Impact of SAR Data Filtering on Crop Classification Accuracy

Transcription

1 Impact of SAR Data Filtering on Crop Classification Accuracy M. Lavreniuk, N. Kussul, M. Meretsky Dept. of Space Information Technologies and Systems Space Research Institute NAS Ukraine and SSA Ukraine Kyiv, Ukraine V. Lukin, S. Abramov, O. Rubel Dept of Transmitters, Receivers and Signal Processing National Aerospace University KhAI Kharkov, Ukraine Abstract For many applied problems in agricultural monitoring and food security, it is important to provide reliable crop classification maps. In this paper, we aim to compare performance of different filters available in ESA SNAP toolbox and compare them with our approach with applying to reduce speckle in multitemporal synthetic-aperture radar (SAR) Sentinel- 1 imagery. For this, we evaluate an impact of SAR data filtering on crop classification accuracy. We have found that overall classification accuracy without any filtering is 82.6% whilst the use of different despeckling methods achieves gain of crop map accuracy from +3.2% to +5.1% compared to classification of original data. The most accurate crop map has been obtained for SAR images pre-processed by DCT-based filter. Keywords SAR; speckle filtering; crop mapping; Sentinel-1 I. INTRODUCTION Agriculture is one of the key areas where Remote Sensing (RS) techniques can be efficiently implemented for solving wide range of tasks (crop mapping, crop monitoring, crop yield forecasting, etc.) on regular basis [1] [6]. In general studies, for crop classification and crop monitoring, researchers mainly utilized optical data. One of the main obstacles in utilizing optical imagery is the presence of clouds and shadows that introduce missing or severely distorted values. At local scale, it is usually possible to acquire cloudfree images in the crucial period of vegetation cycle. However, this is not the case for large territories. Now, Sentinel mission represents really new opportunities in agricultural monitoring it has become possible to use free of charge (for non-commercial use) weather independent synthetic-aperture radar (SAR) satellite images form Sentinel- 1A/B with 10 m spatial resolution. Those satellites have offered global coverage by dual-polarization data with six days revisit frequency. Meanwhile, SAR data also have some specific issues relating to their processing and use [7]. The basic and inevitable drawback of SAR images is the presence of noise-like phenomena called spackle that arises due to coherent mode of backscattered signal processing [8], [9]. Speckle presence and its properties should be taken into consideration for almost all operations of image processing applied to SAR data such as edge detection [10] [11], denoising [12] [13], image segmentation and object detection [14], lossy compression [15], etc. There are several known peculiarities of speckle. Firstly, it is supposed to be pure multiplicative noise [8], [12]. Secondly, it has probability density function (PDF) that is usually non- Gaussian [8] [13]. Thirdly, speckle possesses spatial correlation [16] that has to be taken into account. Thus, it is desirable to reduce speckle and many existing image processing packages allow doing this. For example, ESA SNAP toolbox provides such filters as: Boxcar, Frost, Gamma- MAP, Intensity driven adaptive neighborhood (IDAN), Lee, Lee-Sigma, Refined Lee, Median [17] [20] where some filters are specially designed for despeckling (e.g., Lee, Lee- Sigma, and Frost). Analysis of speckle characteristics is important (can be considered as a pre-requisite) for choosing a proper technique of radar image processing before utilizing those images for solving applied tasks. Thus, the aim of this paper is to analyze characteristics of Sentinel-1A/B images similarly to methodology presented in [16] and to estimate impact of filtering techniques on crop classification accuracy. II. PROPERTIES OF SENTINEL SAR IMAGES Initial assumptions on speckle that can be met in many fundamental books are that it is pure multiplicative and non- Gaussian [8]. Analysis of validity for these assumptions can be carried out in different manner, e.g., manually by analysis carried out by experts and using special automatic tools [16]. Existing software such as ENVI or its analogs allow manual selection of image fragments that are considered homogeneous by an expert. Examples of such fragments (all of rectangular shape) are shown by frames for VV (Vertical- Vertical) and VH (Vertical-Horizontal) polarization components of SAR image fragments presented in Fig. 1 (image fragment size is 512x512 pixels). Analysis of data for such fragments that have different means have confirmed the assumption that speckle is pure multiplicative. The estimated value of its relative variance σ 2 μ is about Automatic blind estimation [16] has resulted in almost the same estimates of σ 2 μ that varied only slightly (no more than by 10 15%) from one processed image to another. Histogram analysis and Gaussianity tests have shown that PDF of speckle is close to Gaussian as this often occurs for multilook SAR images [8]. This is favorable for despeckling since some filters do not work properly for non-gaussian /17/$ IEEE 912

2 noise resulting in possible artifacts [21] or leading to possible bias in homogeneous image regions. As it has been mentioned above, speckle can be spatially correlated. This can be proven in different many by special tests [16], by analysis of 2D spatial autocorrelation function, by considering 2D Fourier spectrum or 2D DCT spectrum. We prefer the latter approach since the results of analysis can be useful for filter parameter setting (see Section III for more details). values close to unity where Dp ( k, l ) for low spatial frequencies exceed unity for spatially correlated noise. Thus, consider the estimates of 2D DCT spectra obtained for real-life Sentinel SAR images. Examples of estimates for VV polarization component are represented in Fig. 2. Analysis clearly shows that speckle is really spatially correlated (the estimates for different VV SAR images are very similar). Thus, spectrum is practically of constant shape. Fig. 1. An example of Sentinel SAR images Not all people know well the peculiarities of 2D DCT spectra. These spectra can be calculated for blocks (fragments, areas) of different size. Since 8x8 pixels is the basic size of blocks used in many despeckling, we have used this size, i.e. indices of spatial frequencies are k 1,...,8; l 1,...,8. Small indices correspond to low spatial frequencies. Since DCT is orthogonal transform, it has uniform spectrum for spatially uncorrelated (white) noise. If power DCT spectrum is not uniform, then the noise is not white. If the basic part of power is concentrated in low frequencies, noise is spatially correlated [22]. Note that 2D DCT spectrum can be presented in normalized form. Then the values D ( k, l), k 1,...,8; l 1,...,8 have p Fig. 2. Examples of normalized DCT spectrum estimates for VV Sentinel SAR images. The spectrum estimates have been also obtained for VH polarization data. They are given in Fig. 3. As it is seen, they are similar to each other showing that spatial spectral properties are quite stable (do not vary a lot from one to another image). Secondly, they are the estimates similar to those ones obtained earlier for VV polarization (Fig. 2) and this is important for some of despeckling techniques analyzed below. Besides, images in polarization components are quite similar to each other. Cross-correlation factor for them (before denoising) is about 0.8, this property can be employed as well. 913

3 ( nm, ) I( nm, ) Fig. 3. Example of normalized DCT spectrum estimates for VH Sentinel SAR images. III. A NALYSIS OF DESPECKLING EFFICIENCY It is worth to recall that two-polarization radar images can be denoised in two basic ways: component-wise (separately) [8], [12] and in 3D (vectorial) manner [7], [23], [24]. Here we analyze the former approach since it is simpler (although less efficient). Certainly, it is possible to apply known despeckling techniques, in particular, those ones available in ESA SNAP toolbox. Below we also consider denoising methods based on DCT more in detail since they allow quite easy taking into account spatial correlation of the noise and its signal dependence. Spatial correlation is taken into consideration using frequency dependent thresholds. In turn, signal dependence (multiplicative nature) of the speckle can be easily taken into account using locally adaptive thresholds. In other words, local thresholds in nm-th block are determined as T n m k l n m D k l k l 0.5 (,,, ) (, ) p (, ), 1,...,8; 1,...,8 where ( nm, ) is local noise standard deviation for nm-th block that can be approximately calculated as where I( n, m ) denotes mean for nm-th block of original (unprocessed) image, is filter parameter usually set equal to 2.7 for hard thresholding. There are also other ways to cope with multiplicative character of the noise. It is possible to apply logarithmic type [8], [13], [21] homomorphic transform to original image with obtaining the image corrupted by additive but still spatially correlated noise, to denoise this image and then to apply inverse homomorphic transform with (possible) corrections of image values [21]. Note that DCT-based denoising can be implemented in several forms. The basic, sliding window, DCT denoising performs locally in blocks which are fully overlapping that is the neighbor block positions are shifted by only one pixel in horizontal or vertical direction. The new tendency is to use non-local approach [21] where, first, similar blocks are found and transformed to 3D data array. Second, these blocks are denoised together using DCT and other operations. Third, processed blocks are put to their original places and aggregated. Some problems with SAR images for this approach are with multiplicative nature of the speckle and its spatial correlation. They are solved by applying homomorphic transform [21], using specific metrics for searching similar blocks [25] and frequency dependent thresholding [13], [25]. Thus, it is, in general, possible to study three DCT-based approaches to component-wise processing. The first is sliding DCT filter (DCTF) with frequency dependent and locally adaptive thresholds [13]. The second approach presumes the use of logarithmic transform, sliding DCT-based denoising with frequency dependent thresholds and inverse homomorphic transform (HT) of exponential type. The third is similar to [21] and [25]. It presumes the use of logarithmic transform, the use of the modified BM3D (MBM3D) filter [25] with frequency dependent thresholds and inverse homomorphic transform. To start, below we analyze only the first approach (DCTF). Since we deal with real-life data, it is difficult to have quantitative criteria for comparison of filter efficiency. Therefore, we propose to compare the filters quality on crop classification map accuracy. The filter outputs are represented in Fig. 4. Comparison of crop classification maps obtained after different filters can be done by visual inspection of data shown in Fig. 5. IV. RESULTS We preprocessed time series of ten Sentinel-1A images with different filters available in ESA SNAP toolbox and DCTF. Comparison of those filters with no-filter image for 07 July 2016 is shown in Fig. 4. It is not easy to evaluate quality of filtering visually (user s opinions can be subjective). Therefore, we provide crop classification maps based on the same in-situ data and on the same time series of images. Classification was done using an ensemble of neural networks, namely, multilayer perceptrons (MLPs) [1], [26] [29]. During the neural network training, cross-entropy error 914

4 function was minimized. After classification, each neural network gave a posteriori probability of the input pixel belonging to each class. In an ensemble, we estimated the average a posteriori probability from all networks and assign to the pixel class with the highest probability. For training ensemble of neural networks, we collected 153 data samples for nine classes and all the accuracies were evaluated on independent set, that consists of 146 samples for nine classes. [30] from crop classification map using different filters for SAR data in The overall accuracy of the crop classification map without any filtering is 82.6%. The lowest accuracy among all the available filters is provided by Median filter, classification accuracy is higher by +3.2% compared to classification of original (non-filtered) data. The most accurate crop map was obtained based on images pre-processed by the DCTF. The gain of using this method is +5.1% compared to classification of original data. Important to emphasize that DCTF gains not only overall accuracy, but increases PA and UA for each class, excluding forest and bare land. Fig. 4. Example of using different filters applied to Sentinel-1A images: A) image without filtering; B) Median filter; C) Refined lee filter; D) DCTF. The obtained crop classification maps after applying different filters are shown in Fig.5. In Table 1, we present the comparison of user accuracy (UA), producer accuracy (PA), overall accuracy (OA) and kappa coefficient for all classes Fig. 5. Example of crop classification maps based on different filters for Sentinel-1A images: A) without filtering; B) Median filter; C) Refined lee filter; D) DCTF. TABLE I. Class COMPARISON OF USER ACCURACY (UA), PRODUCER ACCURACY (PA), OVERALL ACCURACY (OA) AND KAPPA COEFFICIENT FOR DIFFERENT FILTERS FOR SAR DATA IN 2016 Gamma No filter Boxcar Frost IDAN Lee Lee Sigma Median Refined Lee DCTF map UA, % PA, % UA, % PA, % UA, % PA, % UA, % PA, % UA, % PA, % UA, % PA, % UA, % PA, % UA, % PA, % UA, % PA, % UA, % PA, % Artificial Winter wheat Maize Sunflower Soybeans Forest Grassland Bare land Water OA, % / Kappa 82.6 / / / / / / / / / /

5 V. CONCLUSIONS In this paper, we used multitemporal SAR images from Sentinel-1A satellite for 2016 to estimate impact of different filters to crop classification accuracy. All available filters from ESA SNAP toolbox have been evaluated and compared with our method. All of them are useful for solving the applied problems, in particular, crop mapping. Overall accuracy without any filters is 82.6%. At the same time, the use of different methods for speckle reducing results in better classification maps - accuracy ranges from 85.8% to 87.7%. The most accurate and useful filter from ESA SNAP toolbox for crop mapping task is the Refined Lee one. However, the proposed DCT-based filtering approach has outperformed all available filters in ESA SNAP toolbox. Importantly, that this filter increases UA and PA for each class, excluding forest and bare land. This opens up a large number of new possibilities to integrate high-resolution SAR imagery in operational crop monitoring applications, food security analysis. REFERENCES [1] N. Kussul, S. Skakun, A. Shelestov, M. Lavreniuk, B. Yailymov, and O. Kussul, Regional Scale Crop Mapping Using Multi-Temporal Satellite Imagery, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XL- 7/W3, pp , DOI: /isprsarchives-XL-7-W [2] N. Kussul, A. Shelestov, and S. Skakun "Grid technologies for satellite data processing and management within international disaster monitoring projects," Grid and Cloud Database Management. Springer Berlin Heidelberg, pp , [3] A. N. Kravchenko, et al. "Water resource quality monitoring using heterogeneous data and high-performance computations," Cybernetics and Systems Analysis vol. 44, no. 4, pp , [4] A. Kolotii, N. Kussul, A. Shelestov, S. Skakun, B. Yailymov, R. Basarab, M. Lavreniuk, T. Oliinyk, and V. Ostapenko, Comparison of biophysical and satellite predictors for wheat yield forecasting in Ukraine, Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XL- 7/W3, pp , DOI: /isprsarchives-XL-7-W [5] Kussul, N., Skakun, S., Shelestov, A., & Kussul, O., The use of satellite SAR imagery to crop classification in Ukraine within JECAM project, IEEE International Geoscience and Remote Sensing Symposium (IGARSS), pp , [6] N. Kussul, G. Lemoine, F. J. Gallego, S. V. Skakun, M. Lavreniuk, and A. Y. Shelestov, Parcel-Based Crop Classification in Ukraine Using Landsat-8 Data and Sentinel-1A Data, IEEE J. of Select. Topics in Appl. Earth Observ. and Rem. Sens., vol. 9, no. 6, pp , [7] J.-S. Lee, E. Pottier, Polarimetric Radar Imaging: From Basics to Aplications, CRC Press, 2009, p [8] C. Oliver, S. Quegan. "Understanding Synthetic Aperture Radar Images, SciTech Publishing, 2004, p [9] G. M. Bakan, and N. N. Kussul "Fuzzy ellipsoidal filtering algorithm of static object state," Problemy Upravleniya I Informatiki (Avtomatika) vol. 5, no. 5, pp , [10] X. Kang, C. Han, Y. Yang, T. Tao, SAR image edge detection by ratiobased Harris Method, ICASSP 2006 Proceedings, vol. 2., pp , May [11] A. Naumenko, V. Lukin, Egiazarian K., SAR-image edge detection using artificial neural network, Proceedings of MMET 2012, Kharkov, Ukraine, pp [12] R. A. Touzi, Review of Speckle Filtering in the Context of Estimation Theory, IEEE Trans. on GRS., 2002, vol. 40, 11, pp [13] D.V. Fevralev, S.S. Krivenko, V.V. Lukin, R. Marques, F. de Medeiros, Combining Level Sets and Orthogonal Transform for Despeckling SAR Images, Aerospace Engineering and Technology, vol. 2, 99, 2013, pp [14] R. Marques, F. Medeiros, and D. Ushizima Target Detection in SAR Images Based on a Level Set Approach, IEEE Trans. on Systems, Man and Cybernetics, vol. 39, no. 2, pp , [15] O. K. Al-Chaykh, and R. M. Mersereau, Lossy compression of noisy images, IEEE Transactions on Image Processing, Vol. 7(12), pp , [16] S. Abramov, V. Abramova, V. Lukin, N. Ponomarenko, B. Vozel, K. Chehdi, K. Egiazarian, J. Astola, Methods for Blind Estimation of Speckle Variance in SAR Images: Simulation Results and Verification for Real-Life Data, Book Chapter in Computational and Numerical Simulations, ISBN , edited by Jan Awrejcewicz, InTech, Austria, 2014, pp [17] [18] P. Kupidura, Comparison of Filters Dedicated to Speckle Suppression in SAR Images, ISPRS-International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, pp , [19] G. Vasile, E. Trouvé, J. S. Lee, and V. Buzuloiu Intensity-driven adaptive-neighborhood technique for polarimetric and interferometric SAR parameters estimation, IEEE Transactions on Geoscience and Remote Sensing, vol. 44, no. 6, pp , [20] Y. Huang, and J. L. Van Genderen Evaluation of several speckle filtering techniques for ERS-1&2 imagery, International archives of photogrammetry and remote sensing, vol. 31, pp , [21] M. Makitalo, A. Foi, D. Fevralev, V. Lukin, Denoising of single-look SAR images based on variance stabilization and non-local filters, CD- ROM Proceedings of MMET, Kiev, Ukraine, Sept. 2010, pp. 1-4, [22] Lukin V., Bataeva E., Challenges in Pre-processing Multichannel Remote Sensing Terrain Images, Importance of GEO initiatives and Montenegrin capacities in this area, The Montenegrin Academy of sciences and arts Book. No 119, the Section for Natural Sciences Book No. 16, 2012, pp [23] D. Fevralev, V. Lukin, N. Ponomarenko, S. Abramov, K. Egiazarian, and J. Astola, Efficiency analysis of color image filtering, EURASIP Journal on Advances in Signal Processing, Vol. 2011:41, doi: / , [24] R. Kozhemiakin, V. Lukin, B. Vozel, K. Chehdi, Filtering of Dual- Polarization Radar Images Based on Discrete Cosine Transform, Proceedings of IRS, Gdansk, Poland, June 2014, pp [25] A. Rubel, V. Lukin, K. Egiazarian, Block Matching and 3D collaborative filtering adapted to additive spatially correlated noise, Proceedings of VPQM, Scottsdale, USA, Feb [26] S. Skakun, N. Kussul, A. Y. Shelestov, M. Lavreniuk, and O. Kussul, Efficiency Assessment of Multitemporal C-Band Radarsat-2 Intensity and Landsat-8 Surface Reflectance Satellite Imagery for Crop Classification in Ukraine, IEEE J. of Select. Topics in Applied Earth Obser. and Rem. Sens., vol. 9, no. 8, pp , [27] F. Waldner, et al. Towards a set of agrosystem-specific cropland mapping methods to address the global cropland diversity, International Journal of Remote Sensing, vol. 37, no. 14, pp , DOI: / [28] N. Kussul, N. Lavreniuk, A. Shelestov, B. Yailymov, and I. Butko, Land Cover Changes Analysis Based on Deep Machine Learning Technique, Jour. of Automation and Information Sciences, vol. 48, no. 5, pp [29] M. S. Lavreniuk, S. V. Skakun, A. J. Shelestov, B. Y. Yalimov, S. L. Yanchevskii, D. J. Yaschuk, and A. I. Kosteckiy, Large-Scale Classification of Land Cover Using Retrospective Satellite Data, Cybernetics and Systems Analysis, vol. 52, no. 1, pp , [30] R. G. Congalton, A review of assessing the accuracy of classifications of remotely sensed data, Remote sensing of environment, vol. 37, no. 1, pp ,