International Journal of Remote Sensing ISSN: 0143-1161 (Print) 1366-5901 (Online) Journal homepage: http://www.tandfonline.com/loi/tres20 A modified mean filter for improving the classification performance of very high-resolution remote-sensing imagery Lv ZhiYong, WenZhong Shi, Jón Atli Benediktsson & LiPeng Gao To cite this article: Lv ZhiYong, WenZhong Shi, Jón Atli Benediktsson & LiPeng Gao (2018) A modified mean filter for improving the classification performance of very high-resolution remote-sensing imagery, International Journal of Remote Sensing, 39:3, 770-785, DOI: 10.1080/01431161.2017.1390275 To link to this article: http://dx.doi.org/10.1080/01431161.2017.1390275 Published online: 23 Oct 2017. Submit your article to this journal Article views: 5 View related articles View Crossmark data Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=tres20 Download by: [University of Florida] Date: 25 October 2017, At: 02:38 INTERNATIONAL JOURNAL OF REMOTE SENSING, 2017 VOL. 39, NO. 3, 770–785 https://doi.org/10.1080/01431161.2017.1390275 A modiﬁed mean ﬁlter for improving the classiﬁcation performance of very high-resolution remote-sensing imagery Lv ZhiYonga, WenZhong Shib, Jón Atli Benediktssonc and LiPeng Gao d Downloaded by [University of Florida] at 02:38 25 October 2017 a School of Computer Science and Engineering, Xi’An University of Technology, Xi’An, China; bLand Surveying & Geo-Informatics Hung Hom, The Hong Kong Polytechnic University, Kowloon, Hong Kong; c Faculty of Electrical and Computer Engineering, University of Iceland, Reykjavik, Iceland; dSchool of Remote Sensing and Information Engineering, Wuhan University, Wuhan, China ABSTRACT ARTICLE HISTORY Very high resolution (VHR) remote-sensing imagery can reveal ground objects in great detail, depicting the colour, shape, size and structure of the objects. However, VHR also leads to a large amount of noise in the spectra, which may reduce the reliability of the classiﬁcation result. This article presents an extension of the mean ﬁlter (MF), which is named ‘modiﬁed mean ﬁlter (MMF)’, for smoothing the noise of VHR imagery. First, the MMF is a shapeadaptive ﬁlter that is constructed by gradually detecting the spectral similarity between a kernel-anchored pixel and its contextual pixels through an extension detector with eight neighbouring pixels. Then, because pixels of an objective are usually homogeneous with spatial continuity, the pixels located at the hollow area within an extended region are rectiﬁed to enhance the smoothing eﬀect. Finally, the spectral value of the kernel-anchored pixel is determined by the mean of the group of pixels within the adaptive region. Despite the proposed ﬁlter is a simple extension of MF, it has an advantage in preserving the edge between diﬀerent classes, and smoothing the noise of intra-class. The MMF approach is investigated through comparing with the classiﬁcation of VHR images based on ﬁlter processing, including the traditional mean ﬁlter (MF), the median ﬁlter (MedF) and the recursive ﬁlter (RF) which has been proposed for image classiﬁcation in Kang, Li, and Benediktsson (2014). The experimental results obtained by considering two VHR images show the eﬀectiveness of the proposed of MMF, which improves the performance of the classiﬁcation and implies more potential applications. Received 9 May 2017 Accepted 3 October 2017 1. Introduction At present, a large number of remote-sensing images with very high resolution (VHR) are available in urban areas. Compared with low or moderate resolution images, VHR images can depict ground object in more detail, such as shape, structure, size and texture. Therefore, these VHR images open new perspectives for remote-sensing applications CONTACT Lv ZhiYong Lvzhiyong_ﬂ[email protected] School of Computer Science and Engineering, Xi’An [email protected] Land Surveying & Geo-Informatics University of Technology, Xi’An, China; WenZhong Shi Hung Hom, The Hong Kong Polytechnic University, Kowloon HK 999077, Hong Kong © 2017 Informa UK Limited, trading as Taylor & Francis Group Downloaded by [University of Florida] at 02:38 25 October 2017 INTERNATIONAL JOURNAL OF REMOTE SENSING 771 such as in urban monitoring, environment assessment, and decision-making (Moser, Serpico, and Benediktsson 2013; Li, Zhang, and Zhang 2014). Among these applications, as is well known, many applications depend on the results of land cover classiﬁcation. However, the use of higher resolution images does not necessarily mean that higher classiﬁcation accuracies are obtained (Huang and Zhang 2013; Wilkinson 2005). The reason can be concluded brieﬂy as following: Despite an advantage is that VHR images can increase the correlation among pixels, which results in the VHR image having many spatial features that can be utilized for visual interpretation and classiﬁcation (Wang et al. 2012), due to the limitation of the remote-sensing technique, images with a high spatial resolution are typically limited to three or four spectral bands (worldview-2 has a 0.41 m resolution and eight spectral bands) (Gianinetto et al. 2014). Thus, the high spatial resolution and low spectral properties result in an increased intra-class variance and a decreased inter-class variance, which introduces spectral noise into the classifying thematic map and enhances the diﬃculty of separation (Blaschke 2010). For example, Ouyang et al. showed that as the spatial resolution of remote-sensing data increases, the spectral noise of the pixel-wise approach becomes more serious (Ouyang et al. 2011). Many factors, such as sensor and spatial uncertainty, also bring noise into the VHR image, and many pixel-based approaches are very sensitive to this noise (Huang and Zhang 2008; Duro, Franklin, and Dubé 2012). To reduce the noise and improve the classiﬁcation performance of VHR image, spectral-spatial method usually proposed for VHR image classiﬁcation. That is because spatial information in a remote-sensing image is inherited from the real world, especially in VHR images (Li et al. 2014). Therefore, spatial information can be employed to couple with the spectra to remedy for the spectral information’s insuﬃciencies, resulting in the smoothing of the noise and improved the accuracy of VHR image classiﬁcation. For example, Tarabalka, Benediktsson, and Chanussot (2009) proposed a new spectral-spatial classiﬁcation scheme for hyperspectral images. Huang et al. proposed a multi-feature model including spectra, structure and semantic features for VHR image classiﬁcation (Huang and Zhang 2013). Xia et al. proposed a new spatial-spectral classiﬁcation method to enhance the performance of hyperspectral images by integrating rotation forests and Markov random ﬁelds (Xia et al. 2015). In addition, spatial-spectral kernels have been investigated for the classiﬁcation of VHR images in many studies, such as morphological kernels and sparse kernels (Liu et al. 2013). To explore the spatial information and improve the performance of the VHR image classiﬁcation, mathematical morphology was applied and extended by increasing the size of the structuring element (Benediktsson, Pesaresi, and Amason 2003) and stacking diﬀerent shapes of structuring elements (Lv et al. 2014). Several advances of the spectral-spatial approach for classiﬁcation of high resolution images were summarized in previous research (Fauvel et al. 2013). However, the spatial-spectral feature-based approach relies on the performance of the spatial feature extraction technique. To obtain accurate spatial features, advanced feature extraction algorithms are usually required, but those processes may be timeconsuming and experience-dependent. Diﬀerent from spatial-spectral feature-based classiﬁcation approach described previously, image ﬁltering is also adopted to smooth the noise of the hyperspectral imagery with high resolution. For example, Angelos et al. developed an image classiﬁcation framework that was integrated with a nonlinear scale-space ﬁlter (Tzotsos, Karantzalos, Downloaded by [University of Florida] at 02:38 25 October 2017 772 Z. LV ET AL. and Argialas 2011). Recently, image ﬁlter has become a very hot topic in hyperspectral image processing and applied successfully in many applications (Kang, Li, and Benediktsson 2014). Although a ﬁlter is a powerful tool for denoising an image, the practical application of ﬁlters poses two problems: (1) most of the ﬁlters have been investigated on hyperspectral images, while VHR images have been more or less ignored and (2) many ﬁlters, such as the mean ﬁlter (MF) and median ﬁlter (MedF), are related to a regular window as the ﬁlter’s convolution. However, a single-size regular window may be unable to cover the multifarious spatial information of the various ground targets in an entire image. Therefore, it should be noted that the traditional MF is related to the spectral and regular domains alone, especially for VHR remote-sensing image. In this article, a modiﬁed mean ﬁlter (MMF) is proposed for reducing the noise and improving the performance of the VHR image classiﬁcation. Compared with MF, MedF and RF (Kang, Shutao, and Benediktsson 2014), the proposed ﬁlter runs through the whole VHR image through convolution in an adaptive manner. To verify the performance of the proposed approach, experiments were designed on the basis of a VHR false colour image of Pavia University and an IKONOS panchromatic image obtained from Reykjavik, Iceland. In these experiments, to demonstrate the advantage of the proposed MMF for VHR image classiﬁcation, the results of classiﬁcation based on the proposed MMF are compared with that of the unﬁltered process, the MF, MedF and RF. 2. New image denoising approach of an MMF for VHR images In this section, an image ﬁlter, called MMF, is proposed for reducing the noise of the VHR image. The MMF adds an extension to the traditional mean ﬁlter. However, unlike the single size and regular window of the mean ﬁlter, the shape and size of the convolution region of the MMF are extracted in a pixel-by-pixel manner, wherein the region of each pixel has a higher homogeneity. The shape of an adaptive region represents the contextual features surrounding a kernel-anchored pixel, and the size of the adaptive region is constrained by two thresholds in the spectral and spatial domains. For each single region extension, it seems like region growing algorithm. However, to our knowledge, it is the ﬁrst time that region growing is introduced for noise reduction of VHR image classiﬁcation. The proposed MMF contains three blocks, and they will be detailed as following: (1) extension of the MMF’s region; (2) rectiﬁcation-operation within the MMF’s region; and (3) calculation of the mean value of the labelled pixels. 2.1. Algorithm of region-extension for MMF An extension is used to detect adaptively the contextual features surrounding a pixel, an example of extension is shown in Figure 1. First, a kernel-anchored pixel (KAPi;j ) is a pixel of a VHR image at location: ði; jÞ. Then, in one chosen KAPi;j , the spectral diﬀerence between the KAPi;j and its eight neighbouring pixels is measured to decide whether the neighbouring pixel belongs to the homogeneous area around KAPi;j . The homogeneity of the pixel is deﬁned using the formula (1): Downloaded by [University of Florida] at 02:38 25 October 2017 INTERNATIONAL JOURNAL OF REMOTE SENSING 773 Figure 1. A ﬁltering example based on an MMF with an eight neighbouring extension detector. (a) An eight neighbouring extension detector, and (b) is an m n VHR remote-sensing image with R– G–B bands. (c) A sample extended region, where ‘X-X’ represents the detecting and extending order, with the ﬁrst number a KAP or a CAP and the second number the detecting order surrounding the current anchored pixel. Δs ¼ ðKAPÞi;j Psur ; (1) where Δs represents the spectral similarity between a KAPi;j and its surrounding pixels Psur . The greater Δs is, the greater diﬀerence between the KAPi;j and its surrounding pixels (Psur ) is, and the lower homogeneity between KAPi;j and Psur . And Psur is one of the eight neighbouring pixels, sur 2 ½0; 7. The shape and size of the region around a KAP is extended gradually if the following conditions are met. (1) Δs is less than a predeﬁned threshold T1 ; (2) The total number of labelled pixels that comprise the adaptive region is less than another predeﬁned threshold T2 . As shown in Figure 1, if the surrounding pixel Psur and KAPi;j met condition-1, Psur is labelled as a homogeneous pixel and stack it into the set of candidate-anchored pixels (CAP). A pixel in the set of CAPk is prepared as another level of anchored pixel for the region extension in the spatial domain, as the eight neighbouring pixels surrounding a CAP is compared with KAPi;j in the spectral domain. Therefore, this ensures that all of the labelled pixels are spectrally homogeneous, and the extension of the MMF region is adaptive in the spatial domain. In other words, CAP is used to extend the adaptive region in a recursive manner. Based on the aforementioned algorithm, the extension of the homogeneous region for an MMF will cease if either of the two conditions is not met. In this case, the extension of the KAP will be terminated, the value of the KAP is replaced with the mean of the pixels within the adaptive region and the algorithm will skip to the ith + 1 KAP. The whole image is scanned in this manner, and each pixel will be taken as once KAP. 774 Z. LV ET AL. Downloaded by [University of Florida] at 02:38 25 October 2017 2.2. Rectiﬁcation-operation within the MMF region It is worthy noting that when processing a VHR image using MMF, due to the spatial complexity of pixels or the spectral heterogeneity of the ground objects, a hollow contour of the region may be generated in the processing of an adaptive region-extension. For examples, two diﬀerent extensions are demonstrated in Figure 2. It can be found that the appendages of the buildings (such the pump-house and chimney on the top of building) or the sun’s height will all lead to noise in the VHR remote-sensing image. Moreover, this noise signiﬁcantly aﬀects the extension of the MMF region. If the current KAP is an objective pixel, the ﬁtted parameters (T1 and T2 ) may result in a hollow region of MMF, such as the second column in Figure 2. An object in an image scene, such as a building, a road, is described by a group pixels, which are similar in spectra and continuous in spatial domain. Thus, ideally, when a pixel of an object is taken as a KAP, the extension surrounds the KAP should be a solid region, besides the peculiar shape object (such as a circle shape meadow). However, the value of pixels for an objective object is usually heterogeneous for VHR remote-sensing image. The heterogeneity may cause a hollow region for an MMF’s extension, such as the extension demonstration in Figure 2. To enhance an objective KAP’s ﬁltering eﬀect and weaken the inﬂuence of the noise’s KAP, a rectiﬁcation strategy is proposed here. The strategy used here is to remove the noise pixels from inside a homogeneous objective area, which is deﬁned by (2): OðPi Þ ¼ l¼n 1X vl : n l¼0 (2) The above formula represents the operation to replace the hollow pixel within an adaptive region of MMF, where OðPi Þ is the rectiﬁed value of a noise pixel. A ‘noise pixel’ is the pixel which is within the adaptive region in spatial domain. Above, vl is the value of a pixel that is labelled as being in the extension of the MMF region, n is the total Figure 2. Examples of extension for an MMF and diﬀerent classes. INTERNATIONAL JOURNAL OF REMOTE SENSING 775 number of labelled pixels, and OðPi Þ is the rectiﬁed mean value of labelled pixels. The hollow pixels were rectiﬁed by the mean of the labelled pixels, which ensures that the pixels of the MMF region have a higher homogeneity. In other words, when the hollow pixels within the region of an MMF are ﬁlled, the adaptive region becomes a solid region, and all of the labelled pixels consisting of the adaptive solid-region are similar in terms of their spectral values. 2.3. Calculation of the mean value of the labelled pixels After determining the adaptive region of an MMF, the MMF is deﬁned as Downloaded by [University of Florida] at 02:38 25 October 2017 MMFðxi Þ ¼ l¼n 1X vl ; L l¼0 (3) where xi is took as a KAP, MMFðxi Þ is the ﬁltering process of the xi , L is the total number of the pixels within the adaptive region, and vl is the value of a pixel. The value obtained with KAP ﬁltering by its MMF is equal to the average mean of the labelled pixels within the adaptive region. It is worthy to note that the MMF diﬀers from the traditional MF in the following aspects: (1) MF simply replaces each pixel value in an image with the mean (‘average’) value of its neighbours. This has the eﬀect of eliminating pixel values that are unrepresentative of their surroundings. The mean ﬁlter is usually thought of as a regular ﬁlter, such as a 3 3 window or a 5 5 window. However, the MMF ﬁlters an image by replacing each pixel value with the mean of its adaptive region. The aim of the MMF is to describe the overall contextual features and let all pixels of an object have a higher similarity in the spectra. (2) MF yields to linear ﬁlters, and it can not only remove noise, but it also smooths the edges and boundaries and may ‘erase’ details whose size is not equal to the window size. As a result, an image ﬁltered by MF becomes ‘blurred’. However, the proposed MMF is adaptive in the spatial domain, and the ﬁltered value of a pixel depends on the spectral diﬀerence between a kernel-anchored pixel and its contextual pixels. Thus, MMF joints the spectral and spatial feature together. The extension of an MMF region is self-adaptive and constrained by two parameters: T1 and T2 . Hence, MMF not only smooths the intra-class noise, but it also retains the boundary between inter-classes. To illustrate the advantage of the proposed MMF, a VHR images is, respectively, ﬁltered with the MMF and the traditional MF, and the results are compared in Figure 3. The local variance of each band is compared using the same window for the ﬁltered image, ‘RGB-Var’ is referenced the variance value for Red, Green and Blue band respectively. A lower variance shows a higher homogeneity of the ground object. For example, as shown in Figure 3(a), the variance of the pixels which are within the red window for band-1, band-2, and band-3 is 98, 45, and 61, respectively. It can be found that the proposed MMF has an advantage in improving the homogeneity of the ground Downloaded by [University of Florida] at 02:38 25 October 2017 776 Z. LV ET AL. Figure 3. Filtering result comparison between the MF and MMF using a VHR image. (a) A raw VHR image with a 1.0 m resolution and three R–G–B bands; (b), (c) and (d) are the results processed by an MF with diﬀerent window sizes; (e), (f), and (g) are the results processed by the MMF with diﬀerent constraint parameters,T1 and T2 . object, and the MMF can also smooth the building lawns and retain the edges between them and their surroundings, as shown by the yellow arrow in Figure 3. 3. Experimental Two VHR images acquired by ROSIS-03 sensor and IKONOS-2 satellite were utilized to validate the feasibility and eﬀectiveness of the proposed MMF approach through classiﬁcation. Three parts were designed to achieve the aims. First, the images were described for each experiment. Second, experimental setup and parameter settings were presented. Finally, the compared results and discussion were given. 3.1. Data sets The ﬁrst data set is the ROSIS-03 Pavia University image scene with a 1.0 m spatial resolution. The original data set is 610 × 340 pixels, and a total of 12 bands were removed because of noise. A false colour composite of the image using channels 10, 27 and 46 for red, green, blue, respectively, is shown in Figure 4(a). The ground truth is shown in Figure 4(b). Nine classes of interest were considered in this article: trees, asphalt, bitumen, gravel, painted metal sheets, shadows, bricks, meadows, and soil. The second data set is an IKONOS-2 image from Reykjavik city, Iceland, which was used in the experiment to assess the proposed MMF. The IKONOS image is a highresolution panchromatic image with spectral coverage from 0.45 to 0.9 µm. The size of the image is 957 × 639 pixels, with a 1.0 m spatial resolution. Six classes were considered in this case: small buildings, open areas, shadows, large buildings, street and residential lawns. The original data and the available ground reference truth are shown in Figure 5. Downloaded by [University of Florida] at 02:38 25 October 2017 INTERNATIONAL JOURNAL OF REMOTE SENSING 777 Figure 4. False colour original image (a) and the ground reference data (b) of Pavia University. Figure 5. IKONOS-02 panchromatic VHR image from Reykjavik, Iceland and the available ground reference samples: (a) original image; (b) the ground reference map. 3.2. Experimental setup and parameter settings The ﬁrst experiment has two purposes, i.e. to test the eﬀectiveness of the proposed MMF in the classiﬁcation of a VHR remote-sensing satellite image and to explore the relationship between T1 , T2 ; and the accuracy of the classiﬁcation. The proposed MMF-based 778 Z. LV ET AL. Table 1. Training and test samples for the ROSIS-03 Pavia university image. Name Downloaded by [University of Florida] at 02:38 25 October 2017 Asphalt Meadows Gravel Trees Painted metal sheet Bare soil Bitumen Self-blocking Shadows No. of training pixels No. of ground reference pixels 603 412 182 382 46 512 189 414 88 6631 18,649 2099 3064 1345 5029 1330 3682 947 VHR image classiﬁcation was compared with that of the raw image and with other similar ﬁlters, such as MF, MedF, and RF. The training and test samples for each class are detailed in Table 1. The parameters for each approach are detailed as follows: Original image: Three bands, 10, 27, and 46, were taken from the original data as the false-colour image according to the guidance of ROSIS-03 sensor, and these bands were placed into the SVM classiﬁer as spectral features for classiﬁcation. The RBF kernel function was used, and the parameters were set by cross-validation. The obtained gamma parameter was 0.33, and the penalty parameter was 100.0. Original image ﬁltered by MF: the original image was smoothed using MF with a 5 5 window size. The processed spectral feature was used for classiﬁcation based on the SVM with the RBF, and the parameters (gamma 0.33, penalty parameter 100.0) of the SVM were acquired through cross validation (CV). Original image ﬁltered by MedF: As in MF, a window size of 5 5 was used to ﬁlter the original image based on MedF. Then, the ﬁltered bands were used for classiﬁcation using the SVM with the RBF. The gamma and penalty parameters of the SVM that were obtained through CV were 0.33 and 100.0, respectively. Original image ﬁltered by RF: the classiﬁcation map based on the original image using SVM is ﬁltered by RF which has been proposed in (Kang, Shutao, and Benediktsson 2014). The relative optimal parameters were δs ¼ 200; δr ¼ 20, and the number of iteration is 3. Original image ﬁltered by the proposed MMF: the proposed MMF was used to ﬁlter the original spectral bands. The parameters of the proposed MMF were set as follows: T1 ¼ 50 and T2 ¼ 100. After the original spectral bands were smoothed by the proposed MMF with the preset parameters T1 and T2 , the ﬁltered spectral bands were put into the SVM for classiﬁcation. The SVM classiﬁer with RBF was adopted, and the parameters were obtained through CV. The gamma parameter was 0.33, and the penalty parameter was 100.0. In the second experiment, an IKONOS-02 image with a very high spatial resolution from the city of Reykjavik, Iceland, was used in the experiment to demonstrate the proposed MMF is suitable for processing panchromatic image with high spatial resolution. The training samples and test pixels are detailed in Table 2. The training samples were selected randomly, and the available reference samples are shown in Figure 5. In addition, the parameters used in this experiment are detailed as follows: the window size adopted in MF and MedF is 5 5 pixels, and T1 and T2 are set, respectively, to 35 INTERNATIONAL JOURNAL OF REMOTE SENSING 779 Table 2. Training and test samples for the IKONOS Reykjavik panchromatic images. Name Small buildings Open areas Shadows Large buildings Streets Residential laws No. of training pixels No. of ground reference pixels 1526 7536 1286 2797 3336 5616 34,155 25,806 43,867 39,202 30,916 35,147 Downloaded by [University of Florida] at 02:38 25 October 2017 and 200 for the extension of the MMF region. δs ¼ 300 and δr ¼ 40 are adopted here for the ﬁlter RF. Finally, the parameters of the SVM with the RBF kernel were optimized through CV. 3.3. Experimental results In the ﬁrst experiment, the proposed MMF was evaluated and compared with other approaches, including the traditional MF, MedF, and RF. The classiﬁcation maps are shown in Figure 6, and the class-speciﬁc accuracies for diﬀerent approaches are compared in Table 3. From this table, it can be seen that the raw image ﬁltered by the proposed MMF exhibited an overall accuracy (OA) of 68.8% and a kappa coeﬃcient (κ) of 0.611. Compared with the raw image without any ﬁlter processing, the proposed MMF achieved a higher classiﬁed accuracy. Compared with the raw image processed using MF, MedF, and RF, the classiﬁcation accuracy of the proposed MMF is competitive in terms of OA and κ. The class-speciﬁc accuracies of the second experiment are shown in Table 4, and the visual classiﬁcation maps are shown in Figure 7. From the table and its corresponding classiﬁcation map, it can be found that (1) the VHR panchromatic image processed by a ﬁlter, such as MF, MedF, RF or the proposed MMF, can produce a better classiﬁcation map and higher classifying accuracies than that without any ﬁltering process and (2) comparing the proposed MMF-based classiﬁcation with those of MF, MedF and RF, the proposed MMF-based classiﬁcation achieved a higher classiﬁcation accuracy. Therefore, the proposed MMF can smooth the noise in the VHR panchromatic image, which is suitable for improving the performance of classiﬁcation. 4. Discussion To enhance the applicability of the proposed approach, the sensitivity between the parameters (T1 and T2 ) and the classiﬁcation accuracy was investigated. Parameter T1 indicates the spectral similarity between the kernel-anchored pixel (KAP) and its contextual pixels. Selecting a suitable T1 is the key to obtain a higher accuracy of classiﬁcation. If T1 is too small that will not smooth enough noise pixels, leading to the cessation of the extension of the region, whereas a T1 that is too large will remove inter-class noise and details, also leading to inaccurate inter-class information. Figure 8(a) shows that when the value of T1 ranges from 15 to 60, the accuracy of the proposed MMF-based classiﬁcation initially increases and then stabilizes. When T1 is large enough to extend the region and obtain the optimum accuracy, T2 will control the size of the extension. From Figure 8(b), it can be observed that when T1 is ﬁxed at 50, the value of T2 ranges Z. LV ET AL. Downloaded by [University of Florida] at 02:38 25 October 2017 780 Figure 6. Comparisons of classiﬁcation maps based on diﬀerent approaches for Pavia University VHR false colour image: (a) classiﬁcation map based on the raw image without any image processing; (b), (c), (d), and (e) are classiﬁcation maps based on the ﬁltered image using MF, MedF, RF, and the proposed MMF, respectively. (f) Legend of the map. from 15 to 700, but the accuracy, OA and κ, only range from 64.7% to 68.9% and 0.564 to 0.611, respectively. The reason for this is that a suitable value of T1 ¼ 50 has been set for constraining the extension. Therefore, when T1 is large enough, T2 will constrain the size of the extension and the classiﬁcation accuracy based on the proposed MMF will move towards a stable trend. Thus, T1 and T2 complement each other, in a practical application, T1 and T2 can be adjusted and determined in accordance with the diﬀerent image. INTERNATIONAL JOURNAL OF REMOTE SENSING 781 Table 3. Class-speciﬁc accuracy (%) for diﬀerent ﬁltering approaches in the SVM classiﬁcation of the Pavia University image. Information class Downloaded by [University of Florida] at 02:38 25 October 2017 Asphalt Meadows Gravel Trees Painted metal sheets Bare soil Bitumen Self-blocking bricks Shadows OA κ Raw image MF MedF RF  MMF 72.6 45.0 38.1 61.4 99.3 69.2 76.5 78.7 65.0 59.0 0.498 81.8 50.5 52.7 62.2 92.7 67.2 79.1 79.9 93.3 63.9 0.555 88.5 50.1 59.5 65.1 99.6 65.3 90.8 84.4 92.7 66.1 0.581 89.3 85.3 62.7 98.4 89.9 27.4 55.8 71.4 98.5 65.1 0.572 88.0 86.4 69.1 92.9 90.4 31.0 49.4 78.0 98.3 68.8 0.611 Table 4. Class-speciﬁc accuracy (%) for diﬀerent ﬁltering approaches in the SVM classiﬁcation of the IKONOS Reykjavik panchromatic images. Information class Small buildings Open areas Shadows Large buildings Streets Residential lawns OA κ Raw image 20.2 67.14 89.0 43.1 14.6 66.0 48.6 0.377 MF 31.9 60.3 84.2 35.6 41.4 61.1 51.6 0.412 MedF 22.0 60.4 90.8 43.4 51.3 57.0 53.1 0.431 RF  52.3 53.8 93.1 40.1 55.3 41.8 54.7 0.452 MMF 21.3 59.6 93.0 48.7 66.4 56.6 56.6 0.475 Overall, by comparing the result of the two experiments, it is clear that the proposed MMF-based approach is superior to the classiﬁcation of the raw images (without any image ﬁlter). Compared with the traditional MF, MedF, and RF, the proposed MMF is competitive for improving the classiﬁcation performance of VHR image. The discussion above reveals that the proposed MMF is an extension of the traditional MF, it is more eﬀective than the traditional MF for smoothing the noise of VHR images. Furthermore, it is clear that MMF has an advantage over the MF in preserving edge information. This is not only helpful for improving the accuracies of VHR image classiﬁcation but also has the ability to improve the visual performance of raw VHR images. In comparison to the traditional mean ﬁlter, the proposed MMF is competitive in terms of classiﬁcation accuracies when applied on VHR images. Currently, VHR images are used widely for land-cover classiﬁcation. As a novel and simple spatial ﬁlter, MMF may imply more potential applications in image processing. 5. Conclusion In this article, we propose a modiﬁed mean ﬁlter (MMF) for noise reduction of VHR remote-sensing image classiﬁcation. The proposed MMF progressively and adaptively extends the ﬁlter’s contextual shape from a kernel-anchored pixel (KAP) to a labelled pixel group whose members are spectral similar and spatial contiguously. The eﬀect of the proposed MMF is investigated by the classiﬁcation of VHR images in two experiments. The main contributions of this study can be summarized brieﬂy as follows: Downloaded by [University of Florida] at 02:38 25 October 2017 782 Z. LV ET AL. Figure 7. Classiﬁcation maps obtained based on diﬀerent image ﬁlters for the IKONOS-02-Reykjavik city image: (a), (b), (c), (d), and (e) are classiﬁcation map based on the raw image without any image processing, MF, MedF, RF, and MMF, respectively. (1) It is the ﬁrst time that region growing algorithm integrated with mean ﬁlter for noise reduction of VHR image classiﬁcation. Due to the beneﬁt of this integration, the proposed MMF is adaptive according to the shape, size and spectral diﬀerence of a considered ground object. Thus, the proposed MMF not only can reduce the noise of intra-class, but also can preserve the boundary of inter-classes. (2) The proposed MMF-based classiﬁcation system provides competitive classiﬁcation accuracies when compared with the MF, MedF and RF for VHR image classiﬁcation. In addition, the proposed MMF based classiﬁcation system performs better than the MF, MedF, RF in an experiment on the high spatial resolution panchromatic image. This indicates that the proposed MMF is not only eﬀective for a VHR false colour image, but also demonstrates robustness for a panchromatic image. Thus, the proposed MMF may imply more potential applications than the MF, MedF and RF. Downloaded by [University of Florida] at 02:38 25 October 2017 INTERNATIONAL JOURNAL OF REMOTE SENSING 783 Figure 8. Relation between the overall accuracy (OA) and the Kappa coeﬃcient (κ) against the spectral similarity (T1) and Region size (T2) for an MMF-based classiﬁcation of the Pavia University image: (a) T1 vs. OA and κ; (b) T2 vs. OA and κ. In theory, more adaptive region features can be explored than that of a single feature. Therefore, in future research, such additional features will be explored and investigated. In addition, the automation of parameters (T1 and T2) can be considered. If T1 and T2 are acquired in an automated manner by considering the spectral diﬀerence of the contextual pixels, then the result will be an image ﬁlter with a higher degree of automation. The advantage of such an approach may be more practicable than that of the proposed approach for VHR image classiﬁcation. In addition, with the fast development of remote-sensing technology, many VHR images are available conveniently, including the unmanned aerial vehicle (UAV) image, the proposed MMF may play an important role in image processing, such as the UAV images with very high spatial resolution. Acknowledgements The authors would like to thank the editor-in-chief, the anonymous associate editor, and the reviewers for their insightful comments and suggestions. This work was supported by National Natural Science Foundation of China (41331175 and 61701396), National Administration of Surveying, Mapping and Geoinformation, P.R. China (Technical Leading Talents) and the Hong Kong Polytechnic University (1-ZE24, 1-ZVF2), Shaanxi Natural Science Foundation (2017JQ4006) and the project from the China Postdoctoral Science Foundation (2015M572658XB). Disclosure statement No potential conﬂict of interest was reported by the authors. 784 Z. LV ET AL. Funding This work was supported by the National Natural Science Foundation of China [41331175,61701396]; National Administration of Surveying, Mapping and Geoinformation, P.R. China (Technical Leading Talents) and the Hong Kong Polytechnic University [1-ZE24, 1-ZVF2]; the China Postdoctoral Science Foundation [2015M572658XB]; and Shaanxi Natural Science Foundation [2017JQ4006]. ORCID LiPeng Gao http://orcid.org/0000-0003-0026-3719 Downloaded by [University of Florida] at 02:38 25 October 2017 References Benediktsson, J. A., M. Pesaresi, and K. Amason. 2003. “Classiﬁcation and Feature Extraction for Remote Sensing Images from Urban Areas Based on Morphological Transformations.” IEEE Transactions on Geoscience and Remote Sensing 41 (9): 1940–1949. doi:10.1109/ TGRS.2003.814625. Blaschke, T. 2010. “Object Based Image Analysis for Remote Sensing.” ISPRS Journal of Photogrammetry and Remote Sensing 65 (1): 2–16. doi:10.1016/j.isprsjprs.2009.06.004. Duro, D. C., S. E. Franklin, and M. G. Dubé. 2012. “A Comparison of Pixel-Based and Object-Based Image Analysis with Selected Machine Learning Algorithms for the Classiﬁcation of Agricultural Landscapes Using SPOT-5 HRG Imagery.” Remote Sensing of Environment 118: 259–272. doi:10.1016/j.rse.2011.11.020. Fauvel, M., Y. Tarabalka, J. A. Benediktsson, J. Chanussot, and J. C. Tilton. 2013. “Advances in Spectral-Spatial Classiﬁcation of Hyperspectral Images.” Proceedings of the IEEE 101 (3): 652–675. doi:10.1109/JPROC.2012.2197589. Gianinetto, M., M. Rusmini, G. Candiani, G. D. Via, F. Frassy, P. Maianti, A. Marchesi, F. R. Nodari, and L. Dini. 2014. “Hierarchical Classiﬁcation of Complex Landscape with VHR Pan-Sharpened Satellite Data and OBIA Techniques.” European Journal Remote Sens 47: 229–250. doi:10.5721/ EuJRS20144715. Huang, X., and L. Zhang. 2008. “An Adaptive Mean-Shift Analysis Approach for Object Extraction and Classiﬁcation from Urban Hyperspectral Imagery.” IEEE Transactions on Geoscience and Remote Sensing 46 (12): 4173–4185. doi:10.1109/TGRS.2008.2002577. Huang, X., and L. Zhang. 2013. “An SVM Ensemble Approach Combining Spectral, Structural, and Semantic Features for the Classiﬁcation of High-Resolution Remotely Sensed Imagery.” IEEE Transactions on Geoscience and Remote Sensing 51 (1): 257–272. doi:10.1109/ TGRS.2012.2202912. Kang, X., L. Shutao, and J. A. Benediktsson. 2014. “Feature Extraction of Hyperspectral Images with Image Fusion and Recursive Filtering.” IEEE Transactions on Geoscience and Remote Sensing 52 (6): 3742–3752. doi:10.1109/TGRS.2013.2275613. Li, J., H. Zhang, and L. Zhang. 2014. “Supervised Segmentation of Very High Resolution Images by the Use of Extended Morphological Attribute Proﬁles and a Sparse Transform.” IEEE Geoscience and Remote Sensing Letters 11 (8): 1409–1413. doi:10.1109/LGRS.2013.2294241. Li, M., S. Zang, B. Zhang, L. Shanshan, and W. Changshan. 2014. “A Review of Remote Sensing Image Classiﬁcation Techniques: The Role of Spatio-Contextual Information.” European Journal of Remote Sensing 47: 389–411. doi:10.5721/EuJRS20144723. Liu, J., W. Zebin, Z. Wei, L. Xiao, and L. Sun. 2013. “Spatial-Spectral Kernel Sparse Representation for Hyperspectral Image Classiﬁcation.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 6 (6): 2462–2471. doi:10.1109/JSTARS.2013.2252150. Lv, Z. Y., P. Zhang, J. A. Benediktsson, and W. Z. Shi. 2014. “Morphological Proﬁles Based on Diﬀerently Shaped Structuring Elements for Classiﬁcation of Images with Very High Spatial Downloaded by [University of Florida] at 02:38 25 October 2017 INTERNATIONAL JOURNAL OF REMOTE SENSING 785 Resolution.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 7 (12): 4644–4652. doi:10.1109/JSTARS.2014.2328618. Moser, G., S. B. Serpico, and J. A. Benediktsson. 2013. “Land-Cover Mapping by Markov Modeling of Spatial–Contextual Information in Very-High-Resolution Remote Sensing Images.” Proceedings of the IEEE 101 (3): 631–651. doi:10.1109/JPROC.2012.2211551. Ouyang, Z.-T., M.-Q. Zhang, X. Xie, Q. Shen, H.-Q. Guo, and B. Zhao. 2011. “A Comparison of PixelBased and Object-Oriented Approaches to VHR Imagery for Mapping Saltmarsh Plants.” Ecological Informatics 6 (2): 136–146. doi:10.1016/j.ecoinf.2011.01.002. Tarabalka, Y., J. A. Benediktsson, and J. Chanussot. 2009. “Spectral–Spatial Classiﬁcation of Hyperspectral Imagery Based on Partitional Clustering Techniques.” IEEE Transactions on Geoscience and Remote Sensing 47 (8): 2973–2987. doi:10.1109/TGRS.2009.2016214. Tzotsos, A., K. Karantzalos, and D. Argialas. 2011. “Object-Based Image Analysis through Nonlinear Scale-Space Filtering.” ISPRS Journal of Photogrammetry and Remote Sensing 66 (1): 2–16. doi:10.1016/j.isprsjprs.2010.07.001. Wang, L., Q. Dai, L. Hong, and G. Liu. 2012. “Adaptive Regional Feature Extraction for Very High Spatial Resolution Image Classiﬁcation.” Journal of Applied Remote Sensing 6 (1): 063506-1–16. doi:10.1117/1.JRS.6.063506. Wilkinson, G. G. 2005. “Results and Implications of a Study of Fifteen Years of Satellite Image Classiﬁcation Experiments.” IEEE Transactions on Geoscience and Remote Sensing 43 (3): 433–440. doi:10.1109/TGRS.2004.837325. Xia, J., J. Chanussot, D. Peijun, and X. He. 2015. “Spectral–Spatial Classiﬁcation for Hyperspectral Data Using Rotation Forests with Local Feature Extraction and Markov Random Fields.” IEEE Transactions on Geoscience and Remote Sensing 53 (5): 2532–2546. doi:10.1109/ TGRS.2014.2361618.