DETECTION AND CLASSIFICATION OF PLANT SPECIES THROUGH SPECTIR AIRBORNE HYPERSPECTRAL IMAGERY IN CLARK COUNTY, NEVADA BACKGROUND AND INTRODUCTION

Transcription

1 DETECTION AND CLASSIFICATION OF PLANT SPECIES THROUGH SPECTIR AIRBORNE HYPERSPECTRAL IMAGERY IN CLARK COUNTY, NEVADA Rohit Patil, Remote Sensing/ Image Analyst Xin Miao, Postdoctoral Fellow Jill Heaton, Professor C. Richard Tracy, Professor University of Nevada, Reno Reno, NV ABSTRACT The non-native Saltcedar (Tamarix ramosissima Ledeb.) and the native Honey mesquite (Prosopis Glandulosa Torr.), exist in abundance in Clark County, NV. We are using remote sensing to measure changes in distribution and abundance of these species. We collected six strips of 1m-resolution SpecTIR hyperspectral images in Clark County on May, SpecTIR has 227 spectral bands ranging from 0.45 to 2.45 µm. We have explored the properties of these high-spatial resolution hyperspectral images, and we are now testing the potential for detecting and classifying Tamarix, Prosopis and other plant species along Muddy River. Spectral band selection and feature extraction methods are being used to reduce the dimension of the hyperspectral data and the distinguishable spectral characteristics are chosen for SAM (Spectral Angle Mapper) and supervised maximum likelihood (ML) classifiers. Terrestrial information and habitat knowledge of the vegetation are also incorporated into the classifier. The accuracy of classification results are being verified in relation to ground surveys. Preliminary results will be reported on this project. The results will provide evidence for a proposed project of vegetation delineation and change detection in Clark County. BACKGROUND AND INTRODUCTION Remote sensing technique is an important approach to monitor the large-scale ecosystem health. Without possible sampling bias derived from field work, remote sensing can be used to detect and measure crucial changes to habitat. Particularly, remote sensing is useful to identify the indicators for the frail ecosystem. By monitoring indicators, land owners can asses and adjust the current management strategies. Vegetation spectral signatures often share the similar patterns. This causes the difficulty in separating the different vegetation species using conventional multispectral imagery. Recently hyperspectral remote sensing techniques have been broadly applied for detecting and distinguishing the vegetations. Due to its exceptionally high spectral resolution, hyperspectral imaging offers great potential in mapping the abundance of particular species over large spatial extents. For example, several recent publications have applied hyperspectral image analysis to mapping individual invasive species with varying levels of success (Ustin, Scheer et al. 2001; Lass, Thill et al. 2002; Williams and Hunt 2002; Underwood, Ustin et al. 2003). The non-native Saltcedar (Tamarix ramosissima Ledeb.) and other the native vegetation, such as Honey mesquite (Prosopis Glandulosa Torr.), exist in abundance in Clark County, NV. Saltcedar was first introduced in the U.S. from southeastern Europe and Asia to reclaim eroded areas and prevent further loss of stream banks, primarily in the southwest. However, saltcedar has lower value to wildlife habitat and consumes tremendous amount of water, which contribute spring drought. Furthermore, its leaves are concentrated with salts. Once releasing these salts, the surface soil can become highly saline, thus preventing possible colonization by many native plant species. In Clark County, Saltcedar invades riparian habitats and displaces native flora and fauna to a large extent. It is a primary concern of the local government to evaluate the disturbance caused by saltcedar to the arid ecosystem. By applying hyperspectral remote sensing technique, it is our hypothesis that we can survey and monitor the distribution of saltcedar along the riparian corridor and abundance of other plant species in Clark County.

2 METHODOLOGY The objective of this project is mapping vegetation and detecting changes of vegetation through high spatial hyperspectral imagery in Clark County, Nevada. The flowchart for this project is illustrated in Figure 1. We are interested in the non-native Saltcedar (Tamarix ramosissima Ledeb.) and the native Honey mesquite (Prosopis Glandulosa Torr.) in the study area. Besides, two other vegetation classes were identified as newly-grown grassland and agriculture areas. Figure 1. The geo-registration process in the project. Field Measurements Our study area is a relatively flat valley along the Muddy River in Clark County ( Figure 2). Saltcedar mainly distributed along the river forming a dense canopy. Some honey mesquite appeared in the arid area around the northwest corner and the canopies were not continuous.

3 Figure 2. The study area along the Muddy River. The red lines represent the flight lines. The hyperspectral image (HyperSpecTIR) has not been geo-registered. We conducted a field trip from September 7 to September 9, 2005 to collect 20 ground control points (GCPs) and preliminary vegetation canopy samples for registration and classification respectively. The GCPs were recorded using Trimble GeoXT GPS under the differential mode which has sub-meter accuracy. Ten patches of saltcedar and Honey mesquite in the study area were traced using the same GPS unit. These patches were imported into ARC GIS system to compare with the hyperspectral images and served for training algorithms and verifying results. Image Preprocessing We collected preprocessed 1m-resolution HyperSpecTIR (HST) hyperspectral images acquired on May 12, 2005 by Spectral Technology & Innovative Research Corporation (SpecTIR). Raw data was converted to radiance by proper application of gains and offsets for individual pixels. The radiance data was then converted to reflectance data using the MODTRAN4 v3 model. MODTRAN adjusted the radiance values to the mid-latitude, desert summer model using the following parameters: a sensor parameter file, height above ground (1000m), visibility (65 miles), water content (1.1 g/cm 3 ) and CO2 content (330 g/cm 3 ). The raw MODTRAN reflectance output is applied 3x3 kernel Savitsky-Golay algorithm to suppress 1% variation of the data. The data was georectified using the dgps measurements parsed in the data stream taking into account aircraft position, crab angle and altitude however, the observed geopositional accuracy of the 1m-resolution images acquired from SpecTIR only ranges between 15 m 30 m. Since high geo-registration accuracy is crucial for change detection of multidate high spatial resolution images, we had to improve the registration accuracy and control the registration error within 2-3 pixels, so that the mosaiced HST images can be used as a base map for this project. There were several difficulties in registering HST images. First, although 1m-resolution DOQQ is often used as a base map, the DOQQ (92-94) maps acquired from W.M. Keck Earth Sciences & Mining Research Information Center, University of Nevada, Reno were outdated. Due to the time difference in imagery acquisition (13 years), most land cover features in DOQQ and HST images can not be matched, especially in vegetation areas along Muddy River and bare grounds in the south of Muddy River. Besides, although DOQQ had 1m-resolution, pixels were relatively fuzzy compared to HST images, which makes it even more difficult to be used as a base map. Second,

4 HST is an airborne cross-track scanner coupled with a 2-D focal plane array. Instantaneous field of View (IFOV) is 1.0 milliradian and each scene is only 250 by 571 pixels for 1m-resolution imagery. Since our study area covered 15 Hyperspectral scenes from two flight strips, it was not feasible in cost and labor to collect GCPs for each scene in heterogeneous arid areas. Third, although the instrument pitch, roll and yaw mirrors are controlled to optically cancel platform rotations during acquisition, the HST image still has some degree of distortion associate with them. The inconsistency of distortion within image scenes further exacerbates the registration performance. Therefore, we proposed a three-stage registration approach to minimize the registration errors and reduce the goepositional accuracy of HST images down to 2-3 pixels. First, we clipped purchased Quickbird Panchromatic image to our study area and then registered it using DOQQ as base map; second, we subsequently registered all HST scenes to the QuickBird image and mosaic them together; finally, we use 11 GCPs collected from the field to reregister the whole mosaic-ready HST image to further improve the registration accuracy. This registration strategy has several advantages. First, QuickBird Panchromatic image we ordered was taken on June 4, 2004, which is at the same season as HST is acquired and time-delay is less than one year (May 12, 2005) and hence the Land cover features could be matched well. Second, QuickBird being a satellite sensor the swath width of the scene is pretty large around 16.5 km by 16.5 km at nadir. Our study area (around 2 km by 1 km) is located in the south-west corner of the QuickBird scene we ordered. Although the absolute accuracy can be low and claimed RMSE is 14 m for standard Quickbird imagery, the error distribution in the study area should be relatively consistent, because it only occupies a small portion of the IFOV of the QuickBird sensor and sensor in satellite is more stable than in airplane. Therefore, QuickBird can be used as a base map after registering to DOQQ. Third, it is more feasible to register each HST scene to a QuickBird image as more land cover features can be matched between them as both of them belong to close acquisition time frames. By this divide and conquer strategy, the registered HST scenes can be used to create a mosaic of our study area thus reducing overall error to minimum. Finally, 20 GCPs collected were used to re-register this study area mosaic image. Using this technique, the registration accuracy of our study area was improved and the registration error was reduced to 2-3 pixels. Image Analysis First, good bands were visually identified by exploring and examining each band in ENVI 4.2. Furthermore, since vegetation is our primary interest, a Normalized Difference Vegetation Index (NDVI) mask was generated in ENVI using proper threshold so that all the sparse vegetation areas were included. All further processing and classifications were done applying this NDVI mask on Mosaiced study area imagery. The NDVI mask is shown in Figure 3. Figure 3. NDVI mask was generated by setting the threshold.

5 The Spectral Angle Mapper (SAM) is a physically-based spectral classification that uses a multi-dimensional angle to match pixels to training/reference spectra (Kruse, Lefkoff et al. 1993). The algorithm determines the spectral similarity between two spectra by calculating the angle between the spectra. In this study, training class spectra were collected directly from the hyperspectral image for supervised SAM approach. SAM compares the angle between the class spectrum vector and each pixel vector in n-dimensional space. Smaller angles represent closer matches to the reference spectrum. Pixels further away than the specified maximum angle threshold in radians are not classified. Since this method utilizes the shape of the spectral signature, it is relatively insensitive to illumination and albedo effects. On the other hand, the illumination phenomenon and shading effect are obvious and common for high spatial resolution airborne remote sensing imagery. Therefore, we applied SAM algorithms to detect and classify different vegetation species. Since, hyperspectral remote sensing imagery has high correlations among adjacent bands and provides an excessive amount of features to explore. Classification of such data requires large training sample sets to get stable representations of class statistical properties, which are often referred to as the curse of dimension (Landgrebe, Serpico et al. 2001). Furthermore, the computation burden is also increased dramatically. Hence, band selection and feature extraction are important topics in hyperspectral remote sensing image analysis. Principal components transformation is based on the covariance matrix of the full set of image data that contribute little to separability. It often works well in remote sensing because classes are frequently distributed in the direction of maximum data scatter (Richards and Jia 1999). Jeffreys-Matusita distance (JM-distance) is a commonly used separability measure of band selection for classification (Richards and Jia 1999): Jij B e ij = 2(1 ) (1) where B represents Bhattacharyya distance (B-distance), another measure of class separability (Kailath 1967). Jia and Richards proposed the segmented principal components transformation (SPCT) to select high average separability features (Jia and Richards 1999). This method utilizes the segmented structure of the correlation matrix derived from hyperspectral imagery. Bands within each block along the diagonal suggest a high correlation. Thus PCA can be conducted within each block to extract the first several dominant features. Segmented PCA is a feasible trade-off between global PCA and band selection. It also keeps some spectral information and could provide physical explanation. Kappa coefficient was used to compare the performance of classification based on conventional SAM, PCA- SAM, band-selection-sam and segmented PCA-SAM in this paper. The kappa coefficient is an index of classification accuracy derived from the error matrix: k = N x x x kk k+ + k k k 2 N xk+ x+ k k (2) where x ij represents the element of the error matrix (Richards and Jia, 1999). The kappa coefficient is commonly used in remote sensing as a measure of map accuracy, since this omnibus index indirectly incorporates the omission and commission errors of an error matrix (Congalton 1991). One of the advantages of using the kappa coefficient is that we can statistically compare two classification products. RESULTS After visually examination, the final good band list is shown in Table 1. There are 178 good bands selected from a total of 227 bands. The training samples and test samples were selected directly from the image based on the field data. Training samples are used for supervised classification and test samples are used for verifying the classification results (Table 2).

6 Table 1. Good band list Band# Wavelength nm nm nm nm nm nm nm nm Table 2. Training samples and test samples Classes Color Training Samples Test Samples Saltcedar Green Honey mesquite Red Grass Blue Agriculture Yellow First of all, we conducted a conventional SAM to classify the HST image with all 178 bands. Although SAM has been broadly applied to separate the vegetation species, it did not provide a good result in our case (Figure 4). The major reason is the relative difference in spectral brightness between two flight strips. Although the flight scenes have been radiometrically corrected at the time of acquisition, the north flight strip has a relatively lower brightness than the south flight strip. This brightness difference could not be totally corrected when mosaicking flight scenes. Therefore, some saltcedar on the north were misclassified into agriculture area and the same species across the flight strips border were classified into two different vegetation. Kappa= 80.2%. Figure 4. Conventional SAM classification result. Next, the global PCA analysis was conducted on these 227 bands. And the PCA eigen value and their accumulated percentage are listed in Table 3. It suggested that the first six bands can explain over 99.8% of total data variance. Therefore, we conducted SAM based only on the first six PCA channels. Although the brightness problem was mitigated, but the illumination and shading problem were aggravated (Figure 5). Kappa=75.3%.

7 Figure 5. Global PCA SAM classification result. Table 3. PCA Eigen value statistics #Channel Eigen Value Accumulated Eigen Value Channel Channel Channel Channel Channel Channel JM-distance was applied to select bands that have the maximum separability. We found that wavelength longer than nm did not contribute the separability of the vegetation species. The middle infrared ranges of vegetation spectral signatures were consistent. We selected the first 15 bands with the maximum separability by calculating the average JM-distances among four classes. The spectral information is listed in Table 4. SAM classification was run using these 15 bands. We expected band selection SAM could provide a better classification result, because only bands with maximum separability were used. However, the result was still largely affected by the difference in brightness between different flight strips. Kappa= 82.5.

8 Table 4. Selected bands based on JM-distance # Band Wavelength (nm) Finally, we applied segmented PCA for the masked image. According to the block structure along the diagonal of the correlation matrix, eight groups were identified. They are listed in Table 5. The first group represents visible bands; the second group represents the red-edge of the vegetation spectral signature; the third, fourth, fifth and sixth group represent near infrared bands and the last two groups represent the middle infrared bands. PCA was conducted within each group and a total of 23 bands were selected. Then SAM was applied for these 23 bands. The result is shown in Figure 6. This method eliminated the brightness effects and correctly classified honey mesquite and salt cedar with improved accuracy. We believed this method had a good trade off between the Global PCA and conventional SAM. Kappa= Table 5. Group of bands Group Band Wavelength Total nm nm nm nm nm nm nm nm nm nm nm nm nm nm nm nm 50

9 Figure 6. Segmented PCA SAM classification result. DISCUSSION AND CONCLUSIONS We compared the performance of conventional SAM, PCA-SAM, band-selection-sam and segmented PCA- SAM to detect and classify the non-native Saltcedar, the native Honey mesquite, grass and agriculture area in Clark County, NV in this paper. Although conventional SAM was not sensitive to the illumination and shading problem in the high spatial resolution airborne hyperspectral image, it still cannot totally avoid the difference in brightness between the adjacent flight strips. Besides, there was a heavy computation burden of SAM using the original 178 bands. PCA was an effective approach to reduce the dimensionality of the hyperspectral data. The first six channels can explain over 98.8% of data variance. However, the illumination and shading problem were aggravated although the brightness problem was mitigated. Band selection based on JM-distance was supposed to improve the classification accuracy. However, the result was still affected by the brightness difference between different flight strips. A good result was acquired by applying segmented PCA and SAM. The Kappa coefficient was 88.5%. This approach combined the advantages of global PCA and conventional SAM and was able to preserve the spectral information in extracted channels. Therefore, it was not sensitive to the illumination and shading problems. On the other hand, it also eliminated the brightness difference between different flight strips to the minimum extent. In summary, segmented PCA-based SAM is a suitable classification scheme to distinguish the different vegetation species in our case. REFERENCES Congalton, R. G. (1991). "A Review Of Assessing The Accuracy Of Classifications Of Remotely Sensed Data." Remote Sensing Of Environment 37(1): Jia, X. and J. A. Richards (1999). "Segmented principal components transformation for efficient hyperspectral remote-sensing image display and classification." Ieee Transactions on Geoscience and Remote Sensing 37(1):

10 Kailath, T. (1967). "Divergence and Bhattacharyya Distance Measures in Signal Selection." Ieee Transactions on Communication Technology CO15(1): 52-&. Kruse, F. A., A. B. Lefkoff, et al. (1993). "The Spectral Image-Processing System (Sips) - Interactive Visualization and Analysis of Imaging Spectrometer Data." Remote Sensing of Environment 44(2-3): Landgrebe, D. A., S. B. Serpico, et al. (2001). "Introduction to the special issue on analysis of hyperspectral image data." Ieee Transactions on Geoscience and Remote Sensing 39(7): Lass, L. W., D. C. Thill, et al. (2002). "Detecting spotted knapweed (Centaurea maculosa) with hyperspectral remote sensing technology." Weed Technology 16(2): Richards, J. A. and X. Jia (1999). Remote sensing digital image analysis: an introduction. New York, Springer- Verlag. Underwood, E., S. Ustin, et al. (2003). "Mapping nonnative plants using hyperspectral imagery." Remote Sensing of Environment 86(2): Ustin, S. L., G. Scheer, et al. (2001). "Hyperspectral remote sensing for invasive species detection and mapping." Abstracts of Papers of the American Chemical Society 221: U50-U50. Williams, A. P. and E. R. Hunt (2002). "Estimation of leafy spurge cover from hyperspectral imagery using mixture tuned matched filtering." Remote Sensing Of Environment 82(2-3):