Selective Thematic Information Content Enhancement of LANDSAT ETM Imagery

Transcription

1 Remote Sensing in Earth Systems Sciences (2018) 1: ORIGINAL PAPER Selective Thematic Information Content Enhancement of LANDSAT ETM Imagery George Ch. Miliaresis 1 Received: 1 June 2018 /Revised: 15 September 2018 /Accepted: 23 October 2018 /Published online: 6 November 2018 # Springer Nature Switzerland AG 2018 Abstract A selective variance reduction methodology is presented that reduces the band-to-band correlation observed in Landsat ETM imagery allowing a thematic-based de-correlation stretch. The bands 1 to 5, and band 7 are transformed to principal components (PCs). PC-1 and PC-2 account for the 94.7% of the total variance evident in images. In the current case study, band 3 and band 4 are selected to predict PC-1 and PC-2, respectively, through linear regression models. The PC-1 and PC-2 predicted image accounts for the 94.3% and 91.1%, respectively, of the total variance evident in the regression models. The bands 1 to 5 and 7 images are reconstructed from the two PC-1 and PC-2 residual images as well as PC-3 to PC-6 images. Thus, a thematic-based (band-dependent) decorrelation stretch of ETM imagery is achieved allowing the elimination of variance components that are related to spectral signature similarity of landcover types evident in Central Valley (California). The reconstructed imagery, a new higher order Landsat product, will assist image analysis, photo-interpretation, agricultural terrain analysis, mapping applications, and environmental monitoring at global level. Keywords De-correlation stretch. Landsat imagery. Landcover. Mapping. California 1 Introduction Landsat imagery provides a global record of the Earth s terrestrial, coastal, and polar regions at moderate spatial resolution since 1972 [1], while since 2012, the new Landsat-8 satellite [2], a National Aeronautics and Space Administration (NASA) and United States Geological Survey (USGS) collaboration, extends the Landsat record to nowadays. The imagery in the visible, near-infrared, short wave, and thermal infrared [3] acquired by the Thematic Mapper (TM), the Thematic Mapper Plus (ETM), and Operational Land Imager (OLI) sensors carried by the series of Landsat satellites constitutes the most consistent record of land-surface conditions by remote sensing techniques [4]. The availability for free to the public of a such a long-term data record [5] establishes the need for higher-level ETM products at moderate resolution that will be the outcome of new data processing techniques, supporting * George Ch. Miliaresis miliaresis.g@gmail.com; george.miliaresis@ouc.ac.cy 1 Environmental Conservation & Management, Open University of Cyprus, Latsia, Cyprus both global change and food security studies [6], as well as resource management [7]. While sensors like ETM may have responses with about 1% overlap between adjacent bands, the newest sensors like MODIS, AVIRIS, and HYPERION can have 10% or more overlap among certain spectral bands due to both technological and hardaware limitations [8]. Even through the sensor bands responses do not overlap spectrally, the outputs of the bands are usually correlated due to the spectral signature similarity of the objects in the scene [8]. In the case of the ETM, the imagery suffers mostly from a scene-based significant band-to-band correlation [9] resulting from (a) the spectral signature similarity of certain landcover and lithologic types given the spectral resolution (broad spectral bands) of ETM sensor (Table 1), (b) the effect of topography given the sun illumination conditions during the image acquisition, as well as (c) the atmospheric conditions, and (d) the sensor geometry (for example the stripping effect). In general, the band outputs of an ETM sensor are described with n- dimensional orthogonal vectors, n being the number of bands [11]. Nevertheless, the traditional orthogonal (Kartesian) vector model fails when high levels of correlation do exist in between bands [11].

2 54 Remote Sens Earth Syst Sci (2018) 1:53 62 Table 1 Landsat 7 ETM, band wavelengths, and suggested uses for each band Band Wavelength (μm) Useful for mapping 1 Blue 2 Green 3 Red 4 Near infrared 5 Short infrared 7 Short infrared Bathymetric mapping, distinguishing soil from vegetation, and deciduous from coniferous forest Emphasizes peak vegetation which is useful for assessing plant vigor Discriminate vegetation slopes Emphasizes biomass content and shorelines Discriminate moisture contents of soil and vegetation Hydrothermally altered rocks associated with mineral deposits From [10] Towards this end, de-correlation stretching enhances the separation of a multispectral image with significant band-toband correlation [9]. De-correlation stretching requires the bands to be subjected to a principal components analysis (PCA). PCA rotates the axes of the data space so that they become oriented with the orthogonal principal component (Pcs) axes, the directions of maximum variance in the data [12]. Then, data are stretched within the PC space, and the resulting images are transformed back to the original space [11]. Decorrelation stretching ensures that the effects of stretching are independent for each band since PCs are uncorrelated. The widespread acceptance of the de-correlation stretch method is shown by its consideration as a standard product for data collected by the ASTER sensor carried by the Terra satellite [11]. Nevertheless, similarity in-between the spectral signatures of landcover types, the effect of topography (cast shadows, etc.), the systematic and non-systematic noise evident, etc., are still present in the reconstructed images. Some of these effects (for example non-systematic noise) might be partially removed if the PC band/bands with the least explained variance are not transformed back to the original space [11]. Miliaresis [13] presented a significant improvement to de-correlation stretch method since certain types of variances that are associated to certain variables are selectively removed. His selective variance reduction (SVR) method [13] usesmultiplelinearregressionofcertainpcsinorder to quantify and subtract the variance explained by elevation, latitude, and longitude of multi-temporal (a) land surface temperature (LST) imagery [14] and(b) precipitation imagery [15]. In this approach, the reconstructed imagery is derived from the multiple linear regression residual images of selected PCs by considering the portion of variance that it is not related to elevation, latitude, and longitude. The SVR method is proved particularly successful (a) in mapping seismo-tectonic thermal anomalies in Arabian Peninsula and Iran by the spatial standardization of multitemporal LST imagery [14], and (b) in associating the temporal fluctuations in the reconstructed LST imagery to the amount and the spatial extent of precipitation events in California and Nevada [16]. The SVR method is implemented in Python, and it is available for free under the GNU General Public License v3.0 through the GitHub software repository [17]. The main consideration being that the selected independent variables (for example elevation, latitude, and longitude) should construct valid multiple linear regression models for the dependent PC variables. A major limitation of remote sensing in biophysical modeling-capturing the diurnal temperature range (DTR) with global datasets is also addressed by SVR in the frequency domain [18]. The bottom line being DTR enhancement from Aqua and Terra MODIS data sets is performed in two steps: (a) the multi-temporal data set is filtered and the inverse Fourier transform considers the harmonics with period less than or equal to a day, to reconstruct a new image time series, and (b) clustering of the reconstructed data identified the Fig. 1 SRTM DEM of the study area. Elevation is within the range 32 to 1111 m (the darker a grid point, the greater its elevation)

3 Remote Sens Earth Syst Sci (2018) 1: Fig. 2 The Landsat ETM bands 1 to 5 and 7 of the study area spatial and the temporal pattern of regions with common DTR variability, allowing environmental terrain characterization, and agricultural planning in Greece [18]. The aim of this research effort is to develop a SVR methodology that reduces the significant band-to-band correlation observed in ETM imagery. Such an effort should derive a thematic (according to the ETM suggested uses per band)-based decorrelation stretch of ETM imagery, allowing the elimination of variance components that are related to spectral signature similarity of both landcover and lithologic types as well as topographic effects and ETM sensor-induced noise. Hopefully, the reconstructed imagery, a new higher order Landsat product, will assist image analysis, photo-interpretation, agricultural terrain analysis, mapping applications, and environmental monitoring at global level. Fig. 3 The histogram frequency distribution per ETM band

4 56 Remote Sens Earth Syst Sci (2018) 1:53 62 Table 2 Cross correlation matrix, as well as band mean correlation to other bands (band 4 is excluded) Bands Mean correlation Methodology The SVR method [13, 16, 17] includesfoursteps: 1. Computation of unstandardized PCs 2. Selection of independent variables 3. Regression of PC-1 and PC-2 versus the independent variables 4. Image reconstruction from the two residual images of PC- 1 and PC-2 regression models as well as the remaining PCs The selection of independent variables depends on the multi-dimensional imagery, and so first, the study area is presented and the data is interpreted. 2.1 Study Area and Data The study area is bounded by longitudes to (W) and latitudes to (N). It includes the Northern part (Sacramento Valley) of the Central Valley that is a flat valley that dominates the geographical center of the US state of California [19]. The Sacramento River, along with its tributaries, flows southwards through the Sacramento Valley for about 700 km. The Central Valley is one of the world s most productive agricultural regions [19, 20]. The SRTM ver 3. digital elevation model (DEM) with spacing 1 arc/s [21] of the study area, a latitude-longitude grid referenced to WGS-84 ellipsoid, is presented in Fig. 1. Sutter Buttes (almost in the center of the DEM in Fig. 1) is the one notable exception to the flat valley floor, representing the remnants of an extinct volcano 70 km north of Sacramento. The flatness of the valley floor contrasts with the Sierra Nevada on the east and the California Coast Ranges on the west. A Tri-Decadal Global Landsat Orthorectified Enhanced Thematic Mapper Plus [1] image with scene ID L _ is used. The image was acquired on 9th of July 2000 and has been registered to the UTM Zone 10 North and referenced to WGS84 horizontal datum while spacing equals to 28.5 m. Sun azimuth and elevation equals to and during image acquisition. All image bands have been individually resampled, using a nearest neighbor algorithm, and this data product provides systematic radiometric and geometric accuracy [1]. The bands 1 to 5 and 7 are used in this study (Fig. 2), while suggested uses for each band are provided in Table 1 [10]. Each band of the study area consists of 3501 columns and 3401 rows. Frequency histograms for each band are depicted in Fig SVR for Landsat ETM Imagery The cross-correlation matrix (Table 2) quantifies that the visible, near-infrared, and mid-infrared ETM bands (Table 1) are highly correlated, and so, there is redundant information within the bands. Adjacent bands in the spectrum (Table 1)present the highest correlations (Table 2) and that implies that the redundant information is related more to the similarity of the spectral signatures of the landcover types in the study area, given the spectral resolution of ETM. Near-infrared (NIR) band 4 (Table 1) is an exception since negative correlations are observed (Table 2). In the NIR region of the spectrum, very little energy is absorbed by vegetation, reflection increases greatly and makes images of vegetation very bright at NIR wavelengths [9]. The band mean correlation to other bands (with band 4 is excluded Table 3 Principal components (PCs), eigenvectors, eigenvalues, and percent variance explained per PC Principal components PC-1 PC-2 PC-3 PC-4 PC-5 PC-6 Eigenvectors Eigenvalues Percent variance (%)

5 Remote Sens Earth Syst Sci (2018) 1: Fig. 4 The PC images of the ETM imagery due to negative correlations) is presented in Table 1 and indicates that band 3 presents the highest mean correlation to other bands. Unstandardized PCs are computed, and the eigenvalues, eigenvectors, and the PCs images are presented in Table 3 and Fig. 4, respectively. The first two PCs account for the 94.7% (Table 3) of the variance evident within the ETM bands. Independent variables, for example elevation, slope, aspect, latitude, longitude, etc., should be defined in an attempt to quantify their contribution to PC-1 and PC-2 with empirical models based on multiple linear regression analysis. Unfortunately, the multiple linear regression models based on these variables are not statistically significant. The spectral signature similarity of landcover types seems to be the key issue. Thus, if the bands with the highest mean correlation to other bands (Table 2) are considered as the independent variables, then SVR should reduce the redundant information among ETM bands. Band 3 presents the highest mean band correlation to other bands (Table 2). The linear regression of PC-1 versus band 3 (B3) is presented in Eq. 1. PC 1 ¼ 1:8563 B3 þ 27:5144 ð1þ The ANOVA table verifies the statistical significance of Eq. 1 (Table 4). R 2 that represents the extent of variability in the dependent variable explained by all the independents variables [12], equals to Thus, the 94.3% of the variance is explained by Eq. 1. The F test value indicates the overall significance of the regression model, whether or not the independent variables taken jointly contribute significantly to the prediction of the dependent variable [12]. For Eq. 1, the F-test value 195,886,048 (Table 4) exceeds the F-critical value at 99% confidence interval for (1, 11,906,899) degrees of freedom, and hence, the overall regression is significant. The coefficient and the constant term express the individual contribution Table 4 ANOVA Table of the linear regression for Eq. 1 Source Degrees of freedom Sum of squares Mean square F-test Regression 1 49,167,997,303 49,167,998, ,886,048 Residual 11,906,899 2,988,668, Total 11,906,900 52,156,665,513 Regression statistics: R =0.971andR 2 = Individual regression Degrees of freedom Independent variables t test coefficients 11,906,899 Band 3 13,996 Constant term 2419

6 58 Remote Sens Earth Syst Sci (2018) 1:53 62 Table 5 ANOVA table of the linear regression for Eq. 2 Source Degrees of freedom Sum of squares Mean square F-test Regression 1 6,167,220,086 6,167,220, ,728,232 Residual 11,906, ,249, Total 11,906,900 6,770,469,352 Regression Statistics: R =0.9544andR 2 = Individual regression Degrees of freedom Independent variables t test coefficients 11,906,899 Band 4 11,033 Constant term 2614 of the independent variable to PC-1. The significance of the coefficient is expressed in the form of a t-statistic that verifies the significance of the variables departure from zero [12]. The absolute values of the t test (Table 4) for the band 3 coefficient and the constant term exceed the t-critical value at 99% confidence level with 11,906,899 degrees of freedom, and hence, the independent variable and constant term are significant. The multiple linear regression of PC-1 versus band 4 (B4) is presented in Eq. 2. R 2 equals to (Table 5). Thus, the 91.09% of the variance is explained by Eq. 2. PC 2 ¼ 0:9479 B4 þ 20:6076 ð2þ The ANOVA table verifies the statistical significance of Eq. 2 (Table 5). The F-test value 121,728,232 (Table 5) exceeds the F-critical value at 99% confidence interval for (1, 11,906,899) degrees of freedom, and hence, the overall regression is significant. The absolute values of the t test (Table 5) for the band 4 coefficient as well as the constant term exceed the t-critical value at 99% confidence level with 11,906,899 degrees of freedom and hence the independent variables and constant term are significant. The images that correspond to the residuals and the predicted images for Eqs. 1 and 2 are presented in Fig. 5. The residual images visualize the spatial distribution of the unexplained variance of the multiple regression models. In the current implementation, the residual images (Fig. 5b, d) as well as the PC-3 to PC-6 (Fig. 4) are used for the reconstruction of the ETM bands. The reconstructed bands are presented in Fig. 6. Frequency histograms for the reconstructed bands are depicted in Fig. 7. Mean(μ), standard deviation (s) per band, and S-ratio (the ratio of the s of the reconstructed band versus the s of the initial band, assessing the variance reduction in the Fig. 5 Predicted and residual images of the linear regression models. a, b Predicted PC-1 and residual PC-1 images respectively for linear regression (Eq. 1). c, d) Predicted PC-2 and residual PC-2 images respectively for the multiple linear regression (Eq. 2)

7 Remote Sens Earth Syst Sci (2018) 1: Fig. 6 The reconstructed bands 1 to 5 and 7 of the study area reconstructed bands) as well as μ-ratio (the ratio of the difference of initial band μ minus the reconstructed band μ to the initial band μ) are presented in Table 6. 3 Discussion of the Results The discussion aims (a) to justify the bands selection included in the two linear regression models (Eqs. 1 and 2), (b) to quantify the variance reduction, and (c) to assess the thematic information content of the reconstructed imagery. Unstandardized PCs [11] are computed; thus, the variance co-variance matrix, instead of the correlation matrix, is used for the computation of PCs (Fig. 4). Since a major amount of redundant information is subtracted according to the two regression models (Tables 4 and 5), the frequency histogram distributions (Fig. 3) should present a major decrease in the variance. That is verified in Fig. 7, as the histograms of the reconstructed images indicate. The reconstructed band 3 histogram present Fig. 7 The histogram frequency distribution per reconstructed band

8 60 Remote Sens Earth Syst Sci (2018) 1:53 62 Table 6 Band mean (μ), st. dev (S), S-ratio, and μ-ratio Band Landsat imagery Reconstructed imagery S-ratio μ-ratio μ s μ s the greatest decrease in variance (Fig. 7), since band 3 presents the greater mean correlationtoother bands(table 2) and band 3 do participate in the regression model (Eq. 1). Negative values are observed in the domain of the histograms of reconstructed images (Fig. 7). The scaling of the reconstructed images back to the range 0 to 255 is logically simple and involves calculating the maximum and minimum of the reconstructed values, then scaling this range onto the scale [11]. Scaling is avoided, not to distort the pixel values of the reconstructed bands expressing the deviation from the pixel values of the initial ETM bands if the redundant information is subtracted. Band 3 is selected as independent variable in Eq. 1 because it presents the maximum mean correlation to other bands accordingtotable2, while the R 2 of the regression model (Table 4) is maximized. In addition, band 4 is the independent variables in Eq. 2. Its selection also maximize the R 2 of the linear regression model (Table 5). Band mean (μ), st. dev (s), S-ratio, and μ-ratio are presented in Table 6. The greatest the variance reduction of the reconstructed imagery, the less the S-ratio, while the greatest the μ-ratio, the greatest reduction of μ of the reconstructed imagery. The variance reduction of the reconstructed Band 3 image is maximum, and thus, S-ratio is minimum (Table 6). This is also valid for reconstructed band 4 image that presents the second in order minimum S-ratio (Table 6). So, the bands participating in the linear regression models do present the greatest variance reduction. On the other hand, band 2 presenting the third in order (Table 2) greater mean band correlation to other bands (Table 2) do also present the third in order greater variance reduction (Table 6). The μ-ratios (Table 6) indicate the relative shift of the frequency distributions of the reconstructed images (Fig. 7) due to redundant information reduction from initial frequency distributions (Fig. 3). In addition, the interpretation of frequency histogram distributions of the reconstructed images (Fig. 7) indicates that they differentiate more from one another (on the basis of mean and st. dev.) in comparison to the initial bands frequency distributions (Fig. 3). This observation verifies the removal of redundant information. In order to assess the thematic information content, K- Means cluster analysis [11] is applied, assuming that eight classes are evident in the reconstructed images (Fig. 6). Note that the reconstructed images are not standardized to present mean 0 and standard deviation to 1. The cluster centroids are presented in Fig. 8 and in Table 7, while the spatial extent of each cluster is presented in Fig. 9. The clusters map (Fig. 9) allows the interpretation of the remnants of the extinct volcano Sutter Buttes (Fig. 1), and the agricultural regions in Central Valley from the natural cover of both the Sierra Nevada on the east and the California Coast Ranges on the west (Fig. 1). The centroid coordinates (Fig. 8) indicate that the unstandardized band K-means clustering is biased from reconstructed bands 1, 2, 5, and 7 (presenting the highest absolute mean centroid coordinates). In this context, bands 3 and 4 (participating in the linear regression models) do present the minimum absolute mean centroid coordinate (almost equal to zero) for all clusters. The differentiation in between clusters is maximized for bands 5 and 7 since the st. dev. of the reconstructed band coordinates 5 and 7 are maximized according to Table 7 and Fig. 8. So, if unstandardized reconstructed images are used in clustering, then the cluster map observed in Fig. 9 is biased from the four bands that do not participate in the two regression models. In this context, SVR of ETM imagery might be used for thematic context variance reduction according to user needs and the problem under consideration. Although the reconstructed bands 3 and 4 do present the lowest absolute mean values per image band (Table 6) and Fig. 8 The eight cluster centroids

9 Remote Sens Earth Syst Sci (2018) 1:53 62 Table 7 Cluster centroids, number of pixels, and percent occurrence per cluster, as well as st. dev. per cluster coordinate (reconstructed band) Cluster 61 Centroids coordinates for the reconstructed bands Occurrence ,988, , , ,566,004 3,179, ,849,565 1,200, % , st. dev ,906, cluster (Fig. 8, Table 7), they should include non-redundant information in relation to the other reconstructed bands (1, 2, 5, and 7). Thus, first, the thematic information content of reconstructed bands 3 and 4 is examined by the ratio of reconstructed band 4 to the reconstructed band 3 in Fig. 10. The interpretation of Fig. 10 does allow to identification of the agricultural patterns in the Central Valley and the natural vegetation cover on the mountainous terrain surrounding the valley. Thus, in another context and in order to give a proper weight to reconstructed bands 3 and 4 during the clustering process, the six reconstructed bands 1 to 5 and 7 should be standardized to present mean value 0 and st. dev. equal to 1. Fig. 9 The cluster spatial distribution Pixels 4 Conclusion The presented SVR methodology reduces the band-to-band correlation observed in Landsat ETM imagery. The bands 1 to 5, and band 7 are transformed to principal components, and PC-1 and PC-2 are found to account for the 94.7% of the total variance evident in images. Then, band 3 and band 4 predict PC-1 and PC-2, respectively, through linear regression models. PC-1 and PC-2 predicted images maximize the interpreted variance. The bands 1 to 5 and 7 images are reconstructed from the two PC-1 and PC-2 residual images as well as PC-3 to PC-6 images. Fig. 10 The image ratio of the reconstructed band 4 to the reconstructed band 3. The whiter a pixel, the greater its band ratio

10 62 Remote Sens Earth Syst Sci (2018) 1:53 62 Variance statistics of the reconstructed images as well as clustering verify the redundant variance reduction that is related to spectral signature similarity of landcover types evident in the study area. A thematic information content variance reduction (band 3 and band 4 thematic information contend elimination is presented in this case study) is achieved and mapped through the un-standardized reconstructed images used in the clustering process. Publisher s NoteSpringer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Tatem A, Nayar A, Hay S (2006) Scene selection and the use of NASA s global orthorectified Landsat dataset for land cover and land use change monitoring. Int J Remote Sens 27: Roy D, Wulder M, Loveland T et al (2014) Landsat-8: science and product vision for terrestrial global change research. Remote Sens Environ 145: Markham B, Barker J (1985) Spectral characterization of the LANDSAT thematic mapper sensors. Int J Remote Sens 6: Goward S, Arvidson T, Williams D, Faundeen J, Irons J, Franks S (2006) Historical record of Landsat global coverage. Photogramm Eng Remote Sens 72: Kovalsky V, Roy D (2013) The global availability of Landsat 5 TM and Landsat 7 ETM+ land surface observations and implications for global 30m Landsat data product generation. Remote Sens Environ 130: Sonobe R, Yamaya Y, Tani H, Wang X, Kobayashi N, Mochizuki K (2017) Mapping crop cover using multi-temporal Landsat 8 OLI imagery. Int J Remote Sens 38: Gill T, Johansen K, Phinn S, Trevithick R, Scarth P, Armston J (2017) A method for mapping Australian woody vegetation cover by linking continental-scale field data and long-term Landsat time series. Int J Remote Sens 38: Wang Z, Tyo J, Hayat M (2007) Data interpretation for spectral sensors with correlated bands. J Opt Soc Am A Opt Image Sci Vis 24: Lillesand T, Kiefer R, Chipman J (2008) Remote sensing and image interpretation. Wiley, New York 10. ETM (2017) Landsat enhanced thematic mapper plus spectral resolution. U.S.G.S., Available online at: what-are-best-landsat-spectral-bands-use-my-study? (accessed 15 Sept 2018) 11. Mather P, Koch M (2011) Computer processing of remotely-sensed images, 4th edn. John Wiley & Sons, Chichester 12. Landam S, Everitt B (2004) A handbook for statistical analyses using SPSS. Chapman and Hall/CRC Press, New York 13. Miliaresis G (2012) Elevation, latitude and longitude decorrelation stretch of multi-temporal near-diurnal LST imagery. Int J Remote Sens 33: Miliaresis G (2013) Terrain analysis for active tectonic zone characterization, a new application for MODIS night LST (MYD11C3) dataset. Int J Geogr Inf Sci 27: Miliaresis G (2016) Spatial decorrelation stretch of annual ( ) Daymet precipitation summaries on a 1-km grid for California, Nevada, Arizona and Utah. Environ Monit Assess Miliaresis G (2017) Iterative selective spatial variance reduction of MYD11A2 LST data. Earth Sci Inf 10: /s Miliaresis G (2017) Selective variance reduction (SVR). GitHub Project. Available online at: (accessed 15 Sept 2018) 18. Miliaresis G (2014) Daily temperature oscillation enhancement of multi-temporal LST imagery. Photogramm Eng Remote Sens 80: Soulard C, Wilson T (2015) Recent land-use/land-cover change in the Central California Valley. J Land Use Sci 10: org/ / x CALCC (2015) Land cover and agriculture in California s Central Valley. California Landscape Conservation Cooperative. Available online at: (accessed 15 Sept 2018) 21. Slater J, Garvey G, Johnston C, Haase J, Heady B, Kroenung G, Little J (2006) The SRTM data Bfinishing^ process and products. Photogramm Eng Remote Sens 72: /PERS