifiers exceeded a few tens. Given the high temporal dimensionality of the SITS defined in our work we addressed the question of how many images comprising the defined SITS are enough to achieve an optimal classification result. Taking into account the classifier and the feature set that yielded the best accuracy at a training sample size of 200 fields we found that a combination of only four images arise as the best tradeoff between number of images and classification accuracy. This result suggests that temporal resolution of the SITS is not so critical to achieve a reliable classification of the crop types of interest. This finding contrasts with results of Ozdogan (2010), who found that the ability to classify crop types smaller than the image pixels of a MODIS time series (500 m) by applying a temporal unmixing technique (based on independent component analysis) was dependent on temporal resolution more than spectral resolution. In his work, winter wheat, corn and soybean crops subject to different management practices were classified by using only the red or NIR reflectance or their arithmetic combinations in the form of simple ratio (SR = NIR/red) and NDVI. Classification accuracy results were assessed in terms of per-crop area estimation by RMSE, which produced values around 15 and 30% depending on the crop type. As previously discussed, in cases where timings of the phenological stages of a given set of crops markedly differ between them it is likely to expect a relatively more important role in the image temporal resolution. The optimal image acquisition dates found in our work are located toward the extremes of the considered growing season, i.e., the greenup and senescence phenological stages of the crops of interest. As stated before, during these stages the chances of discriminate crop types may increase as a result of its different growing and senescing timings, which are expressed as different intensities in the spectral signals of the wavelengths sensitive to the vegetation state (Jones & Vaughan, 2010). For instance, Arvor et al. (2011); Wardlow et al. (2007) and Wardlow and Egbert (2008), found some of the best separabilities between the NDVI temporal profiles of different crop types in spring, as a result of the different crop planting dates, resulting in different interclass foliage amounts and hence, different interclass NDVI signals. However, these characteristic signals may not be necessarily the strongest at greenness wavelengths (i.e., red and NIR bands). As a matter of fact, for the previously mentioned image combination we found some of the most meaningful bands in terms of crop type discrimination at the green, red, blue and SWIR (1.57–1.65 μm) bands of image date 1 (acquired at the early greenup stage: October 5th, 2013, Table 1), while the NIR band of that image was ranked far away from the top list (Fig. 5). As mentioned in the introduction of this work, the visible and SWIR bands are well-known for their sensitive to the bulk pigment and water content in the foliage, respectively. Regarding the former, green band has been used to construct the GNDVI (green normalized difference vegetation index), whose formulation resembles the NDVI but replacing the red band with the green band (Gitelson, Kaufman, & Merzlyak, 1996). The reason to do that is because the green band saturates at higher foliar chlorophyll concentrations than the red band. According to this, foliar chlorophyll content of the crops of interest

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

might be relatively high at the greenup stage, explaining the best performance of the green band. Results obtained in this work should be compared with SITS-based classifications of the same crop types subject to similar and different agrological settings throughout different years in order to assess the transferability of the methods employed here to a wider range of the crop types of interest. Regarding this, attempts to compare NDVI temporal profiles of graminaceous crop types from different years have shown to be consistent if scene dependent factors are controlled (e.g., using the same type of images subject to the same preprocessing procedures) (Masialeti et al., 2010). 5. Conclusions We addressed the SITS-based classification of tree-fruits crops by adopting as feature sets not only the widely used NDVI temporal profile but also a NDWI temporal profile and the complete spectral resolution of a Landsat-8 time series corresponding to the 2013–2014 growing season of the crops of interest. For all the feature sets the overall results were good (MER ≤ 0.21), but showed that classification accuracies are substantially improved – typically by 5–10 percentage points – if fullband SITS are used (MER 0.14–0.05). Meanwhile, the worst overall classification accuracies were obtained by using the NDVI temporal profile. For a training sample size of 200 fields, LDA using the full-band SITS of image dates 1, 3, 6 and 8 produced the best tradeoff between the number of images and classification accuracy (MER = 0.06), being the green, red, blue and SWIR bands of image date 1 (acquired at the early greenup stage), the most relevant for crop type discrimination. These results emphasize the usefulness that the complete image spectral resolution may have for SITS-based crop type classifications, a methodological approach that is often overlooked in the scientific literature. Moreover, NDWI temporal profile performed slightly better than NDVI temporal profile, which highlights a potential has also been overlooked in SITS-based crop type classification. In light of the encouraging classification accuracies obtained in this work, the proposed methodology constitutes a reliable and less costly approach to the design and updating of crop inventories compared with costly field mapping. Acknowledgments The use of facilities of the Shared Hierarchical Academic Research Computing Network (SHARCNET; www.sharcnet.ca) and Compute/ Calcul Canada is acknowledged. References Arvor, D., Jonathan, M., Simões, M., Meirelles, P., Dubreuil, V., & Durieux, L. (2011). Classification of MODIS EVI time series for crop mapping in the state of Mato Grosso, Brazil. International Journal of Remote Sensing, 32(22), 7821–7847. Badhwar, G.D., Gargantini, C.E., & Redondo, F.V. (1987). LANDSAT classification of Argentina summer crops. Remote Sensing of Environment, 21, 111–117. Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. Brenning, A. (2012). Spatial cross-validation and bootstrap for the assessment of prediction rules in remote sensing: The R package ‘sperrorest’. 2012 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 23–27 July 2012 (pp. 5372–5375). Brenning, A., Kaden, K., & Itzerott, S. (2006). Comparing classifiers for crop identification based on multitemporal Landsat TM/ETM data. Proceedings, Second Workshop of the EARSeL Special Interest Group on Remote Sensing of Land Use and Land Cover, Bonn, 28–30 September 2006 (pp. 64–71). Brenning, A., Long, S., & Fieguth, P. (2012). Detecting rock glacier flow structures using Gabor filters and IKONOS imagery. Remote Sensing of Environment, 125, 227–237. Ceccato, P., Flasse, S., & Gregoire, J. (2002). Designing a spectral index to estimate vegetation water content from remote sensing data: Part 2. Validation and applications. Remote Sensing of Environment, 82, 198–207. Ceccato, P., Flasse, S., Tarantola, S., Jacquemoud, S., & Grégoire, J. -M. (2001). Detecting vegetation leaf water content using reflectance in the optical domain. Remote Sensing of Environment, 77, 22–33. Chenab, C.F., Sonb, N.T., Changab, L.Y., & Chenb, C.R. (2011). Classification of rice cropping systems by empirical mode decomposition and linear mixture model for time-series MODIS 250 m NDVI data in the Mekong Delta, Vietnam. International Journal of Remote Sensing, 32(18), 5115–5134.

243

Esch, T., Metz, A., Marconcini, M., & Keil, M. (2014). Combined use of multi-seasonal high and medium resolution satellite imagery for parcel-related mapping cropland and grassland. International Journal of Applied Earth Observation and Geoinformation, 28, 230–237. Galvão, L.S., Epiphanio, J.C.N., Breunig, F.M., & Formaggio, A.R. (2012). Crop type discrimination using hyperspetral data. In P.S. Thenkabail, J.G. Lyon, & A. Huete (Eds.), Hyperspectral remote sensing of vegetation (pp. 397–421). Boca Raton, FL: CRC Press. Gao, B. (1996). NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment, 58(3), 257–266. Giri, C.P. (2012). Remote sensing of land use and land cover, principles and applications. Boca Raton, Florida: CRC Press (477 pp.). Gislason, P.O., Benediktsson, J.A., & Sveinsson, J.R. (2006). Random Forests for land cover classification. Pattern Recognition Letters, 27, 294–300. Gitelson, A., Kaufman, Y.J., & Merzlyak, M.N. (1996). Use of a green channel in remote sensing of global vegetation from EOS-MODIS. Remote Sensing of Environment, 58, 289–298. Huang, C., Davis, L.S., & Townshend, J.R.G. (2002). An assessment of support vector machines for land cover classification. International Journal of Remote Sensing, 23(4), 725–749. Jakubauskas, M.E., Legates, D.R., & Kastens, H. (2002). Crop identification using harmonic analysis of time-series AVHRR NDVI data. Computers and Electronics in Agriculture, 37, 127–139. James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An introduction to statistical learning with applications in R. New York: Springer (426 pp.). Jewell, N. (1989). An evaluation of multi-date SPOT data for agriculture and land use mapping in the United Kingdom. International Journal of Remote Sensing, 10, 939–951. Jones, H., & Vaughan, R. (2010). Remote sensing of vegetation: Principles, techniques and applications. New York: Oxford University Press (353 pp.). Liang, S., Li, X., & Wang, J. (2012). Advanced remote sensing: Terrestrial information extraction and applications. Oxford: Academic Press (799 pp.). Liaw, A., & Wiener, M. (2002). Classification and regression by randomForest. R News, 2(3), 18–22. Lo, T.H.C., Scarpace, F.L., & Lillesand, T.M. (1986). Use of multitemporal spectral profiles in agricultural land-cover classification. Photogrammetric Engineering and Remote Sensing, 52(4), 535–544. Masialeti, I., Egbert, S., & Wardlow, B.D. (2010). A comparative analysis of phenological curves for major crops in Kansas. GI Sciences and Remote Sensing, 47(2), 241–259. Mathur, A., & Foody, G. (2008). Crop classification by support vector machine with intelligently selected training data for an operational application. International Journal of Remote Sensing, 29(8), 2227–2240. Meyer, D., Dimitriadou, E., Hornik, K., Weingessel, A., & Leisch, F. (2014). e1071: Misc functions of the Department of Statistics (e1071), TU Wien. R package version 1.6-4. http://CRAN.R-project.org/package=e1071 Mingwei, Z., Qingbo, Z., Zhongxin, C., Jia, L., Yong, Z., & Chongfa, C. (2008). Crop discrimination in northern China with double cropping systems using Fourier analysis of time-series MODIS data. International Journal of Applied Earth Observation and Geoinformation, 10, 476–485. Moguerza, J.M., & Muñoz, A. (2006). Support vector machines with applications. Statistical Science, 21(3), 322–336. Mountrakis, G., Im, J., & Ogole, C. (2011). Support vector machines in remote sensing: A review. ISPRS Journal of Photogrammetry and Remote Sensing, 66, 247–259. Murakami, T., Ogawa, S., Ishitsuka, K., Kumagai, K., & Saito, G. (2001). Crop discrimination with multitemporal SPOT/HRV data in the Saga Plains, Japan. International Journal of Remote Sensing, 22(7), 1335–1348. Nguyen, T.T.H., De Bie, C.A.J.M., Ali, Smaling, E.M.A., & Chu, T.H. (2011). Mapping the irrigated rice cropping patterns of the Mekong Delta, Vietnam, through hypertemporal SPOT NDVI image analysis. International Journal of Remote Sensing, 33(2), 415–434. Odenweller, J.B., & Johnson, K.I. (1984). Crop identification using Landsat temporalspectral profiles. Remote Sensing of Environment, 14(1–3), 39–54. ODEPA (Oficina de Estudios y Políticas Agrarias) (2014). Región metropolitana: Información regional 2014. Santiago, Chile: Ministerio de Agricultura (14 pp.). Ozdogan, M. (2010). The spatial distribution of crop types from MODIS data: Temporal unmixing using independent component analysis. Remote Sensing of Environment, 114, 1190–1204. Pal, M. (2005). Random forest for classification in remote sensing. International Journal of Remote Sensing, 26(1), 217–222. Pal, M., & Mather, P.M. (2003). An assessment of the effectiveness of decision tree methods for land cover classification. Remote Sensing of Environment, 86, 554–565. Pal, M., & Mather, P.M. (2005). Support vector machines for classification in remote sensing. International Journal of Remote Sensing, 26(5), 1007–1011. Peña, M.A., & Altmann, S.H. (2009). Use of satellite-derived hyperspectral indices to identify stress symptoms in an Austrocedrus chilensis forest infested by the aphid Cinara cupressi. International Journal of Pest Management, 55(3), 197–206. R Core Team (2014). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing (URL http://www.R-project.org/). Roberto, C., Lorenzo, B., Michele, M., Micol, R., & Cinzia, P. (2012). In P. S. Thenkabail, J. G. Lyon, & A. Huete (Eds.), Optical remote sensing of vegetation water content. In Hyperspectral Remote Sensing of Vegetation (pp. 227–244). Boca Raton: FL: CRC Press. Rodriguez-Galiano, V.F., Ghimire, B., Rogan, J., Chica-Olmo, M., & Rigol-Sanchez, J.P. (2012). An assessment of the effectiveness of a random forest classifier for land-cover classification. ISPRS Journal of Photogrammetry and Remote Sensing, 67, 93–104. Roy, D.P., Wulder, M.A., Loveland, T.R., Woodcock, C.E., Allen, R.G., Anderson, M.C., ... Zhu, Z. (2014). Landsat-8: science and product vision for terrestrial global change research. Remote Sensing of Environment, 145, 154–172.

244

M.A. Peña, A. Brenning / Remote Sensing of Environment 171 (2015) 234–244

Sakamoto, T., Yokozawa, M., Toritani, H., Shibayama, M., Ishitsuka, N., & Ohno, H. (2005). A crop phenology detection using time-series MODIS data. Remote Sensing of Environment, 96, 366–374. Shao, J., Lunetta, R.S., Ediriwickrema, J., & Liames, J. (2010). Mapping cropland and major crop types across the Great Lakes Basin using MODIS-NDVI data. Photogrammetric Engineering and Remote Sensing, 75(1), 73–84. Sibanda, M., & Murwira, A. (2012). The use of multi-temporal MODIS images with ground data to distinguish cotton from maize and sorghum fields in smallholder agricultural landscapes of southern Africa. International Journal of Remote Sensing, 33(16), 4841–4855. Simonneaux, V., Duchemin, B., Helson, D., Er-Raki, S., Olioso, A., & Chehboun, A.G. (2007). The use of high-resolution image time series for crop classification and evapotranspiration estimate over an irrigated area in central Morocco. International Journal of Remote Sensing, 29(1), 95–116. Sun, H., Xu, A., Lin, H., Zhang, L., & Mei, Y. (2012). Winter wheat mapping using temporal signatures of MODIS vegetation index data. International Journal of Remote Sensing, 33(16), 5026–5042. Szuster, B.W., Chen, Q., & Borger, M. (2011). A comparison of classification techniques to support land cover and land use analysis in tropical coastal zones. Applied Geography, 31, 525–532. Turker, M., & Arikan, M. (2005). Sequential masking classification of multi-temporal Landsat7 ETM + images for field-based crop mapping in Karacabey, Turkey. International Journal of Remote Sensing, 26(17), 3813–3830. Van Niel, T.G., & McVicar, T.R. (2004). Determining temporal windows for crop discrimination with remote sensing: A case study in south-eastern Australia. Computers and Electronics in Agriculture, 45(1–3), 91–108. Wardlow, B., & Egbert, S. (2010). A comparison of MODIS 250-m EVI and NDVI data for crop mapping: A case study for southwest Kansas. International Journal of Remote Sensing, 31(3), 805–830.

Wardlow, B.D., & Egbert, S.L. (2008). Large-area crop mapping using time-series MODIS 250 m NDVI data: an assessment for the U.S. Central great plains. Remote Sensing of Environment, 112, 1096–1116. Wardlow, B., Egbert, S., & Kastens, J.H. (2007). Analysis of time-series MODIS 250 m vegetation index data for crop classification in the U.S. Central Great Plains. Remote Sensing of Environment, 108, 290–310. Xavier, A.C., Rudorff, B.F.T., Shimabukuro, Y.E., Berka, L.M.S., & Moreira, M.A. (2006). Multi-temporal analysis of MODIS data to classify sugarcane crop. International Journal of Remote Sensing, 27(4), 755–768. Xie, Y., Sha, Z., & Yu, M. (2008). Remote sensing imagery in vegetation mapping: A review. Journal of Plant Ecology, 1(1), 9–23. Zandler, H., Brenning, A., & Samimi, C. (2015). Quantifying dwarf shrub biomass in an arid environment: Comparing empirical methods in a high dimensional setting. Remote Sensing of Environment, 158, 140–155. Zhang, X., Friedl, M.A., Schaaf, C.B., Strahler, A.H., Hodges, J.C.F., Gao, F., ... Huete, A. (2003). Monitoring vegetation phenology using MODIS. Remote Sensing of Environment, 84, 471–475. Zheng, B., Myint, S.W., Thenkabail, P.S., & Aggarwal, R.M. (2015). A support vector machine to identify irrigated crop types using time-series Landsat NDVI data. International Journal of Applied Earth Observation and Geoinformation, 34, 103–112. Zhong, L., Gong, P., & Biging, G. (2014). Efficient corn and soybean mapping with temporal extendability: A multi-year experiment using Landsat imagery. Remote Sensing of Environment, 140, 1–13. Zhong, L., Hawkins, T., Biging, G., & Gong, P. (2011). A phenology-based approach to map crop types in the San Joaquin Valley, California. International Journal of Remote Sensing, 32(22), 7777–7804.