Removing Thick Clouds in Landsat Images

Transcription

1 Removing Thick Clouds in Landsat Images S. Brindha, S. Archana, V. Divya, S. Manoshruthy & R. Priya Dept. of Electronics and Communication Engineering, Avinashilingam Institute for Home Science and Higher Education for Women, Faculty of Engineering, Coimbatore, India sbrindha.raju@gmail.com, archanasubramanian55@gmail.com, divyavillalan@gmail.com, shruthy.selvaraj92@gmail.com, priyaramasamya01@gmail.com Abstract Landsat images provide the satellite observations of the earth s land surface. Due to the nature, Landsat images are highly affected by thick clouds. A new method based on a modified neighborhood similar pixel interpolator (NSPI) approach was developed for improving quality and availability of landsat images. It was developed for filling gaps due to Landsat ETM+ Scan line Corrector (SLC) off problem. The performance of this method was evaluated with both simulated and real cloudy images and compared with that of a contextual multiple linear prediction (CMLP) method. The results show that the modified NSPI approach can significantly diminish the edge effects by CMLP. The modified NSPI approach is more accurate than by CMLP, especially when the cloudfree auxillary and cloudy images are acquired from different seasons and have different spectral characteristics. Keywords Cloud removal, Spectro spatial information, spectro temporal information, Image restoration. I. INTRODUCTION Digital Image Processing is the use of computer algorithms to perform image processing on digital images. Digital Image Processing is used wider than Analog Image Processing as it can avoid build- up noise and signal distortion during processing. Landsat images are images of earth s land surface and coastal region taken by satellite. Cloud is the most common interferer in the satellite images [1].Cloud interference not only brings difficulties in processing the image for further use but also causes image recognition. Commonly cloud can be divided into two categories: Thin cloud and thick cloud. Thin cloud can be removed based on processing of spatial domain and frequency domain. Thick cloud removal methods mainly utilize overlapping images derived from the same region and different acquisition time. Therefore removing thick clouds in landsat images is necessary for improving their quality and availability. The general methods for removing thick clouds require clouds free images as auxillary images to restore spectral information blocked by clouds. E.H.Helmer and B.Ruefenacht proposed a strategy that uses regression tress and histogram matching method. In this method, regression trees are used to predict pixel values underneath the clouds. Then the histogram matching is applied to adjacent pixels. This method requires no manual interpretation. This method can support large volume processing. The histogram matching method just utilizes these common pixels as training data to build a rescaling model and then this model is employed to restore the value of the target pixel by only by using the value in the input image as model input.if the values of common pixel are within a small range, a linear regression model is difficult to build with high statistical significance [2]. Meng et al (CSF) method to replace spectral values of cloudy pixel by cloud free pixel using two conceptions, using location based one to one correspondence and spectral based closest fit. The location based one to one correspondence is applied to identify pixel with the same location in both cloudy image and auxillary image.the spectral based closest fit is applied to determine the most similar pixels in an image.the advantage of CSF technique is that it does not depend on the areas, the thickness, and the density of clouds and cloud shadows in the images. There should be no overlap of cloud pixels in the two images, if there are overlaps of pixels we need to select an additional auxillary image. Water movements have the spectral characteristic change of surface water so CSF may not work well [3]. Melgani introduced a contextual multiple linear prediction (CMLP) process for reconstructing for spectral values of cloudy area in Landsat Enhanced Thematic Mapper plus (ETM+) images and it was proved to be better than previous methods.there are two methods for the reconstruction of images. In the first method, the contextual prediction process is 121

2 implemented by means of an ensemble of linear predictors. In the second method, the local spectrotemporal relationships are reproduced by single non linear predictor [4]. The spatial continuity of ground features may not be preserved in the restored image with this method [5]. Recently a neighborhood similar pixel interpolator (NSPI) approach was developed to restore the values of unscanned pixel very accurately especially for heterogeneous landscapes and when there is a longer time interval between the input image and the target image.this approach can keep the spatial continuity of filled images even using the auxillary image with long time interval.nspi selects similar number of pixels over a certain sample size,which helps to create greater statistical reliability,especially in heterogeneous areas. We introduce a modified NSPI approach for removing thick clouds in landsat images.the performance of the approach is evaluated with both simulated and real clouds in landsat images. II. MODIFIED NSPI ALGORITHM The following Fig. 1 shows that flowchart of the modified NSPI approach. The cloudy image taken at the date t2 in which the cloudy parts are to be removed is initially taken. The cloud free pixel of the original image at t2 is taken for comparison at date t1. The cloudy area in the cloudy image which is taken at date t2 is marked.thick clouds will be brighter in visible band but colder in thermal bands when compared to landsat. In NSPI approach the adaptive window was used to determine the similar pixels but it is not suited for removing thick clouds. It is because here a large window is required to cover the large clouds. So spectral similarity criteria [6] have to be satisfied and those similar pixels are selected in modified NSPI approach. Cloudy free auxiliary image at t1 Cloudy image at t2 Spot the cloudy area in cloudy image Search for cloud free similar pixels Calculate the weight for each of the similar pixels Predict with spectro-temporal information Merge the two predictions to find the similar pixels Predict with spectrospatial information Fig. 1 : Flowchart of the modified NSPI approach. Fig. 1 : Representation of the similar Pixels selection. In the gap filling process, similar pixels are very close to the gap pixels because gaps are very narrow, so the range of spatial distance Di is comparable to the range of spectral distance RMSDi. However in the cloud removal process, the distance between a cloudy pixel and its similar pixels may vary greatly, which could make the range of spatial distances incomparable with that of spectral distances. Therefore, spatial distances (Di) and spectral distances (RMSDi) were both normalized and rescaled as follows: 122

3 Di* RMSDi* Where the subscripts min and max represent the minimum and maximum values, respectively; value 1 is an offset to define the relative importance of the two distances. In the NSPI method, weights are calculated using the two distance. These weights are used to combine the two predictions which may not be appropriate for the removal of clouds. It is because the assessability of local spatial homogeneity is hard. So the weight calculation is modified for the modified NSPI. If the target pixel is near the boundary of the cloud, then the first prediction is given larger weight based on the spectro-spatial information. It is used to keep the spatial continuity in the target image so that it reduces the edge effect. If the target pixel is located in and around the centre of the clouds, then the second prediction based on spectro-temporal information is used. Therefore, the weights combining the two predictions can be determined by the relative spatial distance from the target pixels to its similar pixels and to the cloud center. the modified NSPI approach, we use the similar neighborhood information so if the simulated images are in the edge of the subset images, then it will be rejected. Fig. 3(b) and (d) shows the example for the simulated images. In this case, Images taken on January 25 and August 5 were used as the cloud free auxiliary. This makes us to evaluate the effect of temporal interval between the cloudy images and the cloud free auxiliary. Wi= (1/(Di**RMSDi*))/ The center pixel of the cloud is calculated by taking average of coordinate of all cloudy pixels. Finally the value of the target pixel can be predicted as L(x,y,t2,b)=T1*L1(x,y,t2,b)+T2*L2(x,y,t2,b) Where x, y are the coordinates of cloudy pixel, t2 represent the time at which the image is taken, b is the band and r1,r2 represent the spatial distance between the target pixel and cloud center. Fig. 3 : Landsat 7 ETM+ images acquired on July 4, 2002 with NIR-red-green composites. (a) Hetrogeneous subset (c) Homogeneous subset. Figures (b) and (d) are the simulated cloudy images based on (a) and (b), respectively III. ALGORITHM TESTS Design of the experiment The modified NSPI approach is tested using the both simulated and real cloudy images. Three landsat 7 ETM+ images are acquired on January 25, July 4, and August 5 of 2002 in central Virginia. Two subset scenes that correspond to a heterogeneous landscape and a homogeneous landscape were extracted from the Landsat images for the analysis of the simulation. A Monte Carlo carried out a experiment in which for each subset images, 100 cloudy images were simulated based on the July 4 images. To create a single cloudy image, a cloud patch with a random size at a random location is simulated. The size of the simulated cloudy image is carried based on the power law distribution. In Fig. 4 (a) Landsat 5 TM image which is used in the real cloud removal test. (b) Restored cloud free image from the CMLP approach. (c) Restored cloud free image from the modified NSPI approach. Now consider the real thick clouds which are more complex than the simulated images. Landsat 5 TM real clouds are taken on July 24, 2008 from the place of Ohio. Here, a cloud free image is acquired on June 6, 2008 which is used as auxiliary image for the cloud removal process. CMLP method is also applied to both simulated and real cloudy images. The result of the CMLP is compared to the modified NSPI approach. This comparison is made 123

4 to find the accuracy of the modified NSPI. The result shows that the two methods were not independent. Results of the Simulated Clouds The parameter used in the modified NSPI is the minimum number of similar pixels (N).This N is calibrated using the trial and error procedure. The result shows that the reflectance prediction is more accurate when compared to n when it is larger. But the prediction will be accurate when N value is stable to 20. So for better efficiency, N value is set to 20. Cloud removal result can be explained using the Fig. 5.In this figure (a) and (b) are the images which are restored by the CMLP method.it has more edge effect which is shown in (c).in modified NSPI approach these edge effect is reduced to maximum extent which is shown in (f). The spectral difference in both approaches is similar and improved yet the spatial continuity is more in modified NSPI approach. The modified NSPI approach can perform better than CMLP approach while using a temporally father auxiliary image. So NSPI is far better than CMLP as it is usually difficult to find the auxiliary image that is much close to the cloudy image. Result of Real Cloudy Images Fig. 4 (b) and (c) shows the images of restored images from the CMPL approach and the modified NSPI approach. The result of both method can recover most of the images which are blocked by thick clouds contamination. The cloud free image restored by the modified NSPI approach has the less edge effect than the CMLP method. So edge effect removal places an important in the real cloudy image. method of the modified NSPI approach. (c) Zoomed image of the area marked in (b). (f) Zoomed image of the area marked in (e) respectively. IV. DISCUSSION ON THE RESULTS Comparing to CMLP method, modified NSPI approach can restore the image with less edge effect.next improved concept is the reflectance value. Here the reflectance value is more accurate than the CMLP method even when the cloud free auxiliary image and the cloudy image are obtained from different seasons and when it is having different spectral characteristics. The improved performance of the modified NSPI can be described in three aspects. First is the similar pixels are used to provide the information change between the cloudy and cloud free image. It provide effective radiometric difference between the images and it is effective when the time interval between cloudy and cloud free is more. so it is effective in mountainous areas. Second is the use of spectral information in a sufficient number of similar pixel, which is more reliability. But CMLP method uses only the corresponding pixel from the cloud free image to predict the pixel in the cloudy image. Third, the weighted combination of two predictions makes the spatial and radiometric continuity of the restored image according to the spectro spatial and spectro-temporal information. Table I Standard accuracy assessment of land cover classifications from just the gap areas.the first number relates to accuracies of land cover data generated from NSPI gap filled simulated SLC-off data,whereas the second refers to accuracies developed using the actual imagery. Number of assessment points was 206. Fig. 5 Restored images of the simulated clouds in fig. 3 (a) and (b) by the method of CMLP approach. (d) and (e) by the From the above table overall accuracy of modified NSPI and the real image is 92.7 and

5 V. CONCLUSION The CSF method uses the spectral closest cloud free pixel but the CMLP method use only the corresponding pixels. Modified NSPI provide similar pixel, usage of spectral information and weight combination. This make it a reliable method than other existing approaches. Hence the efficiency graph shows the reliability of various existing method in which the modified NSPI is improved because of less edge effect and exchange of adaptive window with the spectral similarity criteria. However there is a limitation for modified NSPI. The accuracy decreases slightly with the increases in size of cloud. This is due to the use of neighborhood information. When the size increases, then the average distance between target and similar pixel increases which reduce the prediction accuracy. Fig. 6 : Comparison of various existing methods with modified NSPI approach in efficiency. Next, in areas with more clouds it is difficult to get the cloud free auxiliary images using a single sensor. We have modified the approach by taking multiple image of the same landscape. It is made possible by using different sensor using a reference based approach[13] to obtain the cloud free auxiliary image for the cloud removal process.through this modification accuracy is increased in our image. VI. REFERENCES [1] J.C.Ju and D.P.Roy, The availability of cloud free Landsat ETM plus data over the conterminous United States and globally, Remote sens. Envion.,vol.112,pp ,2008. [2] E.H.Helmer and B.Ruefenacht, Cloud free satellite image mosaics with regression trees and histogram matching, Photogramm.Eng.Remote Sens.,vol.71,no.9,pp ,Sep [3] Q.M.Meng,B.E.Borders, C.J.Cieszewski, and M.Madden, Closest spectral fit for removing clouds and cloud shadows, Photogramm.Eng.Remote Sens.,vol.75,no.5,pp ,May [4] F.Melgani, Contextual reconstruction of cloud contaminated multitemporal multispectral images, IEEE Trans.Geosci.Remote Sens., vol.44,no.2,pp ,feb [5] S.Benabdelkader and F.Melgani, Contextual spatiospectral postreconstruction of cloudcontaminated images, IEEE Trans. Geosci. Remote Sens.Lett.,vol.5,no.2,pp ,Apr [6] J. Chen, X. Zhu, J. E. Vogelmann, F. Gao, and S. Jin, A simple and effective method for filling gaps in landsat ETM+ SLC-off images, Remote Sens. Environ., vol. 115, no. 4, pp , Apr [7] F.Geo, J.G. Masek, R. Wolfe, and C. Huang, Building consistent medium resolution satellite data set using moderate resolution imaging spectroradoimeter products as reference, J.Appl.Remote Sens., vol.4,p ,Apr.2010.DOI: /1: