AN APPLICATION OF ROLL-INVARIANT POLARIMETRIC FEATURES FOR CROP CLASSIFICATION FROM MULTI-TEMPORAL RADARSAT-2 SAR DATA

Crops are dynamically changing and time-critical in the growing season and therefore multitemporal earth observation data are needed for spatio-temporal monitoring of the crops. This study evaluates the impacts of classical roll-invariant polarimetric features such as entropy (H), anisotropy (A), mean alpha angle (ᾱ) and total scattering power (SPAN) for the crop classification from multitemporal polarimetric SAR data. For this purpose, five different data set were generated as following: (1) Hᾱ, (2) HᾱSpan, (3) HᾱA, (4) HᾱASpan and (5) coherency [T] matrix. A time-series of four PolSAR data (Radarsat-2) were acquired as 13 June, 01 July, 31 July and 24 August in 2016 for the test site located in Konya, Turkey. The test site is covered with crops (maize, potato, summer wheat, sunflower, and alfalfa). For the classification of the data set, three different models were used as following: Support Vector Machines (SVMs), Random Forests (RFs) and Naive Bayes (NB). The experimental results highlight that HαASpan (91.43% for SVM, 92.25% for RF and 90.55% for NB) outperformed all other data sets in terms of classification performance, which explicitly proves the significant contribution of SPAN for the discrimination of crops. Highest classification accuracy was obtained as 92.25% by RF and HαASpan while lowest classification accuracy was obtained as 66.99% by NB and Hα. This experimental study suggests that roll-invariant polarimetric features can be considered as the powerful polarimetric components for the crop classification. In addition, the findings prove the added benefits of PolSAR data investigation by means of crop classification.


INTRODUCTION
The growth of worldwide population and correspondingly the increasing demand of foods push the countries to develop low-cost and practical solutions for the applications of precision agriculture.For the sustainable agricultural practices in regional and national scale, the obtaining information about the crop acreage is of importance however owing to the very dynamic structure of crops in the growing season, the mapping of crops and obtaining information about the crop yield present a challenge.In comparison to other land cover types, the agricultural targets have the rapid and dynamic structural changes (i.e.phenological changes) within short time periods in the growing season (McNairn andBrisco, 2004, Bargiel, 2017).Synthetic Aperture Radar (SAR) sensors use the signals at microwave wavelengths and thereby have the weather/daylight in-dependent imaging capability.The radar systems, based on its unique calibration procedure, are powerful and advantageous for the information retrieval about land surface characteristics and changing surface dynamics since they benefit from the relationship between the target and back-scattered radar signal in terms of polarization, physics and illuminating geometry (Schmullius et al., 2015).
In this sense, spaceborne SAR sensors can be effectively used for time-critical agricultural applications and provide the crucial information for crop classification especially for multitemporal analysis.For the mono-temporal SAR data, two crops could have similar backscatter information however when the crops grow and move to the next stage in growing period, the backscatter of each crop type changes over the time and this leads to the better discrimination of the crops (McNairn andBrisco, 2004, McNairn andShang, 2016).In several recent studies, the requirements and advantages of multi-temporal SAR images for successful crop classification and agricultural practices were stated (Jiao et al., 2014, Mirzaee et al., 2014, Larrañaga and Álvarez-Mozos, 2016, Bargiel, 2017, Dimov et al., 2017, Kussul et al., 2018, Woźniak et al., 2018).Jiao et al. (2014) investigated an object-oriented classification of multi-temporal RADARSAT-2 images for mapping and monitoring the crops.Mirzaee et al. (2014) tested the effects of phenological changes for the classification of agricultural fields using multitemporal TerraSAR-X images.Larrañaga and Álvarez-Mozos (2016) assessed the added value of quad-pol data in a multitemporal crop classification by using polarimetric features and object based random forest classification algorithm.Bargiel (2017) developed a new classification approach based on multitemporal data and crop phenology information by using the time-series of Sentinel-1 images.Dimov et al. (2017) compared the classification performance of time-series Sentinel-1 and Landsat-8 data for the crop classification.Kussul et al. (2018) investigated the multitemporal SAR and optical data (Landsat-8, Sentinel-1 and Sentinel-2) for crop specific mapping.Woźniak et al. (2018) tested the dual-polarization polarimetric processing of time-series Sentinel-1 data on land cover classification.All these above studies stressed the requirements of multitemporal SAR data set for an adequate crop classification and consequently demonstrated the potential use of radar remote sensing for agricultural applications, especially crop classification and temporal monitoring.
Polarimetric SAR (PolSAR) sensors record the complete characteristics of scattering field (coherent channels and phase information) and enable of extracting the bio-geophysical parameters for the canopy structure of crops.In comparison to single or dualpolarized SAR data, PolSAR data provide more unique and significant details in terms of scattering mechanism for natural and man-made targets (Steele-Dunne et al., 2017).The differences among the scattering mechanisms of crop types enables the seperation of crops since this mechanism is related to crop phenology and crop type (Lopez-Sanchez et al., 2012, Hütt andWaldhoff, 2018).For better understanding and easier interpretation of the scattering mechanism of the targets, polarimetric features can be generated from the complex PolSAR data by using polarimetric decompositions.Polarimetric features were found to be quite helpful for land cover and crop type classification (McNairn andBrisco, 2004, Steele-Dunne et al., 2017).Hariharan et al.( 2016) used the roll-invariant polarimetric features to classify urban areas using random forest (Hariharan et al., 2016).Tao et al. (2017) tested the potential use of roll-invariant and selected hidden polarimetric features in the rotation domain for land cover classification (Tao et al., 2017).Guo et al.( 2018) investigated the rollinvariant polarimetric features (only entropy and mean alpha angle) for the crop classification by using multitemporal quad and dual-polarization SAR data (Guo et al., 2018).All these three aforementioned studies showed the importance of roll-invariant polarimetric features for classification purposes.
This study discusses the impacts of roll-invariant polarimetric features such as entropy, anisotropy, mean alpha angle, and total scattering power (SPAN) for the crop classification from multitemporal polarimetric SAR data.A time-series of four PolSAR data (single look complex full polarimetric Radarsat-2 with fine quad-polarization acquisition mode) were acquired for the study area located in Konya, Turkey.The H-A-ᾱ (Cloude-Pottier) decomposition decomposition were used to obtain the roll-invariant polarimetric features (entropy, anisotropy and mean alpha angle).
The remainder of the paper is organized as follows.The data description and the details of data processing and image classification are given Section 2. In Section 3, the experimental results are presented and discussed.In the end, Section 4 provides some final conclusions and directions for future work.

Polarimetric Features
Classical roll-invariant polarimetric features which are not dependent of target orientations can extracted from the eigenvalue and eigenvector based decomposition of the coherency matrix by using H-A-ᾱ decomposition method.From this decomposition, the coherency matrix are formed as following: where λ1,λ2 and λ3 are the eigenvalues and U represent the decomposed eigenvectors.And the roll-invariant polarimetric features extracted from the H-A-ᾱ decomposition are defined as follows (Cloude andPottier, 1996, Lee andPottier, 2009): where Pi = λi λ2 + λ2 + λ3 (5) SPAN is considered as the total backscattering power and preferred in PolSAR image analysis since the speckle noise is relatively reduced compared to single channels of data (Wang et al., 2017).The SPAN is formulated as follows:   (Morena et al., 2004) The pre-processing of the SAR data is important to extract relevant information from the data for a successful crop classification.The following processing steps were applied to the Pol-SAR data to extract the entropy, mean alpha angle and anisotropy: First, the data were imported into ESA (European Space Agency) SNAP (The Sentinel Application Platform) toolbox and calibrated.Following this step, the Sinclair scattering matrices were converted to the coherency matrix for further polarimetric processing.After this step, H-A-ᾱ decomposition method was applied to the data to extract roll-invariant polarimetric features except SPAN.During the polarimetric decomposition process, window size was determined as 9 × 9.And then the geometric terrain correction was applied to the imagery using the Range Doppler orthorectification method.During the terrain correction,the Shuttle Radar Topography Mission (SRTM) data (1 arc second, approximately 30-m resolution) was automatically downloaded and used.Then, the data were resampled to 10 meter pixel resolution using bilinear interpolation.As the last step of the data preprocessing, the data were scaled to decibels [dB] and exported in GeoTIFF format.For extracting the SPAN band, almost the same pre-processing chain was applied except matrix generation and polarimetric decomposition process.The data were calibrated and then SPAN was generated through SNAP toolbox.The following step were the terrain correction, scaling the data into dB and exporting the data in GeoTIFF format.Figure 2 shows the pre-processing chain for each acquisition.

Study Area and Ground Truth Information
The study area (Figure 3) is located in Konya basin, Turkey and only covered with agricultural lands.Semi-arid climate is dominated in the region.The study area covers with the crop types which are as following: alfalfa, maize, potato, summer wheat, and sunflower for the vegetation season of 2016 summer season.
In-situ data were collected using a handheld GPS at the acquisition dates of each satellite imagery and were recorded to be used in the crop classification and accuracy assessment.The number of pixels for training and testing data are shown in Table 3.

Image Classification
With the major advances in the machine learning and computer vision, there has been an increasing trend in machine learning algorithms for the operational use in the fields of remote sensing and pattern recognition, especially for the classification purposes.
Image classification can still be considered as one of the powerful methods to extract information from the remotely-sensed images however are of some uncertainties such as the redundant (unnecessarily repeated) information among the spectral bands, imbalance of training samples in classes and the adaptation problem of training samples with imagery and implemented model  (Ghamisi et al., 2017).In this study, three different classification model were used as following: Support Vector Machines (SVMs), Random Forests (RFs) and Naive Bayes (NB).The classification models implemented in this study were briefly outlined in the following paragraph.
Naive Bayes classifier is one of the probabilistic classifiers in machine learning and uses Bayes theorem to predict the class of unknown data.This method calculates the posterior probability for each class and assign the particular pixel to the class with the highest posterior probability (Cánovas-García and Alonso-Sarría, 2015).Support Vector Machines are one of the kernel-based learning methods in machine learning and have been successfully and widely implemented in remote sensing and pattern recognition for classification and regression problems.SVMs aim to define the optimal hyperplane separating the two classes with the maximum margin width.case of linearly inseparable two classes, it use the kernel trick (the non-linear version of SVMs) to map the data into higher dimensional space (Melgani andBruzzone, 2004, Kavzoglu andColkesen, 2009).Random forests are one of the ensemble (multiple) classifiers and use many decision trees to predict the final class for each unknown pixel.The final decision boundary is constructed by the majority voting of all the individual trees (Pal, 2005, Gislason et al., 2006).

RESULTS AND DISCUSSION
In this section, classification accuracies of the five different data sets for each model and the class based accuracies via F1-score are presented.The classification accuracies of the models for each data set are reported in The experimental results highlight that the classification of Hᾱ-SPAN (89.19% for SVM, 92.06% for RF and 89.85% for NB) and H-ᾱ-A-SPAN (91.43% for SVM, 92.25% for RF and 90.55% for NB) outperformed all other data sets in terms of classification performance, which also explicitly prove the significant contribution of SPAN parameter for the discrimination of crops.The incorporation of SPAN parameter into H-ᾱ data set increased the overall classification accuracy as 14.18%, 19.23% and 22.87% for SVM, RF and NB, respectively.Highest classification accuracy was obtained as 92.25% by RF and H-ᾱ-A-SPAN while lowest classification accuracy was obtained as 66.99% by NB and H-ᾱ as seen in Table 4.The classification maps of the H-ᾱ and H-ᾱ-A-SPAN data sets are shown in Figure 5.The main reason of choosing this two data sets is that highest and lowest classification accuracies for per model were obtained from these SAR data sets.When F-score values for per class were examined, it is reported that Alfalfa was the most-confusing class among crop types.The maximum separation ratio for Alfalfa was obtained with NB and H-ᾱ-A-SPAN however this value is still less than 0.50.For maize class, the F-score values were obtained as 0.62 and 0.69 for H-ᾱ and by NB and RFs.With the incorporation of SPAN into the Hᾱ, the class accuracy values of maize were obtained as 0.88 and 0.92 for NB and RFs, respectively.The similar increase trend was also reported for potato class.The F-score values of the sunflower class were obtained as 1.00 with all classification models for the data set of and H-ᾱ-A-SPAN.This is the maximum F-score value among other classes.For all data sets, the F-score values are higher than 0.75 for sunflower class.For wheat class, the F-score values were obtained higher than 0.90 for all data set and implemented models.The highest value of F-score was obtained as 1.00 for the sunflower class while the lowest value was obtained 0.20 for the Alfalfa class.

CONCLUSIONS
In this study, the impacts of classical roll-invariant polarimetric features for the crop classification from multi-temporal polarimetric SAR data were investigated.To explore the impacts of polarimetric features, five different data sets were created and used as input data for classification.The incorporation of SPAN parameter into H-ᾱ data set increased the overall classification accuracy as 14.18%, 19.23% and 22.87% for SVM, RF and NB, respectively.These increases explicitly demonstrate the significant contribution of SPAN parameter for the discrimination of crops.SVMs outperformed other methods for the classification of [T] matrix because of the main capability of the SVMs in handling high-dimensional data with a limited number of training samples.RFs highest classification accuracies for all data sets except H-ᾱ and [T] matrix.The results emphasize that the metric features have to be extensively analysed and investigated in terms of the relationship between the changes in the phenological pattern of the crops and the polarimetric scattering.This experimental study suggests that roll-invariant polarimetric features can be considered as the powerful polarimetric components for the crop classification from multi-temporal polarimetric SAR.In addition, the findings prove the added benefits of PolSAR data investigation by means of crop classification.Our future research will address the extensive analysis and interpretation of the polarimetric features obtained from target decompositions for crop classification from multi-temporal SAR data.

Figure 2 .
Figure 2. Workflow diagram for pre-processing chain

Figure
Figure 3. Study area

Figure 4 .
Figure 4. Spatial distribution of training and testing data

Figure 5 .
Figure 5. Classification maps for the study area

Table 1 .
Image characteristics

Table 2 .
All pre-processing steps were implemented using the open source SNAP toolbox v6.0.1.

Table 2 .
Number of features

Table 5 .
Confusion MatrixThe class based accuracies were compared based on F1-score values which were reported for each classification model in Table6.