ABSTRACT 1. INTRODUCTION

Transcription

1 Correlation between lidar-derived intensity and passive optical imagery Jeremy P. Metcalf, Angela M. Kim, Fred A. Kruse, and Richard C. Olsen Physics Department and Remote Sensing Center, Naval Postgraduate School, 833 Dyer Rd, Monterey, CA, USA ABSTRACT When LiDAR data are collected, the intensity information is recorded for each return, and can be used to produce an image resembling those acquired by passive imaging sensors. This research evaluated LiDAR intensity data to determine its potential for use as baseline imagery where optical imagery are unavailable. Two airborne LiDAR datasets collected at different point densities and laser wavelengths were gridded and compared with optical imagery. Optech Orion C0 laser data were compared with a corresponding 1541 nm spectral band from the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS). Optech ALTM Gemini LiDAR data collected at 64 nm were compared to the WorldView-2 (WV-2) nm NIR2 band. Intensity images were georegistered and spatially resampled to match the optical data. The Pearson Product Moment correlation coefficient was calculated between datasets to determine similarity. Comparison for the full LiDAR datasets yielded correlation coefficients of approximately 0.5. Because LiDAR returns from vegetation are known to be highly variable, a Normalized Difference Vegetation Index (NDVI) was calculated utilizing the optical imagery, and intensity and optical imagery were separated into vegetation and nonvegetation categories. Comparison of the LiDAR intensity for non-vegetated areas to the optical imagery yielded coefficients greater than 0.9. These results demonstrate that LiDAR intensity data may be useful in substituting for optical imagery where only LiDAR is available. Keywords: LiDAR, Intensity, Correlation, Orthophotography, Classification 1. INTRODUCTION When Light Detection and Ranging (LiDAR) data are collected, in addition to the typical time-of-flight information, intensity information is recorded for each return. This intensity information gives a measure of relative reflectivity of the objects in the scene, and when viewed as a gridded raster image, produces an image that resembles that of a passive sensor. This information can be especially useful in cases where optical imagery is not available, such as at night; being an active sensing system, LiDAR data can be collected at night or in low-light conditions. LiDAR intensity information is typically not calibrated, and is therefore not useful for true radiometric measurements. It does, however, provide a visually useful product. In this study, we investigated the correlation of LiDAR intensity with passive optical imagery, with the goal of understanding and quantifying the similarities and differences between the data products. We investigated the correlation of the scenes in an overall sense, and then particular classes of materials (such as vegetation, buildings, roads, etc.), and particular classes of LiDAR points (such as nadir scan angles, first returns, etc.). The goal of this work was to determine if there are cases where LiDAR intensity information, although uncalibrated, may be reliably used to provide pseudo-reflectivity information, or if there are cases when the LiDAR intensity information is especially unreliable. Previous studies have demonstrated the utility of LiDAR intensity information (along with spectral information from orthophotos or spectral imagery) for tasks such as image classification or determination of the Normalized Difference Vegetation Index (NDVI) [1, 2]. Laser Radar Technology and Applications XIX; and Atmospheric Propagation XI, edited by Monte D. Turner, Gary W. Kamerman, Linda M. Wasiczko Thomas, Earl J. Spillar, Proc. of SPIE Vol. 9080, 90800U 14 SPIE CCC code: X/14/$18 doi:.1117/ Proc. of SPIE Vol U-1

2 2. DATA 2.1. LiDAR Data LiDAR data were collected over California s Monterey Bay area by Digital Mapping Inc. in (the AMBAG dataset) and Watershed Sciences Inc. (WSI) in 12. General information about each LiDAR dataset is given in Table 1. The two LiDAR datasets used in this study contrast highly in their collection parameters. Having a very high point density, the WSI data allow investigation of laser intensity correlation at a much finer scale than that of the AMBAG data. At 64 nm, the wavelength of the laser used in the AMBAG dataset is closer to typical wavelength ranges of visible to near infrared imaging systems. Table 1. LiDAR dataset descriptions. Dataset Name WSI AMBAG Origin Watershed Sciences Inc. Digital Mapping Inc. Client Naval Postgraduate School Associated Monterey Bay Area Governments (AMBAG) Collection Date Oct 12-Nov 12 Before Aug LiDAR System Optech Orion C-0 Optech ALTM Gemini Wavelength 1541 nm 64 nm Parameters 4 m AGL, 60% sidelap 10 m AGL, % sidelap Scanning 66 khz PRF, FOV 0 khz PRF, 25 FOV Point Density -80 pts/m 2 average 2-4 pts/m 2 average Posted Accuracy 7 cm vertical, cm horiz. 23 cm vertical, 35 cm horiz Image Data Spectral imagery datasets were chosen for comparison with the LiDAR data. A summary of the spectral imagery used is given in Table 2. Table 2. Spectral Imagery dataset descriptions. Sensor Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) WorldView-2 (WV-2) Ultracam Eagle Collection Sep 11 Apr 11 Nov 12 Date Collection Platform Airborne Satellite Airborne Selected 0-600nm (blue), nm(green), 860- nm Spectral nm (Band 126) 580-7nm (red), nm (NIR) (NIR2) Channels Pixel Size 2.4 m 2.3 m 15 cm Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) Hyperspectral Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data [3] were collected over Monterey, CA, in September 11 with a 2.4 m spatial resolution. Atmospheric correction was performed using ACORN [4]. An 10m by 800m image chip was created covering the Naval Postgraduate School (NPS), Monterey, CA. These data were paired with the WSI LiDAR data. Only the band closest to 1541 nm was used in this study as it overlapped the laser wavelength. Proc. of SPIE Vol U-2

3 Figure 1. True color AVIRIS image of the Naval Postgraduate School campus, Monterey, CA WorldView-2 Multispectral WorldView-2 (WV-2) satellite imagery collected in April 11 was paired with the AMBAG intensity data. The WV-2 image has a spatial resolution of 2.3 m and covers a 3.5 km by 2.3 km area of Monterey, CA. Only the NIR2 band was used in this study. The NIR2 band has a minimum and maximum band edge at 860 and nm respectively [6]...;,. Figure 2. True color WorldView-2 image of a portion of Monterey, CA. e Proc. of SPIE Vol U-3

4 2.2.3 UltraCam Eagle Following LiDAR acquisition, WSI collected high spatial resolution aerial imagery using the UltraCam Eagle camera. The UltraCam Eagle is a 260 megapixel, large format digital aerial camera with four spectral bands: (red, green, blue, and near infrared). Spectral responses for this sensor are 0-600nm (blue), nm (green), 580-7nm (red) and nm (NIR). Images were radiometrically calibrated, pan-sharpened, and then orthorectified using a ground model. The corrected images were output with a 16 bit dynamic range with a ground sample distance (GSD) of 15 cm. Figure 3. True color Ultracam Eagle aerial orthophoto of Naval Postgraduate School, Monterey, CA 3. METHODS Two separate methods were employed to evaluate LiDAR intensity and passive optical imagery correlation. In the first method, raster images were created from the LiDAR intensity information and the correlation to the spectral imagery was calculated on an image-to-image basis. This method was used to compare the AVIRIS to the WSI LiDAR and WV- 2 data to the AMBAG LiDAR. In the second method, an attempt was made to avoid the uncertainties introduced by gridding the LiDAR data. The spectral imagery was overlaid on the LiDAR data, enabling the spectral values to be associated with each individual LiDAR point. Correlation was calculated on an image-to-point basis. This method was used to compare the WSI LiDAR to the UltraCam Eagle orthophotos. For both approaches, several assumptions were made: 1. Only the intensity information contained in the first echo was considered as it is assumed to resemble the surface energy received by passive optical sensors. 2. For hard surfaces such as buildings or roads, the first echo is typically the only echo. In objects that are frequently represented by multiple echoes, vegetation for example, laser energy is often divided among each following echo. 3. Because the behavior of laser returns over water tends to be highly variable, they are ignored in this study. Proc. of SPIE Vol U-4

5 The Pearson Product Moment Correlation coefficient is used to evaluate similarity between datasets. Pearson s r is defined as: r = n i= 1 n i= 1 ( X i ( X i X ) X )( Y 2 i n Y) i= 1 ( Y Y) where is image 1 at the th pixel, is the mean of image 1, is image 2 at the th pixel, is the mean of image 2, and is the number of pixels [7]. With this equation, correlation coefficients range from -1 to 1, where -1 is perfect negative linear correlation, 0 is uncorrelated, and 1 is perfect positive linear correlation. Additionally, the correlation coefficient is invariant to changes in scale Image-to-image approach Create LiDAR intensity raster images Intensity images were created from each LiDAR dataset by gridding the average first return intensity values. The number of points that are averaged for each cell depends on the resulting first return point density. The GSD of AVIRIS and WV-2 images were matched at 2.3 and 2.4 m respectively Classify vegetation and non-vegetation The Normalized Difference Vegetation Index (NDVI) was used to segment image data into vegetation and nonvegetation. NDVI is a common measure to identify healthy vegetation in imagery containing red and near infrared channels [5]. NDVI = (NIR RED) (NIR + RED) Where NIR and RED represent the spectral reflectance measured at the near infrared and red visible spectral regions respectively. NDVI values were calculated using the NIR and RED bands of the spectral imagery The values of an NDVI image range from -1 to 1, where pixels with values above 0.3 represent a variety of vegetation cover. Pixels having an NDVI value above 0.3 were classified as vegetation, and pixels having an NDVI value below 0.3 were classified as non-vegetation Calculate correlation Correlation was calculated for the intensity raster images, and for the vegetation and non-vegetation classes for both AVIRIS/WSI and WV-2/AMBAG. Results are given in Section Image-to-point approach Overlay orthophoto on LiDAR point cloud Spectral information was added to the WSI LiDAR dataset by spatially sampling the Ultracam Eagle aerial orthophoto at each LiDAR point location. Both datasets were collected in the same coordinate system making the overlay process fairly simple. LiDAR points in optical shadow present in the orthophoto were identified and removed using a simple threshold of the NIR spectral channel Classify LiDAR data The WSI LiDAR data were delivered with ground points already classified. An automatic classification process was applied to the LiDAR data to produce building, and vegetation classes. Points forming flat surfaces greater than square meters and more than 3 vertical meters from the ground were classified as buildings. Dispersed points with a height minimum of 1.3m and a radius minimum of 2m were classified as trees. Visual inspection of the automatic classification result revealed slight class confusion between building and vegetation classes. A large majority of points i 2 (1) (2) Proc. of SPIE Vol U-5

6 remaining unclassified were found to be clearly associated with building and vegetation classes. Class confusion was manually corrected. During the correction process, an additional road class was introduced using ground points Calculate correlation Correlation was calculated for all LiDAR points, and each of the individual classes (ground, roads, building, and vegetation) Image-to-image correlation 4. RESULTS AND DISCUSSION The results of computing the correlation coefficients for the AVIRIS/WSI and WV-2/AMBAG datasets are presented in Table 3. Table 3. Computed correlation coefficients for AVIRIS/WSI and WV-2/AMBAG image datasets Overall Vegetation Non-vegetation AVIRIS/WSI WV-2/AMBAG Comparing the overall correlation result of both datasets, the WV-2/AMBAG correlation was relatively higher than that of the AVIRIS/WSI test. We initially expected the AVIRIS/WSI dataset to have a much higher correlation since the spectral response of AVIRIS band selected directly overlapped the wavelength of the laser used in the WSI LiDAR collection. At this point, it is unclear what are the major contributing factors to the correlation values we see between datasets. When both image datasets were separated into vegetation and non-vegetation categories, the resulting correlation coefficients improved dramatically. Despite the differences in point density, coverage area, and wavelength overlap of each test case, the correlation within each category was well above 0.9. This increase in correlation is likely explained by variations in how each surface is generally represented in LiDAR data. Impervious surfaces are usually represented by a single return having an intensity value more closely related to the reflectance characteristics of the surface. In vegetation, specifically vegetation having more than one return, the intensity received by the first return can be diminished depending on number of returns, beam width, and object size. To visualize how correlation varies between datasets, composite images were created for AVIRIS/WSI (Figure 4) and WV-2/AMBAG (Figure 5) datasets. AVIRIS and WSI intensity images appear to be aligned rather well as seen in Figure 4. Buildings with split rooftops can readily be identified by the bright red pixels owing to illumination differences inherent in LiDAR and optical data collection methods. The image has an overall blue color because the LiDAR intensity and AVIRIS images were scaled to different ranges. Because AVIRIS was collected in September 11 and WSI LiDAR in November 12, variations found in AVIRIS brightness and LiDAR intensity can be partially explained by seasonal changes between years. Proc. of SPIE Vol U-6

7 , t, ','HIS: :coed CI: LiDAR: scaled 0: ,. It...x.,..2. '''..,,,. Figure 4. Composite image of AVIRIS/WSI image data with AVIRIS image in red, WSI intensity image in blue and green. Visual inspection of the WV-2/AMBAG image composite (Figure 5) reveals discrepancies likely affecting image correlation. At first glance, there appears to be a brightness gradient where LiDAR intensity is higher in the left of the image. After reevaluating the original AMBAG LiDAR dataset, the intensity gradient was found to correspond with elevation for this area. It is possible that the AMBAG LiDAR intensity data were not properly calibrated prior to our analysis. Image alignment errors are apparent along the gradient as well. Where the overall LiDAR intensity decreases, image misregistration increases. Additionally, the WV-2 and AMBAG data were collected at different dates; objects not present in both dates can be seen as appearing very bright red or blue. WV 2 scaled 0:8000 LiDAR Intensity: scaled 0: ' 1- Figure 5. Composite image of WV-2/AMBAG data with WV-2 NIR2 image (red), AMBAG intensity image (blue and green). Proc. of SPIE Vol U-7

8 4.2. Image-to-point correlation The correlation results of the WSI LiDAR intensity and Ultracam Eagle spectral bands are presented here. The WSI LiDAR data were classified manually and then fused with the high-resolution orthorectified aerial imagery. Although additional classes were created, we are mainly concerned with assessing the correlation of vegetation and impervious surfaces (ground, road, and building). Figure 6 shows the LiDAR classification and image fusion results. Table 4 displays the computed correlation coefficients for WSI intensity and orthophoto spectral bands. Figure 7 shows scatterplots for these data. o o Unclassified Ground Vegetation Building Road Power Line Figure 6. WSI LiDAR intensity scaled 0 to (top), classification result (bottom left), orthophoto fused with LiDAR points (bottom right) Table 4. Correlation coefficients for WSI LiDAR intensity and optical imagery All Classes Ground Road Buildings Vegetation Blue Green Red NIR Proc. of SPIE Vol U-8

9 All Classes Ground Road Buildings Vegetation io Blue A 32 m X m 6 8 y v x ' x ' , x " n 93 Green 33 m " " 00 Red in Figure 7. Scatterplots for five classes/four spectral bands: x-axis represents the LiDAR intensity values from 0 to and y-axis represents the Ultracam Eagle spectral band values from 0 to 625. Note: actual intensity values range from 0 to 96, but almost all of the information St is 9 represented R 19 _ in the range from 0 to. o From Table 4, we see the highest correlation coefficients for nearly all classes of the WSI LiDAR data occur with the UltraCam Eagle NIR band. Despite having positive correlation with blue, green, and red bands, the ground class was found to be uncorrelated. Unlike other classes, the ground class contains both impervious surfaces and low vegetation. Ìi Compared to other classes, the highest intensity correlation coefficients are found within the road class. The scatterplots for this class show a much clearer linear relationship than any other class. For all bands, the vegetation class also S Fì FS _ appears to be uncorrelated. In Figure 7, scatterplots for vegetation show almost no linearity between intensity and pixel values. From these results, it appears that the intensity associated with vegetation is quite different from that of other surfaces. Overall, the reported correlation coefficients for the WSI LiDAR and UltraCam spectral data are rather low. Several factors that likely had an adverse effect on correlation include: 1) Although imagery and LiDAR data were collected by WSI within - the same 51 9 week, Fì F3 9 the spectral response of the Ultracam Eagle does not overlap with the LiDAR laser wavelength; 2) Concerning healthy vegetation, it is known that the reflectance of this material can be high within the range of the UltraCam NIR band, but low at laser operating wavelength of 1541 nm; and 3) Additionally, the image e.;.:.l overlay process operates in the x and y dimensions only. This translates to ground, road, and building classes receiving vegetation pixels values when directly underneath trees. The effect of solar illumination and LiDAR scan angle were not addressed in this study. o s RR _ (Yi z Proc. of SPIE Vol U-9

10 5. CONCLUSIONS We presented the results of computing the Pearson s Product Moment correlation coefficient between corresponding LiDAR intensity data and passive optical imagery. Two parameters mainly appear to control the correlation between the LiDAR intensity images and optical data: 1) Properly matching the LiDAR and Optical wavelengths, and 2) Separating the datasets into vegetation and non-vegetation categories. When LiDAR data acquired with a laser wavelength of 1541 nm were compared with mismatched optical data with a nm wavelength range (which does not extend to the laser wavelength), correlations were minimal (between 0 and ~0.5). When the LiDAR and optical data wavelengths were well matched, the correlations averaged between Separation of the vegetation from non-vegetated areas in the AVIRIS/WSI LiDAR test resulted in increasing correlation from approximately 0.5 for the full AVIRIS band to well above 0.9 for segmented vegetation and nonvegetation. There was a similar but less pronounced increase in correlation with the WV-2/AMBAG LiDAR test, where an overall correlation of approximately 0.7 rises above 0.9. It is interesting to note that although vegetation and nonvegetation pixels are highly correlated in their own category, when considered together, the correlation drops significantly. Future work will include the summation of intensity by each pulse having multiple returns, separation of LiDAR points by scan angle to investigate BRDF effects, and the use of full waveform LiDAR collected with a green laser. 6. ACKNOWLEDGEMENTS This paper describes selected research results from a project "Remote Sensing for Improved Earthquake Response" supported by the Science and Technology (S&T) Directorate, Department of Homeland Security (DHS). REFERENCES [1] Causey, R., Kehoe, J. and Slatton, K. C. Airborne laser intensity measurements for vegetation studies: A comparison to passive imagery techniques, Adaptive Signal Processing Laboratory (ASPL) at University of Florida, ASPL Report No. Rep_ , (05). [2] Bandyopadhyay, M., van Aardt, J. A. N. and Cawse-Nicholson, K., "Classification and extraction of trees and buildings from urban scenes using discrete return LiDAR and aerial color imagery", Proceedings of SPIE Vol. 8731, (13). [3] Green, R., "AVIRIS and Related 21st Century Imaging Spectrometers for Earth and Space Science," in High Performance Computing in Remote Sensing, Chapman and Hall/CRC Press, , (07). [4] IMSPEC LLC, "ACORN 6 User's Manual", IMSPEC LLC, 121 p., (). [5] Tucker, C.J Red and Photographic linear combinations for monitoring vegetation, Remote Sensing of the Environment, 8, 127 1, (1979). [6] Satellite Imaging Corporation, Spectral Response for DigitalGlobe WorldView 1 and WorldView 2 Earth Imaging Instruments, (Accessed 28 April, 14). [7] Pearson, K., "Notes on regression and inheritance in the case of two parents," Proceedings of the Royal Society of London, 58 : 2 242, (1895). Proc. of SPIE Vol U-