A new method for crop classification combining time series of radar images and crop phenology information

Remote Sensing of Environment 198 (2017) 369–383

Contents lists available at ScienceDirect

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

A new method for crop classification combining time series of radar images and crop phenology information Damian Bargiel Technische Universität Darmstadt, Institute of Geodesy, Germany

A R T I C L E

I N F O

Article history: Received 18 November 2016 Received in revised form 7 June 2017 Accepted 24 June 2017 Available online 3 July 2017 Keywords: Agriculture Sentinel-1 Radar Classification Phenology

A B S T R A C T Agricultural land cover is characterized by strong variations within relatively short time intervals. These dynamics are challenging for land cover classifications on the one hand, but deliver crucial information that can be used to improve the classifiers performance on the other hand. Since up to date mapping of crops is crucial to assess the impact of agricultural land use on the ecosystems, an accurate and complete classification of crop types is of high importance. In the presented study, a new multitemporal data based classification approach was developed that incorporates knowledge about the phenological changes on crop lands. It identifies phenological sequence patterns (PSP) of the crop types based on a dense stack of Sentinel-1 data and accurate information about the plant’s phenology. The performance of the developed methodology has been tested for two different vegetation seasons using over 200 ground truth fields located in northern Germany. The results showed that a dense time series of Sentinel-1 images allowed for high classification accuracies of grasslands, maize, canola, sugar beets and potatoes (F1-score above 0.8) using PSP as well as standard (Random Forest and Maximum Likelihood) classification method. The PSP approach clearly outperformed standard methods for cereal crops, especially for spring barley where the F1-score varied between zero and 0.43 for standard approaches, while PSP achieved values as high as 0.74 and 0.79 for both vegetation seasons. The PSP based approach also outperformed for oat, winter barley and rye. Furthermore, the PSP classification is more resilient to differences in farming management and conditions of growth since it takes information about each crop types’ growing stage and its growing period into consideration. The results also indicate, that the PSP approach was more sensitive to subtle changes such as the proportion of weeds within a field. © 2017 Elsevier Inc. All rights reserved.

1. Introduction 1.1. Motivation The ongoing worldwide population growth is constantly increasing the demand of foods (FAO, 2007). This consequently results in the intensification of agricultural productions. The increased food production threatens those services provided by the ecosystems, also referred to as ecosystem services, that go beyond the production of food (Foley et al., 2005). Such an overstraining of supporting or regulating ecosystem services leads to the loss of e.g. fresh water or fertile soils or to negative changes in climate regulation (Foley et al., 2005; Poh Sze Choo et al., 2005; Tivy, 1993; Matson et al., 1997). Therefore, it is an indispensable premise to keep all services provided by the ecosystems in balance in order to avoid a collapse of the ecosystem services. For this purpose, decision makers, scientists and

E-mail address: [email protected].

http://dx.doi.org/10.1016/j.rse.2017.06.022 0034-4257/© 2017 Elsevier Inc. All rights reserved.

planners rely on detailed spatial information about the land cover in agricultural areas (Poh Sze Choo et al., 2005; Adamowicz et al., 2005; Feng et al., 2010). Since agricultural land is strongly affected by spatial and temporal dynamics within and between each vegetation season, the mapping of crops represents an especially challenging task. Thanks to regular revisiting intervals and weather independent acquisition capabilities, satellite-based imaging radar is well suited to capture the agricultural land use dynamics for crop mapping. This is shown by a large number of studies using a stack of radar images for multi-temporal classification approaches (McNairn et al., 2002; McNairn et al., 2009b; Bargiel and Herrmann, 2011; Bargiel, 2013; Sonobe et al., 2014; Skriver et al., 2011; Skriver, 2012; Larrañaga and Álvarez Mozos, 2016; Mascolo et al., 2016; Larrañaga and Álvarez Mozos, 2016). Sentinel-1 is a modern C-band imaging radar two satellite constellation, which was launched in 2014 and 2016 for earth observation. It is representing the first sensor designed for the Copernicus Project initiated by the European Union. It provides free data access and unprecedented high temporal resolution, which enables new possibilities to capture the dynamics in agricultural

370

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

Fig. 1. Sketched dynamics of selected crops in the study area (cf. Section 2.1). Images of plants are taken from Meier and Bleiholder (2006).

areas using multi-temporal classification approaches that include information about the crops’ phenology. 1.2. Introduction to the dynamics of agricultural land cover In contrary to most land cover types, agricultural areas vary strongly within very short time intervals. Most crop types’ have a vegetation period of several months. In this time, the plants phenology changes strongly many times between seeding and harvest. Furthermore, the life cycles beginnings and endings differ depending on the crop types (cf. Fig. 1). This is also true for the different stages within the plant’s phenological development, for instance germination, leaf development or flowering as illustrated in Fig. 1. Further dynamics take place during the time intervals before the seeding period (preparation) and after the harvest (post harvest). During this time, the areas are affected by deviating management of the farmers. This is not necessarily unique depending on the types of crops that are planning to be planted or have been harvested. In summary, there are two types of short-term changes: the crop-type specific changes occurring between seeding and harvest (within the life cycle) and the non-crop-type specific differences outside the life cycle depending on the farmers management. Fig. 2 demonstrates the latter one for three fields of sugar beets before seeding in the stage of preparation.

possibilities for improvements of crop classifications. Since the crop’s phenology causes temporal dependencies between the single images of the multitemporal stack, more recent studies incorporate these dependencies using statistical models like Markov Random Fields (Leite et al., 2011; Siachalou et al., 2015) or Conditional Random Fields (Kenduiywo et al., 2015, 2016). The statistical modelling of temporal dependencies improves the classification results, serving as evidence for the potential given by a classification approach that is sensitive to temporal variations caused by agricultural dynamics. The presented study introduces a novel approach to incorporate the full knowledge of the crops’ phenology using a sequence based classification approach. As (Julea et al., 2011) proved in an unsupervised classification of multitemporal images, there are certain characteristic sequences, that are detectable in an image stack of agricultural areas. They called these sequences “frequent sequential patterns”. The presented study transfers this approach to a supervised classification with the incorporation of phenological knowledge to classify crops based on phenological sequential patterns (PSP). This is a new approach that ensures a full implementation of knowledge about the crops’ phenology into the classification approach to improve classification results.

2. Methodology

1.3. Objectives and related works

2.1. Study area and ground truthing

Most existing approaches for multitemporal crop classification based on radar images put a stack of images taken during the whole vegetation season into the classification process. They do not take the knowledge of the dynamics of the crops’ phenology (within their life cycle) or the fields’ appearance (outer the life cycle) into consideration (Bargiel and Herrmann, 2011; McNairn et al., 2002; Sonobe et al., 2014; Skriver et al., 2011; Skriver, 2012; Larrañaga and Álvarez Mozos, 2016). Although these studies show remarkable results for the classification of certain crop types and prove the high potential for radar-based crop classifications, some of them face challenges to differentiate between certain crop types e.g. single cereals (Bargiel and Herrmann, 2011). Current studies have demonstrated an improvement of crop classification results using polarimetric features of multitemporal SAR-images instead of backscatter intensity values only (Skriver et al., 2011; Skriver, 2012; Larrañaga and Álvarez Mozos, 2016; Mascolo et al., 2016). The implementation of knowledge about the development and phenology of crops into the classification process introduces further

The area of interest for the presented study is situated in northern Germany (Fig. 3). It has an extent of 39 kilometres (km) in eastwest expansion and 61 km in north-south expansion. It surrounds the city of Hanover, which is in the centre of the scene. The area is flat and dominated by agricultural land cover. The average annual precipitation is 656 mm and the average annual temperature is 8.9 ◦ C (Deutscher Wetterdienst, 2012). Ground truthing was conducted during the vegetation periods of the years 2014/2015 and 2015/2016. During the 2014/2015 period, 205 fields and grasslands were visited in eight sites distributed throughout the whole study area (several plots per site). The same areas plus several additional fields were surveyed in the period of 2015/2016, resulting in a total number of 257 areas (Fig. 4). For each field the crop type was recorded and its phenology was described according to the BBCH-scale developed by Meier and Bleiholder (2006) . For meadows the mowing status was recorded by defining whether the parcels have been mown or not. This step included an assessment of the mowing time according to the current growth

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

371

Fig. 2. Three photos of sugar beets. The photos on top were taken in March during the preparation stage before seeding. The photos below show the same plots during the growing stage in July. It can be seen clearly, that the fields had a different cover in March before sugar beets were planted. These differences resulted from different management before seeding and can result in different measurements of radar backscatter.

stage of the grass. Furthermore, a GPS tagged photo was taken from each plot. Table 1 illustrates the different crop types as well as the number of plots recorded for each type during field work. The dates of field trips for the 2014/2015 vegetation period were 16– 17 October, 10–12 March, 23–27 April, 28–31 May, 1–2 and 18–23 June, 16–23 July and 11–12 August. For the period 2015/2016 ground surveys were conducted on 2–9 and 29–31 October, 8–14 March, 28–30 May, 24–27 June and 4–9 July. 2.2. Data set Overall a number of 99 dual polarized (VV & VH) Sentinel-1A images taken within the time period of 13th October 2014 until 8th

October 2016 in Interferometric Wide Swath high resolution mode were available for investigations (Table 2). The dates of available images varied considerably between the different years. For instance, only three images were available in May 2016, while five images were available for the same month of the year 2015. The images were ordered as Level-1 Ground Range Detected (GRD) products with a spatial resolution of 22 m in azimuth direction and 20 m in range direction as well as a pixel spacing of ten by ten meters. According to the product specifications, a multilooking in range direction with a number of five looks has been conducted for speckle reduction (Sentinel-1 Team, 2013). The centre of the study area is located at an incidence angle of 40 ◦ . The images were calibrated to s 0 using the ENVI SARSCAPE Software and a digital terrain model from Space

Fig. 3. Location of study area and visualization of agricultural areas.

372

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

Fig. 4. Location of ground truth sites.

Radar Topography Mission (SRTM). This terrain corrected backscatter coefficient of both polarizations (dB scale) was used without further filtering as input data for the classification. 2.3. Definition of phenological stages Different phenological stages for each crop type based on the crops appearance inside the life cycle (cf. Section 1.2) were defined. The definition of phenological stages is based on the phenological appearance of the plants resulting from their morphological and physiological development. The appearance does not change much within one stage but differs strongly between two different stages. The categorization of the plants’ phenology results in a number of six phenological stages for each crop type. The equal number of six stages for each crop type ensures comparable quality of the phenologic description and an equal length of the phenological sequence

for each crop type (cf. Section 2.4). The BBCH-scale developed by Meier and Bleiholder (2006) was used to code the crops plants’ appearance at each phenological stage. The phenological stages and their corresponding BBCH-codes for each crop type are illustrated in Table 3. Since there is no BBCH-scale for grasslands, the phenological stages of grasslands (meadows and pastures) were defined according to the growth rate and recorded mowing events during the growing season (Table 4). Each phenological stage describes a uniform appearance which differs from the uniform appearance of any other phenological stage of the specific crop type. For instance, the appearance of canola differs between the phenological stage 1 when the plant is very small (up to 5 leaves) and the following stages, e.g. stage 5 when the plant is flowering. This is accurate for all stages of other crops. Although some crop types might have similar phenology during some stages, such as stages 1, 2 and 3 for cereal crops, there are always stages

Table 1 Overview of ground truth data. Vegetation season 2014/2015

Vegetation season 2015/2016

Land cover

Number of plots

Total area (ha)

Average size of plots (ha)

Number of plots

Total area (ha)

Average size of plots (ha)

Meadow (once mown) Meadow (twice mown) Pasture Potato Maize Canola Sugar beet Spring barley Winter barley Rye Winter wheat Oat

14 4 12 12 29 15 42 6 6 28 37

35.4 13.2 37.2 73 143.7 86.3 237.2 24.6 45.2 162.4 238.3

2.5 3.3 3.1 6.1 5 5.8 5.6 4.1 7.5 5.8 4.3

19 4 11 16 42 3 26 11 13 19 78 3

51.5 13.2 32.3 77.2 204.3 12.4 133.1 43.3 59.6 88.1 460.3 15.7

2.7 3.3 2.9 4.8 4.9 4.1 5.1 3.9 4.6 4.6 5.9 5.2

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383 Table 2 Dates of Sentinel-1 data acquisition for the study area. Year

2014

2015

2016

373

Table 4 Definition of phenological stages of grasslands according to the growth rate and mowing event during the growing season.

Month

Day

October November December January February March April May June July August September October November December January February March April May June July August September October

13; 22 15; 27 9; 21 2; 14; 29 7; 10; 19; 22 3; 6; 15; 27 8; 11; 20; 23 2; 5; 14; 17; 26 7; 10; 19; 22 1; 4; 13; 16; 25 6; 9; 18; 21; 30 2; 11; 14; 23; 26 5; 17; 20; 29 1; 10; 22; 25 4; 7; 16; 19; 28 9; 12; 21; 24 2; 5; 14; 17; 26; 29 9; 12; 24 2; 5; 14; 17; 26; 29 11; 20; 23 1; 4; 13; 28 7; 19; 22; 31 12; 15; 24 5; 8; 17; 20; 26; 29 2; 8

Stage

Meadow (once mown)

Meadow (twice mown)

Pasture

1 2 3 4 5 6

Winter (no growth) First 8 weeks of season Period until mowing Following 9 weeks Following 4 weeks Rest of season

Winter (no growth) Time until 1st mowing Time until 2nd mowing Following 3 weeks Following 4 weeks Rest of season

Winter (no growth) First 8 weeks of season Following 4 weeks Following 4 weeks Following 4 weeks Rest of season

images of the data stack, which were acquired during the stages’ existence according to the phenological sequence. Any classifier can be used to get this probability p(ys ) (cf. Section 2.5.2). In other words, a probability based model is applied globally, using selected image data (according to the phenological stage s) and training data of all classes. This results in a string of class probabilities for each pixel, where the string value of the investigated class y is addressed in order to be assigned to the pixel value of p(ys ) (cf. Fig. 6). This process is repeated for each phenological stage of the investigated class and summed up according to Eq. (1) resulting in the probability of a classes’ phenological sequence yseq k 

which strongly differ like stages 4, 5 and 6, when the spikes of the different cereals appear. The measurement of the radar backscatter varies throughout the different phenological stages. 2.4. Definition of phenological sequences in the study area The different phenological stages of the crop types occur during different times of the growing season and vary in duration. This process depends on crop type, date of seeding, climatic conditions and soil type in the area. A phenological sequence was defined as the assignment of the growing season’s days of year (doy) to the phenological stages of a crop type. Accordingly, each crop type has a certain phenological sequence for each specific site and year. Fig. 5 illustrates the phenological sequences for crop types within the study area. It is crucial to mention, that the phenological sequences for each study site can change with varying climatic conditions for different years, which did not apply for the investigated seasons of the presented study.

yseq =

Table 3 Definition of phenological stages for each crop type according to the BBCH-scale. Crop type

Stage 1

Stage 2

Stage 3

Stage 4

Stage 5

Stage 6

Cereals Canola Potato Maize Sugar beet

10–15 10–15 10–35 10–15 10–15

16–29 16–29 36–39 16–19 16–19

30–39 30–39 39–59 30–39 31–33

41–59 50–59 59–69 51–59 34–35

61–77 60–69 69–79 61–69 36–37

83–89 70–89 91–95 70–89 38–39

(1)

k

where, • s is a phenological stage, • k is the total number of phenological stages, • y is the corresponding class of the sequence. Calculations of yseq are conducted for each class y. Since the overarching goal of the PSP-classification is to define a local class label to the class with the most probable sequence, it was decided to penalize sequences with high variations within the probabilities of the sequences single stages. This is due to the fact that it is unlikely to have strong variances within the stages of a sequence when the class corresponding to the sequence exists at the local site. For this purpose the value of each local sequence yseq was divided by the coefficient of variance c of the single probabilities of the sequences’ phenological stages by calculating the quotient of the standard deviation and the mean of all p(ys ) of a sequence (cf. Eq. (2)).

2.5. Crop-classification based on phenological sequence patterns 2.5.1. Description of the PSP-classification approach The classification of crops based on phenological sequence patterns (PSP-classification) is a new classification method that incorporates phenological information into the classification process. The probability p(ys ) for a crop-type y (class) at a certain phenological stage s is calculated for each local entity of the investigated area (pixel) based on training data samples from ground surveys (cf. Section 2.5.2). This probability is calculated based on those radar

p(ys )

s=1

1 k

yseq =

k 

p(ys )

s=1

(2)

c

In a final step, the spatial context of the local site is accounted for, using the eight adjacent pixel neighbours. For this purpose, the value of a classes’ phenological sequence is calculated for the local pixel and its eight neighbours and averaged, according to Eq. (3).

yseq

n+1 1  = n+1 i=1

1 k

k 

p(ys )

s=1

c

(3)

where, • c is the coefficient of variance of the sequences’ single stage probabilities, • i is the index of the pixel an its eight nearest neighbours, • n is the the number of considered neighbours.

374

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

Fig. 5. Illustration of phenological sequences for crops and grasslands in the study area.

Fig. 7 illustrates the calculation of a sequence map with consideration of spatial context resulting from Eq. (3). 2.5.2. Calculation of probabilities As mentioned in Section 2.5.1, the calculation of probabilities p(ys ) for each phenological stage of the classes can be realized using any classification algorithm. In the presented study two different classifiers were used for calculations of p(ys ), namely the Naive Bayes and the Random Forest (Breiman, 2001) classifier. Both classifiers represent nonparametric classifiers, which are able to handle training data of various distributions. The Random Forests classifier represents a modern approach, which has proven to perform fast and accurate for large features spaces and variegated training data sets. In

the presented study, for each class, a number of 500 (300 for season 2015/2016) samples were randomly selected from the classes’ training data set which resulted from ground truth data (cf. Section 2.1). The number of samples was based on the training data size of the class with the lowest amount of available ground truth data. This was done to ensure a balanced number of training samples to avoid poor results of minor classes and possible favouring of major classes. The training plots that were used as data pool for the sampling were generated using the random sample function in the research tools of QGIS, which randomly separates 50% of the plots from each class to get training and validation sites. The calculations of the probabilities were realized in R programming language version 3.3.1. Probabilities based on Naive Bayes were calculated using the “klaR”-package

Fig. 6. Calculation of probability maps for the class of canola. A map for each phenological stage of canola is calculated based on images taken during the stages’ appearance. This results in six maps of probabilities, in which each pixel represents the probability of it being canola in the corresponding phenological stage.

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

375

Fig. 7. Calculation of a sequence map based on the probability maps of phenology stages (cf. Fig. 6).

(Weihs et al., 2005). The priors were set to equal a priori probabilities for all classes. The probabilities from the Random Forest classifier were calculated with the “randomForest” package (Liaw and Wiener, 2002). The number of trees was set to 1000 and the number of features randomly selected at each node was set to the square root of features for each phenological stage. The code was parallelized for faster calculations based on the “foreach”- and “doSNOW” packages (Analytics and Weston, 2015; Revolution Analytics and Weston, 2015). 2.5.3. Classification - finding most probable sequences After calculation of probabilities for each classes’ phenological stage as described in Section 2.5.2, the values for each classes’ phenological sequence were calculated according to Eq. (3). In a final

step, the class for each pixel of the site was defined by assigning the class label l of the class with the maximum value for the phenological sequence according to Eq. (4) (cf. Fig. 8). In other words, the phenological sequence pattern that maximally fits to the local site of the image was identified. The classification step was also realized in R programming language version 3.3.1 using the packages “sp”,“rgdal”,“raster” and “maptools” (Pebesma and Bivand, 2005; Bivand et al., 2016; Hijmans, 2016; Bivand and Lewin-Koh, 2016). The classification was conducted for agricultural areas only, which were identified using the federal topographic and cartographic information system of Germany (ATKIS).

l = max(yseq ) y

Fig. 8. Classification step. The class with highest sequence value is considered to be the class on ground assigned to the pixel of the result map.

(4)

376

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

Fig. 9. F1-score values for grasslands and for the merged class of grasslands (including all three grassland classes). Bars in blue colors represent standard methods and those in red colors represent the phenological sequence patterns (PSP) classification method introduced by the presented study.

2.6. Validation of results Producer’s and user’s accuracies (PA and UA) were calculated for each class using test plots from the sampling process described in Section 2.5.2, where the plots for each class were devided randomly into a test and a training set (cf. Section 2.5.2). Furthermore, the F1score was calculated for each class according to Eq. (5). F1 = 2 ×

(5)

The results were compared with classifications of the Maximum Likelihood classifier implemented in ENVI as well as Random Forest realized in R-programming language. The parameters for Random Forest were the same as for the probability calculations of the phenological stages in Section 2.5.2. In the contrary to the PSP classification, all images from Table 2 were taken as input features in one feature space. To consider the spatial context for these standard approaches, the classification result was filtered using a majority filter with a

Table 5 Producer’s accuracies (PA) and user’s accuracies (UA) achieved by Maximum Likelihood (ML) and Random Forest classifier (and Majority Filter postprocessing) compared to Phenological Sequence Patterns (PSP) approach using probabilities generated with Naive Bayes and Random Forest. Results of the investigation in vegetation season 2014/2015 [%].

Grasslands Potatoes Maize Canola Sugar beets S. barley W. barley Rye W. wheat

3. Results 3.1. Results of the vegetation season 2014/2015

PA × UA PA + UA

ML

three by three window (8 neighbours). This was done because the PSP-classification approach does also consider spatial context (cf. Eq. (3)).

Random Forest

PSP

PSP

Naive Bayes

Random Forest

PA

UA

PA

UA

PA

UA

PA

UA

96.5 91.3 86.3 99.3 98.9 0 80.3 81.1 96.0

93.4 88.7 93.9 99.6 92.3 100 99.1 87.2 76.5

95.5 92.2 89 99.2 98.5 16 77.2 95.8 95.6

92.2 93.3 94.9 98.4 92.2 86.7 98.5 85.2 86.8

95.1 79.9 95.5 99.7 93.3 76.4 95.9 81.1 82.7

87.7 85.4 92.6 87 92.8 51.5 84.7 94.9 96.9

96.5 81.3 96 99.9 96.9 74.1 95.5 92.8 89.5

89.3 93.3 93.3 91.3 94.3 73.6 96.9 93.2 97.3

The differentiation between the types of grasslands was neither possible with standard classification approaches nor with the PSP based classification, as can be seen in Fig. 9. The F1-score values for grassland types were around 0.5 or below. On the contrary, the values rose to high F1-socres of above 0.9 for all classification methods when the grassland types were merged to one class comprising all grassland types. Thus, the merged class of grasslands will be considered for further analysis of the results. The classification of different types of grasslands remains a challenging task for future studies. The producer’s accuracies (PA) and user’s accuracies (UA) for merged grasslands class and crop types are listed in Table 5. Additionally, Fig. 10 illustrates the F1-scores in a diagram. The results prove that classifications of agricultural areas based on a dense time series of Sentinel-1 images enable a good classification of crop types. F1-scores of maize, canola and sugar beets were above 0.9 using standard methods as well as the PSP based method. The values for potatoes were lower with PSP having a F1-score value of 0.87 (0.83 for probabilities based on Naive Bayes) compared to 0.93 for nonPSP based approach of Random Forest (0.90 for ML). The reasons for the lower accuracies of potatoes are discussed in Section 4. The PSP based classification clearly outperformed for cereal crops, especially for spring barley where the F1-score value was 0.74 compared to 0.27 with the standard Random Forest approach. The standard Maximum Likelihood (ML) classifier did not classify any pixel of spring barley correctly while the PSP classification based on probabilities from Naive Bayes reached a comparably high F1-score value of 0.62. The F1 score values were also higher for winter barley using the PSP method for classification. In the case of Random Forest based probability calculation the PSP F1-score delivered the highest rate of 0.96 for winter

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

377

Fig. 10. Representation of F1-score values for the vegetation season 2014/2015. Bars in blue colors represent standard methods and those in red colors represent the phenological sequence patterns (PSP) classification method introduced by the presented study.

barley. Fig. 11 illustrates the classification for a test plot of spring barley and winter barley. The PSP based approach also outperformed for winter wheat and rye having the highest F1-score values of 0.93 for the PSP based classification with probabilities based on Random Forest. This is mainly due to a higher rate of user’s accuracies for winter wheat and rye when PSP was used (cf. Table 5). The high error of commission for rye and winter wheat for standard approaches (non-PSP) was mainly due to wrongly classified pixels within the test plots of spring and winter barley. In other

words, the low user’s accuracies for rye and wheat were due to the low producer’s accuracies on barley test fields. For instance, 83.6% of spring barley test plots have been classified wrongly to the class of wheat and 13.6% to the class of rye, using the ML classifier (cf. Table 6). The values for the non-PSP Random Forest classifier were similar (62.8% to wheat, 13.4% to rye). This effect was also recognizable on winter barley test plots: 5.7% to wheat, 10% to rye with ML and 1.8% to wheat, 17.8 % to rye with Random Forest. Accordingly, the main errors for the classes of rye and winter wheat for

Fig. 11. Examples of classification on single fields of spring- and winter barley. Green color illustrates correctly classified pixels.

378

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

Table 6 Confusion matrices of classification for the vegetation period 2014/2015 [%]. Grass-lands

Potatoes

Maize

Canola

Sugar beets

Spring barley

Winter barley

Rye

Winter wheat

Maximum likelihood and Majority Filter Grasslands 96.5 Potatoes 0 Maize 1.6 Canola 0 Sugar beets 0.1 Spring barley 0 Winter barley 0.1 Rye 1.4 Winter wheat 0.3

0.4 91.3 6.1 0 1.9 0 0 0.3 0

0.7 4.5 86.3 0 8.0 0 0.1 0.3 0.1

0.2 0 0 99.3 0.1 0 0 0.3 0

0 0.2 0.6 0 98.9 0 0 0.2 0.1

2.4 0 0.3 0 0 0 0 13.6 83.6

0.8 0.1 0.3 1.0 1.8 0 80.3 10.1 5.7

0.9 0 0.8 0 0 0 0 81.1 17.2

0.5 0.1 0.6 0 0.7 0 0.1 2.1 96.0

Random Forest and Majority Filter Grasslands 95.5 Potatoes 0 Maize 2.5 Canola 0.3 Sugar beets 0.7 Spring barley 0 Winter barley 0 Rye 0.9 Winter wheat 0

0.1 92.2 4.5 0 2.3 0 0 0.8 0

0.2 2.6 89.0 0.2 7.5 0 0.1 0.2 0.4

0 0 0 99.2 0.3 0 0.1 0.3 0

0 0.1 0.4 0.1 98.5 0 0.1 0.3 0.5

6.6 0 0.9 0.1 0 16.0 0.3 13.4 62.8

0.6 0.2 0.2 1.0 1.2 0 77.2 17.8 1.8

1.0 0.1 0.6 0.1 0 0.6 0 95.8 1.8

0.8 0 0.5 0 0.5 0 0 2.4 95.6

PSP based on probabilities from Naive Bayes Grasslands 95.1 0 Potatoes 0 79.9 Maize 1.8 4.1 Canola 1.7 0.7 Sugar beets 0 14.4 Spring barley 0.7 0 Winter barley 0 0 Rye 0.5 0 Winter wheat 0.1 0.8

0.3 0.9 95.5 2.6 0.2 0.2 0.1 0.1 0

0 0 0 99.7 0.3 0 0 0 0

0.1 3.3 1.4 0.6 93.3 0.1 0.4 0.2 0.7

7.8 0 3.4 2.5 0 76.4 0 3.7 6.3

0.1 1.1 0.2 1.6 0 0 95.9 0.1 1.0

3.3 0.5 2.4 1.4 0.5 10.1 0.7 81.1 0.1

1.0 0.3 0.8 1.7 1.5 6.1 4.2 1.7 82.7

PSP based on probabilities from Random Forest Grasslands 96.5 0.2 Potatoes 0 81.3 Maize 1.6 5.6 Canola 1.4 0.2 Sugar beets 0 12.2 Spring barley 0 0 Winter barley 0 0 Rye 0.5 0.3 0 0 Winter wheat

0.4 0.8 96.0 1.5 0.9 0 0.1 0.1 0.3

0 0 0 99.9 0 0 0 0 0

0 1.1 0.9 0.4 96.9 0 0.1 0.4 0.2

11.5 0.2 1.6 1.0 0.1 74.1 0.2 3.4 8.1

0 0.3 0 1.9 0.6 0 95.5 0.7 0.8

1.7 0.1 1.3 0.4 0 3.2 0 92.8 0.4

0.6 0.1 1.3 1.3 0.6 2.8 0.8 3.0 89.5

Table 7 Confusion matrices of classification for the vegetation period 2015/2016 [%]. Grass-lands Random Forest and Majority Filter Grasslands 95.9 Potatoes 0 Maize 1.9 Canola 0 Sugar beets 0.2 Spring barley 0.8 Winter barley 0 Rye 1.2 Winter wheat 0 Oat 0

Potatoes

Maize

Canola

Sugar beets

Spring barley

Winter barley

Rye

Winter wheat

Oat

0 67.7 26.1 0.1 4.8 0 0 1 0.1 0.1

0.3 0 97.6 0 1.1 0.6 0.1 0.1 0.1 0

0 0.6 0.2 98.3 0.9 0 0 0 0 0

0 0.2 0.3 0 94.3 0 0 0 5.2 0

11.8 0 1.7 0.1 0.1 46.8 8.8 2.8 0 27.8

8 0 3.4 0.4 0.2 0.1 45 42.8 0 0

6.5 0 5.8 0 0.3 4.5 0 82.8 0 0.1

1.4 0.1 0.7 0.1 0.6 2.8 2.3 2.2 89.8 0

2.2 1.4 3.8 0 0.8 88.1 0.3 0 3.5 0

0.7 0.9 96 0.6 0.5 0.3 0.2 0.8 0.1 0

0 0.2 0 99.8 0 0 0 0 0 0

0.1 4.2 1 0.4 89 0 0 0.1 5.1 0

8.2 0 0.4 0.6 0 67.1 7.8 9.2 0 6.8

1.2 0.1 0.2 0.6 0 0 88.2 9.3 0.3 0

1 0 4.7 0.8 0 0 1 88.8 3.5 0.1

0 0.2 0.5 0.6 0.3 0.1 6.8 5.9 85.6 0

0 1.6 0.8 2.2 0 3 0.3 48.5 1.1 42.5

PSP based on probabilities from Random Forest Grasslands 95.7 0.8 Potatoes 0 74.8 Maize 0.5 15.7 Canola 1.1 1.4 Sugar beets 0 5.8 Spring barley 0.1 0.2 Winter barley 0.7 0.1 Rye 1.9 1.2 Winter wheat 0 0.1 Oat 0 0

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

standard approaches resulted from misclassifications of barley fields. It can be stated that the user’s accuracies and the F1-score would decline strongly for these two classes, given a larger size of barley test plots. This is not the case for the presented study because the main cereals in the area are rye and winter wheat, resulting in larger ground truth areas for these two classes (162,4 ha for rye, 238,3 ha for winter wheat) compared to barley (45,2 ha for winter barley, 24 6 ha for spring barley) (cf. Table 1). The PSP based classification on the other hand is not sensitive to these kinds of errors. The classification accuracies for barley plots were substantially higher, which resulted in lower misclassifications of barley plots in general and errors that resulted from other crops than wheat and rye (cf. Table 6). The PSP based classification clearly outperformed for the classification of cereals and offered high accuracy rates for the other crop types. This again, resulted in a very good classification of agricultural sites, which is not equally realizable through standard approaches. 3.2. Results of the vegetation season 2015/2016 For the vegetation season 2015/2016 solely Random Forest (RF) based probability calculations were conducted. This is due to the better performance of RF based probability calculations obtained from the results of the previous season 2014/2015 (Section 3.1). The acquisition dates of Sentinel-1 images and the number of plots vary between the two vegetation seasons (Tables 1, 2). This is also true for the number of plots for each class as illustrated in Table 1. Moreover, one additional class of oat, was classified for the second vegetation season. Detailed results of the classification for the season 2015/2016 are illustrated within the confusion matrices of Table 7. Additionally, classification results for selected areas are mapped in Fig. 13. The results are comparable to the ones from the previous season. A better performance of PSP classification was proved for this season as well for cereals crops, where F1 score value was clearly higher for spring and winter barley and oat (Fig. 12). For spring barley, the improvement of F1 score was 0.36 for PSP, which was slightly lower than in the previous season, where the improvement for Random Forest based PSP was 0.47 (cf. Fig. 10). A great improvement compared to the previous season was observed for winter barley, where

379

the PSP classifier performed better by a F1 score value of 0.17 compared to 0.09 of the previous season. The improvement was on a comparable level for the class of rye in both seasons (0.02 compared to 0.03 in first season). For the class of winter wheat, the classification results in 2015/2016 were lower by a F1 value of 0.03 for the PSP approach while the PSP outperformed for the previous season by a value of 0.02. Despite a small number of ground truth samples, the class of oat achieved the F1 score value of 0.44 for PSP. This is not a high value for a land cover classification, but indicates a better performance clearly compared to standard Random Forest (RF) results, where the class of oats could not even be detected on the test fields (F1 score of zero). The F1 score value of 0.79 for the classification of canola using PSP was relatively low, when compared to standard RF which had a value of 0.97. The values were also higher for the vegetation season 2014/2015 (F1 score 0.99 for RF and for PSP). The reason for the low F1 score of PSP in 2015/2016 was a lower user’s accuracy of 66% for canola even though the producer’s accuracy remained on a high level of above 99% (Table 8). The low user’s accuracy for the canola class resulted from the small number of available ground truth fields (3) in this season compared to 15 fields in the vegetation season before. Therefore, the error of commission had a stronger impact on the class in comparison with the season of 2014/2015. The F1 score value for grasslands was higher for PSP (0.94) compared to standard RF (0.86). This was different from the season of 2014/2015 where both classifiers performed above 0.93. The low value of 0.86 also resulted from a lower user’s accuracy of 77.2% for grasslands compared to 92.7% of PSP as illustrated in Table 8. The reason for the lower user’s accuracy was not a small number of ground truth plots between the seasons, as observed for the canola class. In both vegetation seasons the number of ground truth plots for grasslands was comparable. Thus, a better performance of PSP for grasslands in the season 2015/2016 can be stated. Moreover, the classification accuracies for potatoes were at a high level for both classifications of the 2015/2016 season. A lower accuracy for the PSP approach as a result of a field covered in weeds as observed in 2014/2015 (cf. Section 4) did not occur on the ground truth fields of the 2015/2016 season.

Fig. 12. Representation of F1 score values for the vegetation season 2015/2016. Bars in blue color represent results from Random Forest classification and those in red color represent the phenological sequence patterns (PSP) classification method introduced by the presented study.

380

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

Fig. 13. Maps of classification results for selected areas in the vegetation season 2015/2016.

4. Discussion This paper suggested and presented a novel classification approach incorporating the dynamic changes on croplands. It uses knowledge of the plants’ phenological changes over time in order to extract crop type specific sequences of features based on dense time series of Sentinel-1 data. The identification of these sequences within the global image based on probability calculations enabled for an improved cropland classification. This new approach has been applied to Sentinel-1 data for two different vegetation seasons, classifying nine, respectively ten, different crop types and validating against a high number of ground truth plot. Generally, the results show that multitemporal classifications based on dense time series of radar images enable for very good results with regard to noncereal crop types. Studies based on lower densities of sigma nought

backscatter time series could not reach comparably high accuracies for maize, potatoes or sugar beets (Bargiel and Herrmann, 2011). Furthermore, the PSP based classification introduced in this study strongly improves the accuracies for the different types of cereal. Standard approaches, namely Maximum Likelihood and Random Forest classification, could not achieve a comparable differentiation within the group of cereals. This is a challenging task as shown by studies that could not differentiate within the group of cereals (Bargiel and Herrmann, 2011; Jiao et al., 2014). Accordingly, hitherto many existing studies do not focus on classes of single cereal types, but instead classify other crop types than cereals (McNairn et al., 2014; Villa et al., 2015; Navarro et al., 2016; Guarini et al., 2016), only one single cereal crop type (McNairn et al., 2009a; Sonobe et al., 2015) or one grouped class of cereals (Bargiel and Herrmann, 2011; McNairn et al., 2009b; Jiao et al., 2014).

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383 Table 8 Producer’s accuracies (PA) and user’s accuracies (UA) achieved by Random Forest classifier (and Majority Filter postprocessing) compared to Phenological Sequence Patterns (PSP) approach using probabilities generated with Random Forest in vegetation season 2015/2016 [%]. Random Forest

PSP Random Forest

Class

PA

UA

PA

UA

Grasslands Potatoes Maize Canola Sugar beets Spring barley Winter barley Rye Winter wheat Oat

95.91 67.74 97.64 98.29 94.31 46.81 45.02 82.81 89.76 0

77.17 97.82 84.59 95.44 91.37 39.5 62.89 75.78 98.73 0

95.67 74.77 96 99.79 89 67.08 88.23 88.82 85.61 42.55

92.13 86.69 89.49 66.03 93.83 95.76 56.35 74.06 97.56 46.18

Multitemporal SAR polarimetry features seem to be a crucial factor for the improvement of crop classification results as proved by recent studies (Skriver et al., 2011; Skriver, 2012; Larrañaga and Álvarez Mozos, 2016; Mascolo et al., 2016). Thus, the implementation of polarimetric features into the presented PSP approach is very promising for further improvements in the classification of croplands. In addition, the fusion with data provided by optical sensors, such as Sentinel-2, should be tested for the PSP approach. This is also true for radar sensors of different frequency domains, for instance Xor L band radar. The PSP classification approach focuses on the crop development on ground. It considers the fact that the beginning and ending of vegetation time vary for different crop types. Thus, it analyzes only those temporal data features, which are relevant according to the crop types’ occurrence on ground. This is one aspect which explains the better results of the classification approach. Beyond that, the PSP classification subdivides the feature space with regard to the main stages of a crop types’ phenological development. This allows for a subtle analysis of the time series with accordance to the main morphological changes within the phenology of the plants. Each phenological stage of each crop plant potentially carries particular information that allows for a better discrimination of the crop type. This consideration of subtle information is believed to be a further reason for the good classification results. Further research must analyze how strong the data of each phenological stage contributes to the classification result. Possibly, a further refinement of the definition of phenological stages can even improve the classification results. This must be carried out in further research work. A further topic that needs to be investigated in the future is the redundancy of image data. This can be realized by the calculation of feature importance within each phenological stage to select mostly important features for the PSP based classification. Such an analysis can reduce the data amount and at the same time it can avoid a decline of classification accuracy. This is a crucial point in order to handle the high data amount provided by modern satellite missions for earth observation. The fact that the presented PSP approach takes the dynamics of agricultural lands into account and also considers the phenology of crops, ensures a high resilience of the method. This is in terms of agricultural management changes or natural factors such as the climate and the soil. This creates possibilities for highly accurate crop classifications on large scales or different climate conditions between the vegetation seasons. For this purpose, the phenological

381

sequences can be adjusted to local site conditions using models for the definition of agricultural growth stages based on soil and climate data. A further possibility to receive information about crops’ local phenology, which is inevitable for the PSP classification, is to derive this information from satellite radar measurements. Conradsen et al. (2016) demonstrated a statistical approach to detect changes on grasslands and rye crops based on polarimetric L-band measurements from EMISAR. Further studies showed the suitability of X- and C-band SAR data for the detection of growth stages for rice crops (Lopez-Sanchez et al., 2012; Lopez-Sanchez et al., 2014; Yuzugullu et al., 2015; Küçük et al., 2016; Yuzugullu et al., 2017; Erten et al., 2016). Mascolo et al. (2016) introduced a method for the determination of phenological stages of oat, barley, wheat and maize based on polarimetric C-band measurements of RADARSAT-2. Vicente-Guijalba et al. (2015) developed a dynamic model based on an extended Kalman filter for real time determination of crop stages and showed its applicability to barley, wheat and oat. These methods of data based phenology determination are of very high importance for the PSP classification method since it relies on frequent and accurate data of the local phenology, which cannot be provided by ground truthing in operational mode. The PSP based classification also seems to be more sensitive to subtle anomalies of certain field parcels. For instance, for the class of potatoes (vegetation season 2014/2015) low classification accuracy for a field that has been overgrown by weeds in late summer could be observed. The same field was classified with a high accuracy using standard classifiers. Fields that were not covered by weeds were classified correctly at a high rate by PSP and non-PSP approaches (cf. Fig. 14). These kinds of anomalies are also responsible for a lower classification accuracy of two potato fields that suffered from droughts during late summer. The lower accuracies for the class of potatoes (cf. Section 3) result from this sensitivity of the PSP approach. The higher sensitivity of the PSP classification establishes possibilities for the detection of organic farming or drought monitoring which need to be investigated in the future. 5. Conclusions and outlook The presented study demonstrates a promising approach for crop classification, which considers detailed information on the crops’ phenology. A new classification approach (PSP) that identifies crop specific sequences within a dense stack of multitemporal Sentinel-1 images has been developed. The high performance of this approach has been demonstrated through the classification results for two different vegetation seasons and numerous different classes. The approach can serve as a base for the classification of very large areas, where phenology can vary due to different climatic conditions over the area. Further studies can also test the performance of the developed approach with use of different multitemporal input data. These could be polarimetric features as well as optical sensor data or imaging radar data of different frequencies. Acknowledgments Acknowledgments are addressed to Pouya Hedayati for the support in data processing as well as Isabel Karst and Saori Miyake for proof reading the article. I am also grateful to my mentor, Prof. Sörgel, for his helpful comments on the manuscript. Moreover, I am grateful to the reviewers whose comments helped to improve the quality of the manuscript remarkably. This work was funded by the Federal Ministry for Economic Affairs and Energy of Germany (Project 50EE1326). Sentinel-1 data was provided by the European Space Agency (ESA).

382

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383

Fig. 14. Examples of classification on two different potato plots of vegetation season 2014/2015. Green color represents correctly classified pixels. Plot 167 (a) has a high rate of accuracy for all classifiers while plot 33 (b) shows lower accuracy for PSP based classification. This is due to a high amount of weeds, which grew on field 33 between June (e) and July (f). On the contrary, there are no weeds on field 167 neither in June (c) nor in July (d).

References Adamowicz, W., Akçakaya, H., Arcenas, A., Babu, S., Balh, D., Confalonieri, U., Cramer, W., Falconí, F., Fritz, S., Green, R., Gutíerrez-Espeleta, E., Hamilton, K., Kane, R., Latham, J., Matthews, E., Ricketts, T., Tian Xian, Y., 2005. Analytical approaches for assessing ecosystem condition and human well-being.. In: Assessment, M.E. (Ed.), Global Assessment Reports Volume: 1 Current State & Trends Assessment chapter 2. Island Press., pp. 37–71. Analytics, R., Weston, S., 2015. foreach: Provides Foreach Looping Construct for R. https://CRAN.R-project.org/package=foreach. R package version 1.4.3.

Bargiel, D., 2013. Capabilities of high resolution satellite radar for the detection of semi-natural habitat structures and grasslands in agricultural landscapes. Ecol. Inf. 13, 9–16. http://www.sciencedirect.com/science/article/pii/ S1574954112001021. http://dx.doi.org/10.1016/j.ecoinf.2012.10.004. Bargiel, D., Herrmann, S., 2011. Multi-temporal land-cover classification of agricultural areas in two European regions with high resolution spotlight TerraSAR-X Data. Remote Sens. 3, 859–877. http://www.mdpi.com/2072-4292/3/5/859/. http://dx. doi.org/10.3390/rs3050859. Bivand, R., Keitt, T., Rowlingson, B., 2016. rgdal: Bindings for the Geospatial Data Abstraction Library. https://CRAN.R-project.org/package=rgdal. R package version 1.1-10.

D. Bargiel / Remote Sensing of Environment 198 (2017) 369–383 Bivand, R., Lewin-Koh, N., 2016. maptools: Tools for Reading and Handling Spatial Objects. https://CRAN.R-project.org/package=maptools. R package version 0.8-39. Breiman, L., 2001. Random Forests. Mach. Learn. 45, 5–32. Conradsen, K., Nielsen, A.A., Skriver, H., 2016. Determining the points of change in time series of polarimetric SAR data. IEEE Trans. Geosci. Remote Sens. 54, 3007–3024. http://dx.doi.org/10.1109/TGRS.2015.2510160. Deutscher Wetterdienst, 2012. Mittelwerte der Temperatur und des Niederschlags bezogen auf den aktuellen Standort. http://www.dwd.de. Online: http://www. dwd.de. Erten, E., Lopez-Sanchez, J.M., Yuzugullu, O., Hajnsek, I., 2016. Retrieval of agricultural crop height from space: a comparison of sartechniques. Remote Sens. Environ. 187, 130–144. http://www.sciencedirect.com/science/article/pii/ S0034425716303790. http://dx.doi.org/10.1016/j.rse.2016.10.007. FAO, 2007. A Review of the Current State of Bioenergy Development in G8 +5 Countries. Food and Agriculture Organization of the United Nations, Feng, X., Fu, B., Yang, X., Lü, Y., 2010. Remote sensing of ecosystem services: an opportunity for spatially explicit assessment. Chin. Geogr. Sci. 20, 522–535. Foley, J.A., DeFries, R., Asner, G.P., Barford, C., Bonan, G., Carpenter, S.R., Chapin, F.S., Coe, M.T., Daily, G.C., Gibbs, H.K., Helkowski, J.H., Holloway, T., Howard, E.A., Kucharik, C.J., Monfreda, C., Patz, J.A., Prentice, I.C., Ramankutty, N., Snyder, P.K., 2005. Global consequences of land use. Science 309, 570–574. Guarini, R., Bruzzone, L., Santoni, M., Vuolo, F., Dini, L., 2016. Analysis of the potentiality of multi-temporal COSMO-SkyMed data for classifying summer crops. 2016 IEEE International Geoscience and Remote Sensing Symposium (IGARSS). pp. 3170–3173. Hijmans, R.J., 2016. raster: Geographic Data Analysis and Modeling. https://CRAN.Rproject.org/package=raster. R package version 2.5-8. Jiao, X., Kovacs, J.M., Shang, J., McNairn, H., Walters, D., Ma, B., Geng, X., 2014. Object-oriented crop mapping and monitoring using multi-temporal polarimetric RADARSAT-2 data. ISPRS J. Photogramm. Remote Sens. 96, 38–46. http://www. sciencedirect.com/science/article/pii/S0924271614001695. http://dx.doi.org/10. 1016/j.isprsjprs.2014.06.014. Julea, A., Méger, N., Rigotti, C., Trouvé, E., Bolon, P., L˘az˘arescu, V., 2011. Mining pixel evolutions in satellite image time series for agricultural monitoring. In: Perner, P. (Ed.), Advances in Data Mining. Applications and Theoretical Aspects chapter Mining Pixel Evolutions in Satellite Image Time Series for Agricultural Monitoring. Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 189–203. Kenduiywo, B.K., Bargiel, D., Soergel, U., 2015. Spatial-temporal conditional random fields crop classification from TerraSAR-X Images.. ISPRS Annals Photogramm. Remote Sens. Spat. Inf. Sci. 2, Kenduiywo, B.K., Bargiel, D., Soergel, U., 2016. Crop type mapping from a sequence of TerraSAR-X images with dynamic conditional random fields. ISPRS Annals Photogramm. Remote Sens. Spat. Inf. Sci. 59–66. Küçük, C., Taskin, ¸ G., Erten, E., 2016. Paddy-rice phenology classification based on machine-learning methods using multitemporal co-polar X-Band SAR Images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 9, 2509–2519. Larra naga, A., Álvarez Mozos, J., 2016. On the added value of Quad-Pol data in a multi-temporal crop classification framework based on RADARSAT-2 imagery. Remote Sens. Leite, P.B.C., Feitosa, R.Q., Formaggio, A.R., da Costa, G.A.O.P., Pakzad, K., Sanches, I.D., 2011. Hidden Markov models for crop recognition in remote sensing image sequences. Pattern Recogn. Lett. 32, 19–26. http://dx.doi.org/10.1016/j.patrec. 2010.02.008. Liaw, A., Wiener, M., 2002. Classification and regression by Random Forest. R News 2, 18–22. http://CRAN.R-project.org/doc/Rnews/. Lopez-Sanchez, J., Cloude, S., Ballester-Berman, J., 2012. Rice phenology monitoring by means of SAR polarimetry at X-Band. IEEE Trans. Geosci. Remote Sens. 50, 2695–2709. Lopez-Sanchez, J.M., Vicente-Guijalba, F., Ballester-Berman, J.D., Cloude, S.R., 2014. Polarimetric response of rice fields at C-Band: analysis and phenology retrieval. IEEE Trans. Geosci. Remote Sens. 52, 2977–2993. http://dx.doi.org/10.1109/TGRS. 2013.2268319. Mascolo, L., Lopez-Sanchez, J.M., Vicente-Guijalba, F., Nunziata, F., Migliaccio, M., Mazzarella, G., 2016. A complete procedure for crop phenology estimation with PolSAR data based on the complex Wishart classifier. IEEE Trans. Geosci. Remote Sens. 54, 6505–6515. Matson, P.A., Parton, W.J., Power, A.G., Swift, M.J., 1997. Agricultural intensification and ecosystem properties. Science 277 (5325), 504–509.

383

McNairn, H., Champagne, C., Shang, J., Holmstrom, D., Reichert, G., 2009a. Integration of optical and Synthetic Aperture Radar (SAR) imagery for delivering operational annual crop inventories. ISPRS J. Photogramm. Remote Sens. 64, 434–449. http:// www.sciencedirect.com/science/article/pii/S0924271608000944. http://dx.doi. org/10.1016/j.isprsjprs.2008.07.006. McNairn, H., Ellis, J., Van Der Sanden, J.J., Hirose, T., Brown, R.J., 2002. Providing crop information using RADARSAT-1 and satellite optical imagery. Int. J. Remote Sens. 23, 851–870. http://dx.doi.org/10.1080/01431160110070753. McNairn, H., Kross, A., Lapen, D., Caves, R., Shang, J., 2014. Early season monitoring of corn and soybeans with TerraSAR-X and RADARSAT-2. Int. J. Appl. Earth Obs. Geoinf. 28, 252–259. http://www.sciencedirect.com/science/article/ pii/S0303243413001839. http://dx.doi.org/10.1016/j.jag.2013.12.015. McNairn, H., Shang, J., Champagne, C., Jiao, X., 2009b. TerraSAR-X and RADARSAT-2 for crop classification and acreage estimation. 2009 IEEE International Geoscience & Remote Sensing Symposium, Proceedings. vol. 5. pp. 898–901. Meier, U., Bleiholder, H., 2006. Growth stages of mono-and dicotyledonous plants. BBCH Monograph, Federal Biological Research Centre for Agriculture and Forestry.. Navarro, A., Rolim, J., Miguel, I., Catalão, J., Silva, J., Painho, M., Vekerdy, Z., 2016. Crop Monitoring based on SPOT-5 Take-5 and Sentinel-1A data for the estimation of crop water requirements. Remote Sens. 8, 525. http://www.mdpi.com/20724292/8/6/525. http://dx.doi.org/10.3390/rs8060525. Pebesma, E., Bivand, R., 2005. Classes and methods for spatial data in R. R News 5, 9–13. http://CRAN.R-project.org/doc/Rnews/. Poh Sze Choo, H., Cooper, D., Devendra, C., Dixon, J., Gaskell, J., Khan, S., Lal, R., Lipper, L., Pretty, J., Primavera, J., Ramankutty, N., Viglizzo, E., Wiebe, K., 2005. Cultivated systems. In: Assessment, M.E. (Ed.), Global Assessment Reports Volume: 1 Current State & Trends Assessment chapter 26. Island Press., pp. 747–787. Analytics, Revolution, Weston, S., 2015. doSNOW: Foreach Parallel Adaptor for the ‘snow’ Package. https://CRAN.R-project.org/package=doSNOW. R package version 1.0.14. Sentinel-1 Team, E., 2013. Sentinel-1 User Handbook. European Space Agency. Siachalou, S., Mallinis, G., Tsakiri-Strati, M., 2015. A Hidden Markov models approach for crop classification: linking crop phenology to time series of multi-sensor remote sensing data. Remote Sens. 7, 3633. http://www.mdpi.com/2072-4292/ 7/4/3633. http://dx.doi.org/10.3390/rs70403633. Skriver, H., 2012. Crop classification by multitemporal C- and L-Band single- and dual-polarization and fully polarimetric SAR. IEEE Trans. Geosci. Remote Sens. 50, 2138–2149. Skriver, H., Mattia, F., Satalino, G., Balenzano, A., Pauwels, V.R.N., Verhoest, N.E.C., Davidson, M., 2011. Crop Classification Using Short-Revisit Multitemporal SAR Data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 4, 423–431. Sonobe, R., Tani, H., Wang, X., Kobayashi, N., Shimamura, H., 2014. Random forest classification of crop type using multi-temporal TerraSAR-X dual-polarimetric data. Remote Sens. Lett. 5, 157–164. % tt arXiv:http://dx.doi.org/10.1080/2150704X.2014.889863. Sonobe, R., Tani, H., Wang, X., Kobayashi, N., Shimamura, H., 2015. Discrimination of crop types with TerraSAR-X-derived information. Phys. Chem. Earth Parts A/B/C 8384, 2–13. Emerging science and applications with microwave remote sensing data. Tivy, J., 1993. Landwirtschaft und Umwelt. - Agrarökosysteme in der Biosphäre. Spektrum Akademischer Verlag GmbH Heidelberg, Berlin, Oxford., pp. 344. Vicente-Guijalba, F., Martinez-Marin, T., Lopez-Sanchez, J.M., 2015. Dynamical approach for real-time monitoring of agricultural crops. IEEE Trans. Geosci. Remote Sens. 53, 3278–3293. http://dx.doi.org/10.1109/TGRS.2014.2372897. Villa, P., Stroppiana, D., Fontanelli, G., Azar, R., Brivio, P.A., 2015. In-season mapping of crop type with optical and X-Band SAR data: a classification tree approach using synoptic seasonal features. Remote Sens. 7, 12859. http://www.mdpi.com/20724292/7/10/12859. http://dx.doi.org/10.3390/rs71012859. Weihs, C., Ligges, U., Luebke, K., Raabe, N., 2005. klaR analyzing German business cycles. In: Baier, D., Decker, R., Schmidt-Thieme, L. (Eds.), Data Analysis and Decision Support. Springer-Verlag, Berlin, pp. 335–343. Yuzugullu, O., Erten, E., Hajnsek, I., 2015. Rice growth monitoring by means of X-Band co-polar SAR: feature clustering and BBCH scale.. IEEE Geosci. Remote Sens. Lett. 12, 1218–1222. Yuzugullu, O., Erten, E., Hajnsek, I., 2017. Estimation of rice crop height from X- and Cband PolSAR by metamodel-based optimization. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 10, 194–204. http://dx.doi.org/10.1109/JSTARS.2016.2575362.