2007 Land-cover Classification and Accuracy Assessment of the Greater Puget Sound Region

Transcription

1 2007 Land-cover Classification and Accuracy Assessment of the Greater Puget Sound Region Urban Ecology Research Laboratory Department of Urban Design and Planning University of Washington May

2 1. Introduction This report describes the classification methods applied to Landsat satellite imagery for the purpose of classifying land cover in 2007 for the Greater Central Puget Sound Region. This work was carried out by the Urban Ecology Research Lab (UERL), an inter-disciplinary UW research lab focused on describing and modeling coupled human-natural interactions. Contained in this report is detailed information about data preparation, processing, classification, and overall accuracy assessment. 2. Data Acquisition Landsat Thematic Mapper (TM) and imagery covering the extent of the Central Puget Sound region was used for this project. This area includes the Seattle-Everett-Tacoma Metropolitan Statistical Area and extends east to the crest of the Cascade Range. We acquired a total of 2 images for 2007 with the goal of having leaf-on (June-July) and leaf-off images (March-April) for the 2007 study period. While the winter months of December and January are ideal for leaf off images, the low sun angle, snow cover in the mountains, and cloudy conditions make these images less than practical for our analysis. The low sun angle increases the level of shade in high relief areas, which are plentiful in this region. Increased shade from the low sun angle, combined with a large winter snow pack in higher elevations, precludes advantages that can be reaped from seasonal analysis for much of the study area. We instead acquired a leaf-off image from March 2007 when the sun elevation angle is not as severe and the snow pack has typically receded to higher elevations. The only drawback to using spring images for leaf-off is that some deciduous green-up typically starts during this time of the year. Table 1 includes the Landsat image acquisition information. Table 1. Image ID Acquisition Date WRS Path LT EDC00 3/39/ , row 27 LT EDC00 6/01/ , row Data Preparation A series of geographic, radiometric, data conversion, image inter-calibration and data management steps were applied to both scenes to prepare them for land-cover classification. Many of the processing steps were performed using RSI s Environment for Visualizing Images (ENVI), a comprehensive image processing software used for advanced hyper-spectral image analysis, radiometric and geometric corrections, and image classification. Other steps, including all classification methodologies were performed using ESRI s Erdas Imagine, a comprehensive, geographic imaging suite. These steps are outlined in the following sections. 3.1 Geographic registration Each image was registered to the 07/07/1991 image because it has already been registered to a 30-m DEM. For both images at least 50 Ground control points (GCP) are picked throughout the images. Ground control points with high RMS (Root-Means-Squared) errors were discarded, so that only those GCP s with an RMS less 2

3 than ½ pixel were used in the resampling. The images are registered and projected simultaneously to UTM GRS 1980, NAD 83 using nearest neighbor resampling. 3.2 Radiometric corrections A series of radiometric corrections were applied to the image data, including instrument calibration, atmospheric compensation and image-to-image inter-calibration. These steps are explained below Instrument calibration The Puget Sound images are derived from Landsat 5 (TM) satellites. All Landsat TM images are pre-processed by USGS prior to our acquisition using standard geometric and radiometric correction methods. These methods include correcting for geo-location and terrain errors. Although the correction methods are compatible for both the TM and ETM+ instruments, ETM+ is an upgraded and superior radiometric instrument. All TM images can be converted to the radiometric calibration of ETM+ using simple data conversion algorithms provided by USGS. First, we converted TM data number values (DN5) to ETM+ data number values (DN7) using the following formula: DN7 DN5 slope intercept The slope and intercept values for each band are provided by USGS. Table 2. Coefficients for Converting DN Landsat 5 to DN Landsat 7 Band # Slope Intercept We then converted the DN7 values to radiance at satellite (L S ) values using the following formula: L S DN7 gain bias These gain and bias values for each band are again provided by USGS. 3

4 Table 3. Radiance at Satellite Coefficients. Band # Gain bias Dark Object Subtraction: A dark object subtraction method was used to correct for atmospheric scattering in the path. Dark object subtraction is an image-based approach that assumes certain objects in a scene should have zero reflectance (such as water bodies), and that radiance values greater than zero over these areas can be attributed to atmospheric scattering and thereby subtracted from all pixel values in an image. To account for a high degree of variation in atmospheric scattering between images, the minimum value to the right of zero was determined from each band. This value was then subtracted from each band so that dark objects had radiance values at or approaching zero Earth -Sun Distance Normalization Each image received a correction to account for seasonal illumination differences caused by varying earth sun distance and sun elevation angles. The distance of the sun s orbit to the earth varies depending on the time of year, with it being closest to earth during the winter months and farthest away during the summer months. The angle of the sun in the in relation to the earth, however, is much higher in the summer providing more incident radiation thus making summer images appear brighter than their winter counterparts. The combination of these two factors affects the incoming energy from the sun and the amount of solar irradiance that is reflected by the earth s surface. Seasonal illumination differences caused by these factors can be normalized using the following at-satellite reflectance correction: (Markahm and Barker, 1986). 4

5 Where: = Unitless planetary reflectance = Spectral radiance at the sensor's aperture = Earth-Sun distance in astronomical units from nautical handbook = Mean solar exoatmospheric irradiances 3.3 Topographic correction Variation in the aspect and slope of the earth s surface, combined with the suns position relative to the earth, creates a corresponding variation in the solar radiation incidence angle. The variation in solar irradiance, in turn, creates a variation in the amount of radiance leaving the earth. Simply stated, a given object will reflect varying degrees of light with respect to its position on the earth in relation to the sun. Holding all else equal, the same patch of forest on a shaded slope will appear less bright than it will on a perfectly illuminated flat surface. As should be expected, illuminated (south-facing) slopes are brighter in radiance than shaded (northfacing) slopes. Assuming surfaces in the images are Lambertian in nature, we can normalize for topographic effects using a shaded relief image, which is the cosine of the solar incidence angle for each pixel. The incidence angle is calculated from the Digital Elevation Model (DEM) together with the solar elevation angle and solar azimuth for the date and time of image acquisition. The correction involves dividing each band of image data by the shaded relief image. On shaded slopes the cosine of the incidence angle approaches zero, thus the calculation increases the brightness values for shaded pixels in each band. On the other hand, flat areas that are well illuminated have shaded relief values that approach one and therefore receive little if no correction. 3.4 Intercalibartion Spectral intercalibraton is the final step that attempts to correct any lingering spectral differences between images that are not the result of land cover change. In an attempt to further control for image differences, each image was spectrally intercalibrated to the summer 1999 image. Pseudo invariant objects with varying degrees of albedo were identified and digitized. Because we were dealing with images from different years and seasons, objects were chosen carefully to make sure they were indeed invariant. For each band in each image, the average spectra of each object were recorded. These values were then regressed against the corresponding object-values from the 1999 summer image. The resulting gain and offset from the regression equation was applied to each band to intercalibrate to their corresponding band in the 1999 image. 4. Classification: 4.1 Selecting AOI s (Area of Interest) to Extract Signatures We applied supervised classification by extracting spectral signatures from various land-cover types in the summer image, using high-resolution ortho-photography for referencing the materials. The various land-cover types included dense urban surfaces (e.g. pavements, roofing surfaces), mixed-urban areas (e.g. residential), 5

6 grasses, clear-cut, conifer and bare ground and agricultural lands. Training sites are digitized using the AOI tool in Erdas Imagine. Signatures are extracted from the image using the AOI s for each class. At each supervised classification interval, we masked water and vegetation using a three-class image comprised of water, vegetation and non-vegetation, derived from spectral unmixing, discussed below. The vegetation class was later disaggregated using spectral unmixing and seasonal change strategies, also discussed below. Our general approach to classification is both interactive and hierarchical. We start with broad classes and working within these classes to disaggregate them into more detailed classes or subclasses. We use several classification techniques throughout the process to achieve a final classification for each time step. Our starting point is a top-level classification in which each summer image is segmented into three categories: vegetation, non-vegetation, and water. This is done through a combination of spectral unmixing and supervised classification. 4.2 Top Level Supervised Classification We applied a top level supervised classification by extracting spectral signatures from various land-cover types in the summer image, using high-resolution ortho-photography as a reference to locate training sites. All six spectral bands plus an NDVI band derived from the image are used in the classification. AOI s representing vegetation and non-vegetation land cover types were digitized and signatures were extracted. These include multiple AOI s for each of the following categories: urban, mixed urban, bare soil, deciduous, conifer, grass, green agriculture, dry grass, grasslands, & water. Signature seperability scores were calculated and signatures combinations were reconfigured only when scores indicated poor seperability. This classification results were recoded into vegetation (deciduous, conifer, grass, and green grass) non-vegetated (urban, mixed urban, bare soil, dry grass, and grasslands), and water. Linear spectral unmixing was applied to the same image (without the NDVI band) to derive end member fraction images for non-vegetation, vegetation, and shade. End members were identified by analyzing a subsample of the image in spectral space (through an n-dimensional viewer) and locating clusters of pixels that bounded the majority of pixels in band space. End members were selected based on the following criteria: 1) they must exists as unique spectra that bound the majority of the pixels in the data cloud, 2) exists on a general linear plane between the other end members and 3) are not extreme pixels. Some of the literature regarding end member selection suggests using the extreme pixels that bound all other pixels in mixing space. This, however, may be inappropriate for several reasons. Extreme end members can be thought of as statistical outliers. They do not represent the end of mixing space but rather pixels that do not occur very often such as clouds. Typically, extreme pixels exist on the far end of a vector delineating the mixing space between itself and two other end members. Using these spectra instead of the pure pixels that bound the majority of the data may cause skewed results. End member selection yielded three general end members representing non-vegetation (soil and urban pixels), vegetation (agricultural fields), and shade (water). Spectra for these end members were extracted and used as inputs into the linear unmixing model. Running the model produces a series of fraction images, one for each end member and an RMS image representing model error for each pixel. The vegetation end member image was normalized for shade by dividing it from the sum of itself and the non-vegetation image. The shade 6

7 normalized vegetation image, the raw end-member fraction images, and the results of the supervised classification were used in the following rules to divide the image into vegetation, non-vegetation, and water. Table 4. Top Level Rules Preliminary Class Rule Water Shade EM >=.75 and Veg EM <=.20 Vegetation Non-Vegetation Shade Normalized Veg >.80 and/or pixel is Vegetation in supervised classification. All other pixels 4.3 Non-Vegetation Classification The first step in the Non-vegetation classification is to classify clear-cut patches. Previous experience with Landsat in this area indicate that clear cuts are often spectrally confused with urban classes and are thus hard to separate with a high level of accuracy. Clear cuts do, however, exist as large homogenous patches and have temporal characteristics that are quite different from urban objects. A recent clear cut will have very little vegetation, but will have been forested in earlier images and exhibit regrowth in later images. Our collection of images from numerous time steps provided an opportunity to take advantage of temporal patterns that help separate classes that are difficult using single date spectra alone. To classify clear cuts, we created a stacked raster file with bands 3, 4, 5, and 6 from the date of classification and the two endpoints of our study period, 1985 and We then apply a supervised classification using signatures from urban, mixed urban, dry grass, grasslands, bare soil and clear cuts. The result of this classification is masked so that only those pixels classified as non-vegetation in the top-level classification are included. A 7x7 focal majority filter is then applied to the masked image so that only pixels classified as non-vegetation are impacted by the filter. The results of the filter were converted to a binary image, where clear cuts are assigned a value of one and everything else is zero. These pixels are considered classified as clear-cut and do not receive further consideration at this point. By and large, this process was able to identify the clear-cut patches and sharply reduced the number of misclassified pixels. Another common problem in urban remotes sensing is the spectral confusion between bare soil and urban surfaces. Within our study area, there are numerous agricultural areas that often have tilled or unplanted fields and are spectrally similar to urban areas. However, similar to clear cuts, the temporal characteristics of agricultural areas are quite different from urban pixels in that they typically go back and forth from vegetated to non-vegetated. In order to capture this pattern, we calculated the variance of band 4 for all the summer images combined. This variance band was stacked with 6 spectral bands for each summer image. We hypothesized that agricultural areas would display greater variance in the near-infrared portion of the spectrum than would urban areas. A supervised classification was run using signatures for urban, mixed urban, dry grass, grasslands, and agricultural bare soil. The classification was masked by the non-vegetation pixels and a 3x3 majority filter was applied. Again, this filtering only affected pixels identified as non-vegetation in the top-level classification. 7

8 4.4 Urban Classes Spectral unmixing is once again applied to the summer image using a refined set of endmembers representing urban, vegetation and shade. The methodology for extracting signatures was the same as in the top-level classification. This time, however, only pixels that contained 100% of urban materials were considered as potential endmembers from the area in feature space that previously had been identified as non-vegetation. Locating spectral end members for urban surfaces present a unique problem in that there is a tremendous degree of spectral variance among urban reflectance. Previous research has shown that pure impervious surfaces tend to lie along an axis of low to high albedo spectrum. As a result, pure urban pixels often exist without actually being end members because they fall somewhere between low and high albedo (nonvegetation) end members (Small 2001, Wu and Murray 2002). Furthermore, the low and high albedo pixels at the end of this range may represent end members but not necessarily distinct urban end members. For example, the low albedo end member is spectrally similar to shade and is therefore hard to differentiate. Similarly, high albedo spectra could be from clouds, bare soil or a bright urban surface. To further confound this problem, pixels that are a mixture of impervious surface and vegetation, for instance, will contain some degree of shade. Using a low albedo end member as an impervious proxy will most likely model this shade component as impervious surface, when in reality it cannot be determined whether the shade is caused by a building, tree or is in fact part of the impervious surface spectra. From our experience, bare soil does not seem to exist as a truly distinct urban end member in spectral space, but rather exists along the same vector between the urban end member and shade/water end member (site sources). This plane extends from high albedo bare soil/urban surface to low albedo bare soil/urban surface to water. The pixels along this plane (except for water pixels at the end of the plane) are some combination of urban surface, bare soil and shade. Because the supervised classification separates whole pixel bare soil from urban pixels, it can be assumed (from the accuracy assessment) that most bare soil pixels in the image are already classified and removed from the analysis. Furthermore, because the shade fraction image is solved to derive impervious estimates, it can be assumed that urban pixels along this plane will be estimated at close to 100 percent impervious. This leaves pixels that contain both impervious surface and bare soil to be somewhat difficult to detect. Since there was little evidence that bare soil existed as an end member distinct from urban end members, it is inappropriate to use it in a spectral unmixing model with an urban end member. Therefore, if an urban pixel does contain bare soil, this component will most likely be modeled as shade and impervious surface during the unmixing. 4.5 Vegetation Classification The first step in the vegetation classification is to separate vegetation pixels into Forest and Grass/shrub/green agriculture. Using the same summer image + NDVI used in the top level classification, a supervised classification is run using several vegetation signature resulting in a two class image comprised of either forest pixels or grass, shrub, crops pixels. Once masked by the vegetation pixels classified in the top-level classification, the grass/shrub/crops pixels are considered classified. 8

9 A file consisting of bands 3, 4, 5, and 6 from the leaf-off and leaf-off image is generated for the particular year of classification. A supervised classification is run using signatures for conifer and mixed forest classes. The resulting classification is masked by those pixels identified as forest in the previous step. The classification is then compiled using the following rules resulting in 10 classes. Table 5. Final Land Cover Classes Land Cover ID Land Cover Classes 1 Developed High Intensity 2 Developed Medium Intensity 3 Developed Low Intensity 4 Land Cleared for Development 5 Grass, Grasslands, Agriculture 6 Mixed Forest 7 Coniferous Forest 8 Regenerating Forest 9 Water and Wetlands, Shorelines 10 Ice/Snow, Bare Rock 5. Final Classification The final 2007 classified land cover image (Figure 1) is comprised of 14 land cover classes: dense urban (>80% imperviousness); medium urban (50%-80% imperviousness); light urban(25%-50% imperviousness); cleared for development; grass/grassland; mixed/deciduous forests; coniferous forests; clear-cut forest; regenerating forest ; agriculture, wetland, water, snow/rock/ice and shoreline. 9

10 Figure 1: 2007 Classified Land cover 6 Accuracy Assessment We assessed the accuracy of the land-cover image by comparing the actual and known land-cover types (as detected by high-resolution (1.5ft), digital, ortho-photography) to the results of an automated, supervised classification as described in a confusion matrix (Figure 11). We applied high-resolution (1.5ft), digital, orthophotography covering the spatial extent of the spectral images as reference data for our assessment. This set of ortho-photographs was obtained from the National Agricultural Imagery Program (NAIP) and were taken in Our method consisted of constructing a continuous, 90-meter grid covering the spatial extent of the image data, selecting a random sample of 10% of the grids, and converting these to a polygon cover, which we overlaid onto the ortho-photography. We conducted an accuracy assessment at two levels: for the Six County Area and for the Hood Canal Area. These areas posed different challenges during the post-classification process. The six county area of the image (Island, Snohomish, Kitsap, King, Pierce and Thurston Counties) was covered by a polygon feature generated for the 2002 accuracy assessment which contains 162,197 records. For the Hood Canal area (not classified in 2002), a 90-meter vector grid was generated. For the area 10

11 corresponding to the classified image, a 10 percent random sample of this grid was produced. Given the large number of samples in the two 10% random sample features, a further random subsample was produced containing twice the number of samples that would ultimately be assessed (roughly 9000 polygons for the entire area). The following step was to find those polygons that were located in an area with a dominant land cover class. The zonal area tool in ArcMap was used to find the area of the land cover classes under each polygon. Only those polygons that were composed of 66% or more of a given land cover class were kept. Figure 2: Location of Hood Canal and Six County Areas The distribution of land cover classes was compared to that of the 2007 classification image. In some cases, a limited number of randomly selected polygons were added or deleted to the final set of sample polygons in order to approximate the land-cover distribution of the classified image (See Table 6). The final polygon feature contains 4135 samples. 11

12 Table 6. Percent areal coverage, ideal number of samples (proportional to class area), final number of samples. Class # Class Description Hood Canal (% Area) 6 County (% Area) Entire Area (% Area) # Samples (proportional to area) Final number of samples 1 heavy urban 0.7% 3.6% 3.1% medium urban 1.4% 7.0% 6.1% light urban 2.8% 8.0% 7.2% cleared for development 0.06% 0.03% 0.04% grass, grasslands 11.7% 6.2% 7.1% deciduous/mixed forest 18.7% 13.1% 13.9% coniferous forest 36.5% 25.5% 27.2% clearcut forest 1.9% 1.2% 1.3% regenerating forest 2.8% 12.1% 10.7% agriculture 1.6% 2.4% 2.3% non-forested wetlands 0.8% 0.4% 0.5% open water 18.3% 12.8% 13.7% snow, bare rock 0.1% 7.4% 6.2% shorelines 2.7% 0.1% 0.5% Total 100.0% 100.0% 100.0% We interpreted these polygons using the 2006 ortho-photos and a 14-class classification scheme. The National Wetland Inventory (NWI) dataset was used to help identify non-forested wetland areas. We exported the attribute table associated with the processed feature and created a confusion matrix in MS Excel. The statistics calculated as part of this confusion matrix are described in Table 7. After an initial confusion matrix was generated for the Hood Canal area, it was noted that the winter image contained two large hazy areas that seemed to be affecting the classification. This resulted in confusion between deciduous and coniferous forest. Furthermore, light urban pixels were found to be common in forested areas. Thus, the Hood Canal area was reclassified and the trajectory rules were re-applied. We have included a confusion matrix for this initial classification as well as one for the reclassification (Tables 8 & 9). It is important to note, that the samples are proportional to the land cover class distribution for the reclassification. These sample polygons were then used in analyzing the initial classification. However, because of differences between the classifications some samples no longer met the criteria that 2/3 of the area in a sample polygon be of one class. Thus, the total number of samples in the confusion matrix corresponding to the initial classification is less. 12

13 Table 7. Accuracy Assessment Statistics Used Accuracy Statistics Producer's Accuracy: Describes how much of each validation class is correctly classified compared to the total validation amount. It is defined as the diagonal number in the error matrix divided by the number in the column total. User's Accuracy:The user accuracy describes the accuracy of each radar derived class individually. It is defined as the number of correct classifications divided by the total number of classifications in the category, i.e. the diagonal number in the error matrix divided by the number in the row total. Overall Accuracy: The overall accuracy is described by the total number of correct classifications divided by the total number of classifications. The weighted overall accuracy weights the number of correct classifications by the proportion of that class in the entire image. Kappa coefficient: Kappa Coefficient is a statistical measure of agreement, and describes an index of strength of agreement, compared to chance. Kappa values range from -1.0 to +1.0, with a value of 0.0 representing a random relationship and a value of +1.0 representing a non-random relationship. Equations Kappa = (observed expected)/(1 expected) Tables 10 and 11 show the confusion matrices for the Six County Area and the classification of the entire study area. The classification of the Hood Canal area presented some challenges. After a reclassification of the Hood Canal area, some improvement was seen especially in the accuracy of the light urban class. The Six County area had a higher Kappa coefficient and overall accuracy. Because this part of the classification makes up close to 85% of the total area, these higher statistics are also seen for the complete classification. 13

14 Table 8: Confusion Matrix for Hood Canal Initial Classification (refer to Table 6 for class descriptions) Classified Image Ortho-photo interpretation Total User's Accuracy Total Producer's Diagonal Accuracy Sum Overall Accuracy Kappa Coefficient Table 9: Confusion Matrix for Hood Canal Reclassification (refer to Table 6 for class descriptions) Classified Image Ortho-photo interpretation Total 14 User's Accuracy Total Diagonal Producer's Sum Accuracy Overall Accuracy Kappa Coefficient

15 Table 10: Confusion Matrix for Six County Area Classification (refer to Table 6 for class descriptions) Classified Image Ortho-photo interpretation Total User's Accuracy Total Diagonal Producer's Sum Accuracy Overall Accuracy Kappa Coefficient Table 11: Confusion Matrix for Complete Classification (refer to Table 6 for class descriptions) Classified Image Ortho-photo interpretation Total User's Accuracy Total Diagonal Producer's Sum Accuracy Overall Accuracy Kappa Coefficient