Evaluation of Landsat 7 SLC-off Image Data for Forest Cover Mapping

Suzanne Furby and Xiaoliang Wu CSIRO Mathematical and Information Sciences

CMIS Technical Report 06/154

September 2006

2

Executive Summary Landsat 7 SLC-off image products were obtained for three test regions to evaluate the issues that arise if this imagery were to be used in place of the Landsat 5 imagery in the Land Cover Change Program (forest cover mapping) of the Australian Greenhouse Office. Both individual ‘path’ level images and the SLC-off composite products produced by the Australian Centre for Remote Sensing were evaluated. All aspects of the forest cover mapping were considered; including scene selection, ortho-rectification, calibration, mosaicing and thresholding to produce forest cover maps. The recommendations for the Land Cover Change Program are:

• The standard scene selection criteria should be applied to select the primary image for each path/row. The recommended criteria for selecting the ‘fill’ image are closeness to the primary image acquisition date and low cloud cover. Between 94% and 99.5% of the path/row extent is filled by compositing two individual images for these test areas. Operationally, a third image may be required to achieve this level of coverage for a small proportion of path/rows, however the availability of reasonably cloud-free image data will be limiting for the 2006 epoch.

• Individual Landsat 7 SLC-off ‘path’ images can be adequately registered and calibrated to the

AGO base images with minimal additional ‘per image’ effort. However there will be twice as many images to be processed.

• The ACRES Landsat 7 SLC-off composite products are not always well registered to the AGO

rectification base. The amount of displacement appears to vary regionally. In the NSW test area, the displacement was also observed between the ACRES composite and the NSW state DEM, affecting the terrain-illumination correction. The local calibration performed as part of the ACRES compositing process is also of concern. The ACRES process can make quite different cover types (cloud / no cloud and dry / wet) look identical. This makes a good ‘picture’ product, but it is of concern when trying to map ground cover. Although more effort is involved, the products would be more reliable if the individual images were obtained and composited.

• Some modifications to the mosaicing methodology are required. A new method for tracking

image date boundaries within the mosaics needs to be developed. It is strongly recommended that cloud masking of the individual Landsat 7 SLC-off images be performed before compositing to significantly reduce the effort involved in cloud masking. Such a change will increase the mosaicing effort, but a reduction in the thresholding effort will offset the change.

• Where there is no significant change between the images being composited, the thresholding can

be performed treating each composite as a single path/row image. In the rare event that differences have been flagged between images to be composited, thresholds should be derived from the individual images and mosaiced. Assuming that cloud-masking is being performed as part of the mosaicing stage, there will be no overall increase in thresholding effort and generally the effort will decrease.

• The forest cover extent and change products are consistent with those obtained using Landsat 5

images from similar dates. A slight increase in forest extent is observed, similar to the 2000 and 2002 epochs in which Landsat 7 data is also used. The forest extent and change products formed from SPOT 4 image data were not as consistent with the Landsat 5 products. If missing data remains in any Landsat 7 SLC-off composites, the products may include stripes in the forest extent and change data if missing data partially cover areas of land cover change.

3

Evaluation of Landsat 7 SLC-off Image Data for Forest Cover Mapping

Suzanne Furby and Xiaoliang Wu CSIRO Mathematical and Information Sciences

1. Introduction Landsat 7 SLC-off image products have been acquired for three test regions to evaluate the issues that arise if this imagery were to be used in place of Landsat 5 imagery in the Land Cover Change Program (forest cover mapping) of the Australian Greenhouse Office (AGO). Both individual ‘path’ level images and the SLC-off composite products produced by the Australian Centre for Remote Sensing (ACRES) have been obtained. Landsat 7 SLC-off images were also used in the 2006 update of the Land Monitor II Perennial Vegetation Monitoring program (http://www.landmonitor.wa.gov.au). Experiences with the Land Monitor data are also included in this report. All aspects of the forest cover mapping are considered; including scene selection, ortho-rectification, calibration, mosaicing and thresholding to produce forest cover maps. The suitability of the image data for other products such as sparse cover mapping, plantation type and canopy density will be considered in a follow-up study reported separately. Unless indicated otherwise in the text, all processing was performed according to the standard methodology for the Land Cover Change Program as described in Furby (2006). Other sources of image data are also being evaluated. These include SPOT 4, IRS and CBERS. The evaluations will be reported separately and a summary comparison report produced. 2. The Test Areas The test areas are in New South Wales, Tasmania and Western Australia as shown in figure 1.

Figure 1: Approximate location of the test areas

4

The New South Wales test area, shown in figure 2, was selected to include a black soil stratification zone so that forest / non-forest discrimination can be evaluated in one of the most challenging environments. It also includes a region with significant terrain effects to allow evaluation of registration, BRDF and terrain illumination correction issues in an ‘extreme’ environment. The selected area is in the centre of map sheet SH56. The WRS path/rows are 89/81 and 90/81. Although the full image extent is considered in the discussion of the ‘pre-processing’ issues, the thresholding comparison is restricted to the extent of the SPOT 4 images.

Figure 2: Landsat 7 SLC-off images for the New South Wales test area in map sheet SH56, bands 3, 5, 4 in BGR. The yellow lines show the stratification zone boundaries used in the Land Cover Change Program. The white lines show the boundaries of the two Landsat 7 scenes. The red lines indicate the extent of the SPOT 4 images obtained for the test area. The Tasmanian test area, shown in figure 3, was selected to include mountainous area terrain as well as an agricultural environment in which discrimination between cropped paddocks and plantations can be difficult. The selected area is just to the east of the centre of map sheet SK55. The WRS path/rows are 90/89 and 91/89. Although the full image extent is considered in the discussion of the ‘pre-processing’ issues, the thresholding comparison is restricted to the extent of the SPOT 4 images. The Western Australian test area, shown in figure 4, was chosen to include a significant region of new plantations as well as some of the wheatbelt tree cover that is close to the 20% canopy cover cut-off in the forest definition. The test area is flat. The selected area is in the north of map sheet SI50. The WRS path/rows are 111/83 and 112/83. Although the full image extent is considered in the discussion of the ‘pre-processing’ issues, the thresholding comparison is restricted to the extent of the SPOT 4 images.

5

Figure 3: Landsat 7 SLC-off images for the Tasmanian test area in map sheet SK55, bands 3, 5, 4 in BGR. The yellow lines show the stratification zone boundaries used in the Land Cover Change Program. The red lines indicate the extent of the SPOT 4 images obtained for the area. The large black ‘gaps’ are regions where cloud has been masked.

Figure 4: Landsat 7 SLC-off images for the Western Australian test area in map sheet SI50, bands 3, 5, 4 in BGR. The yellow lines show the stratification zone boundaries used in the Land Cover Change Program. The red lines indicate the extent of the SPOT 4 images obtained for the area. The black ‘gaps’ are regions where cloud has been masked.

6

3. Scene Selection Issues The standard scene selection criteria were applied to select the ‘primary’, or main, image for each path/row. The criteria for selecting the ‘fill’ image were closeness to the primary image acquisition date and low cloud cover. A slightly increased cloud cover was considered better than a longer time between acquisitions. Two images were obtained for each path/row, a primary image and a single fill image. Although some gaps remain after compositing the two images, three relatively cloud-free images are very rare due to extensive cloud cover during the 2006 summer. In the case of 91/89 in Tasmania only one relatively cloud-free image could be obtained. The images used are listed in table 1. They were provided as the usual ‘path’ image product used in the Land Cover Change Program. ACRES also supplied their standard SLC-off composite product for these path/rows. The composites were formed using the same primary and fill images. The composites were provided as standard ACRES ‘ortho-rectified’ products.

Table 1: Selected Landsat 7 Images Path/Row Date Cloud Purpose 89/81 15/03/2006 Clear Primary 89/81 31/03/2006 5% Fill 90/81 07/04/2006 Clear Primary 90/81 10/02/2006 40% Fill 90/89 23/04/2006 55% Primary 90/89 07/04/2006 50% Fill 91/89 24/01/2006 40% Primary 111/83 21/01/2006 25% Primary 111/83 06/02/2006 25% Fill 112/83 17/03/2006 Clear Primary 112/83 13/02/2006 25% Fill

The Tasmanian and Western Australian images have more cloud than would normally be considered acceptable, but they are the best images available. There appears to be more cloud than usual everywhere in 2006.

Table 2: Proportion of Landsat 7 Image Area Filled by Compositing Image Percentage of 89/91 Composite Percentage of 90/81 Composite

89/81 15/03 78.82 0.52 89/81 31/03 11.79 0 90/81 07/04 2.29 80.72 90/81 10/02 1.07 18.10

Unfilled 6.03 0.60 Table 2 shows the remaining ‘gap’ area after compositing for the NSW test area. As discussed in section 7 (Mosaicing Issues), cloud is masked from the individual images before compositing. The percentage of the composite remaining unfilled includes cloud-masked areas. Cloud is a minor component in the filled NSW data but a major component in the other test areas; hence the

7

results are presented for NSW only. In the overlap region between adjacent paths, data from all four images are included in the composite. The 89/81 images are an example where the gaps are largely coincident in the two images. This would represent a near ‘worst case’ scenario when only two images are used. The 90/81 images are an example from the opposite extreme. Even with a high proportion of cloud in the ‘fill’ image, the coverage is almost complete. 4. Raw Image Quality Issues With the exception of the missing data, the image quality of the ‘path’ image products is as expected for Landsat TM data. Samples of data from near the centre and edge of one of the images are shown in figure 5. The gaps are approximately 16 pixels wide at the edge of the images and disappear towards the centre of the image. Along the edges of the gaps there are occasional pixels for which the data for only some of the image bands are missing, e.g, bands 1, 2 and 3 may contain zeros, while bands 4, 5 and 7 contain non-zero data.

Western edge of the image

Centre of the image

Figure 5: Samples of the ‘path’ image product for 89/81 15/03/2006 showing the gaps in the SLC-off products. Bands 3, 5, 4 are displayed in BGR. The image quality of the composite products is generally as expected for Landsat TM data. However, as noted above the individual images contain occasional pixels for which the data for only some of the image bands are missing. In the composites, these pixels end up having intensity values from both images. A ‘mask’ image is provided with each composite. It shows which image, if any, the pixels are derived from. Each image band has a corresponding mask image band.

8

5. Ortho-rectification Issues 5.1 Individual ‘Path’ Images The images were ortho-rectified using the standard processing sequence for Landsat 7 imagery in the Land Cover Change Program, using the Year 2000 base, the merged AUSLIG 3 and 9 second DEMs and the Master GCPs file used for the 2005 update. Issues were encountered with the correlation matching process to generate ground control points (GCPs) from base image features and the PCI resampling algorithm. Both of these issues were resolved as described below. Equivalent registration accuracy can be expected as for unaffected Landsat 7 imagery. The modified procedures are all automated rather than manual, providing only a very minor ‘per image’ increment to the ortho-rectification effort. GCPs are automatically derived by matching features from the rectification base in the new overpass image by maximising image correlation in local windows around the feature. In the current software, if the search window around the estimated feature location in the overpass window contains zeroes, the feature is omitted as being outside the image. The result when applied to the Landsat 7 SLC-off imagery is that GCPs are only found down a narrow central strip of the image. Inadequate registration with the base results. The feature matching calculations were modified to omit just the pixels with zeroes from the correlation calculation rather than omit the whole feature. Where the search window overlaps a gap, fewer pixels are used in the calculation. Acceptable numbers of GCPs with an adequate spatial distribution were generated. As a region is used in the calculation, not a single pixel, the GCP located at the centre of the region may be in a gap. This modification has no effect on the ‘per image’ ortho-rectification effort. The PCI resampling algorithm places a zero in the output image any time its resampling window includes a zero from the input image. With the 8pt sin/x resampling kernel used, each input zero becomes the centre of an 8 pixel by 8 line block of zeroes in the output image. Effectively all gaps are grown by eight pixels, four at the top and four at the bottom causing an unacceptable loss of data. Within PCI the only options are to use alternative resampling algorithms with smaller resampling windows such as cubic convolution or nearest neighbour. Cubic convolution resampling increases the gaps by four pixels but produces more blurred images than 8pt sinx/x. Nearest neighbour resampling does not increase the gap size, but produces a blockier image. The alternative considered was to ‘fill’ the gaps with data that will produce a sensible resampled pixel value. The gaps were filled by interpolating the data above and below the gap. An indicator image tracking the artificial data was created. The image with the interpolated data was ortho-rectified in PCI using the usual resampling algorithm. The indicator image was attached to the path image as an additional band and ortho-rectified separately using the same GCPs and nearest neighbour resampling. The indicator image is then used to mask the extra artificial data from the ortho-rectified overpass image. The interpolation strategy provides the best ortho-rectified imagery. The extra processing is completely automated (and partially hidden in existing processing) and so provides only a very minor ‘per image’ increment to the ortho-rectification effort.

9

5.2 ACRES Composite Products In order to composite two (or more) images, they must be registered to each other. ACRES ortho-rectifies each image using their normal processing sequence. Standard GCP chips with known coordinates (from a national GPS survey) are found in the overpass image. Thus the images are ortho-rectified to an ACRES base rather than the AGO base used in the Land Cover Change Program. The correspondence of the ACRES and AGO bases has not been systematically investigated. As described below there are issues with the registration of some of the images forming composites to each other as well as discrepancies between the ACRES composites and the AGO base of up to 2 pixels (50m).

The 111/83 ACRES composite (WA) is generally well-registered

The 89/81 ACRES composite (NSW) is

displaced from the AGO base by up to two pixels as shown above and below

Figure 6: Band 5 from the ACRES composite image (green) overlaid on the AGO rectification base

(red)

10

Figure 6 shows one band from an ACRES composite image overlaid on the same image band from the AGO rectification base. In some areas the registration is good, but in other areas, such as the 89/81 composite image in NSW, a clear shift between the two images of up to two pixels can be seen. As is shown in the discussion of terrain illumination correction of the imagery in section 6.2, the AGO base and images registered to it match the terrain features in the NSW state DEM. The position of ridges and valley floors in the ACRES composite are displaced from the DEM leading to artefacts in the illumination correction.

Figure 7: Band 5 from the ACRES composite for the 107/83 triple scene (25/01/2006 and 10/02/2006) obtained as part of the Land Monitor project.

11

Figure 7 shows part of an ACRES Landsat 7 SLC-off composite image supplied for the Land Monitor project. There was a displacement of between 8 and 12 pixels (200-300m) between the two images being composited. ACRES investigated the problem and reported that it was caused by the failure to automatically locate sufficient GCPs. A combination of the missing data and cloud cover meant that GCPs were not correctly located and the ACRES QA procedures had not detected the problem. As a result of the review of all composite imagery provided, ACRES reprocessed the 91/89 imagery to correct a similar but much smaller deficiency. These registration issues require that if Landsat 7 SLC-off data is going to be used, the individual images should be obtained rather than the ACRES composite, even though it means more processing. 6. Calibration Issues (including terrain-illumination correction) The standard calibration process consists of three distinct steps:

• top-of-atmosphere and BRDF corrections; • invariant target atmospheric check/correction; and • terrain-illumination correction, if required.

Each of these steps is discussed for both image types. In addition, a calibration step is part of the ACRES compositing process. The effect of this calibration on the composite image is also discussed. 6.1 Individual Path Images There are no issues in applying the top-of-atmosphere and BRDF corrections to the Landsat 7 SLC-off path images using the current procedure. The invariant target check / correction is hampered by some of the targets being located in gaps. Over these test areas the number of targets in gaps was sufficient to compromise the gain and offsets calculated. Typically more targets in one of the dark, mid-range or bright intensity ranges were omitted than the others causing the targets to be unbalanced. Targets had to be manually updated for each Landsat 7 image separately, not just for each path/row, which would add a substantial ‘per image’ amount to the calibration effort. However, particularly with Landsat 7 imagery, the invariant target calculations are performed as a check of the processing rather than an automatic correction. Corrections are applied only in very rare circumstances when very seasonally atypical images are used and it can be clearly demonstrated that the image is more similar to adjacent images with the correction than without it. The invariant target correction was not needed for any of the test imagery. Potentially this step could be omitted for Landsat 7 SLC-off imagery with no consequences for thresholding accuracy, except where the two (or more) images to be composited for a particular path/row appear very different. In this situation it may be better to use the overlapping image area to estimate a correction of the fill image(s) directly to the primary image rather than to try to calibrate each to the calibration base independently. Procedures for doing this require further investigation. Terrain illumination correction is usually performed after the images have been mosaiced into map sheet units. The image date boundaries are used to derive and apply an individual correction

12

to each constituent image as the sun-angles vary. The coefficients for each correction are estimated from the image data using a woody cover mask. If the Landsat 7 images are composited before terrain illumination correction there are two few woody pixels remaining from the ‘fill’ image for reliable coefficient estimation. Instead, the corrections must be calculated and applied before the images are composited, i.e. before the mosaicing stage not afterwards. This change to the processing order leads to other consistency gains and does not change the overall level of effort involved. In summary, the individual SLC-off images can be adequately and consistently calibrated. 6.2 ACRES Composite Products ACRES perform a calibration of the fill image to the primary image as part of their compositing process. As the images are merged, a local calibration is performed for each fill pixel. A local gain and offset is estimated to calibrate the intensity values for the ‘fill’ image to the intensity values in the primary image using data in a small local window centred on the pixel being inserted. The ACRES calibration procedure produces some interesting results. The assumption is that the ground cover has not changed between the two images anywhere the fill image is used. In all of the images provided by ACRES a seamless product is created – the only way to distinguish between pixels from the two input images is to use the mask image provided. However, even with images acquired sixteen days apart it is rarely true that there is no change anywhere in the image. In the extreme, clear pixels in the fill image are ‘calibrated’ to look like the cloud in the primary image as shown in figure 8. Cloud-covered pixels in a fill image are also ‘calibrated’ to the underlying cover in the primary image. While this may not seem so bad for forest cover mapping, the texture of cloud and the normal ground cover are very different and will have a profound affect on sparse cover mapping. The Land Monitor images for the 110/83 triple scene were acquired 16 days apart. The earlier image was shortly after significant rains as a result of cyclone activity in the north of the state and shows residual effects of the flooding. In the later image much of the surface water is gone and the wetness and greenness of many paddocks has changed. As shown in figure 9, the ACRES composite is a much nicer picture, but for mapping and monitoring land cover change it is of concern that water in the fill image appears like dry land in the final merged image. Other undesirable changes are likely to be present, though perhaps not always obvious. Applying the standard calibration sequence to the ACRES composites can also be problematic. A gap mask file is provided that allows the composite image to be separated into the constituent images and so separate corrections could be applied. However, for the invariant target and terrain illumination correction steps there are too few pixels in the fill image to reliably estimate correction coefficients. For this evaluation the images were corrected as if they were a single image. The satellite gains and offsets and sun angles used were from the primary image and coefficients were estimated using the merged image data.

13

The top-of-atmosphere / BRDF correction software relies on reading satellite gain and offset and sun-angle information from ancillary ‘report’ files provided with the image data. The files provided with the composite images do not contain the necessary information. For the test data, the report files provided with the individual primary images were used but these would not be available if only the composite product was provided. The result produced was acceptable.

ACRES composite 90/89

Cloudy 90/89 ‘primary’ image

Merged image without local calibration

Relatively cloud-free 90/89 ‘fill’ image

Figure 8: ACRES composite and constituent images showing the effect of cloud on the local calibration during compositing. Bands 3, 5, 4 are displayed in BGR.

14

ACRES composite 110/83

Dry 110/83 ‘primary’ image

Merged image without local calibration

Wet 110/83 ‘fill’ image

Figure 9: ACRES composite and constituent images showing the effect of land cover change on the local calibration during compositing. Bands 3, 5, 4 are displayed in BGR. There are no complications in performing the invariant target check, but again no correction was deemed necessary. The terrain illumination correction was performed in a straightforward manner, but a comparison of the corrected individual image (mosaiced) with the composite for the 89/81 image in NSW highlighted a consequence of the registration issues noted earlier. Figure 10 shows a section of the terrain corrected image formed from the individual images with the same area in the corrected

15

composite image. Terrain effects remain in the corrected composite image. This is not actually due to any inherent fault in the terrain illumination correction process, simply that the image is slightly displaced from the DEM. Incorrect slopes and aspects area applied near apparent ridgelines and streamlines in the image data.

Terrain–corrected mosaic of individual path images for 89/81.

Terrain–corrected ACRES composite image for 89/81

Figure 10: Comparison of the terrain illumination correction applied to the individual ortho-rectified path images (left) and the ACRES composite (right). Terrain effects remain in the ACRES composite due to misregistration with the DEM. Bands 3, 5, 4 are displayed in BGR. 7. Mosaicing Issues Mosaicing of the ACRES composite images would proceed according to the standard procedures. Some consideration could be given to whether Landsat 5 data should take priority over Landsat 7 data in the overlay order rules if all other considerations of cloud cover and time of year were the same to minimise missing data. With the individual path images mosaicing can be considered to consist of two activities:

• compositing two or more images from the same path/row (to fill the gaps); and • mosaicing images from adjoining path/rows to form map sheet units.

Operationally, both activities would be performed simultaneously. Overlay order (primary and fill images) would be stipulated for the compositing activity; otherwise the usual overlay order rules apply. It is recommended that Landsat 5 data take priority over Landsat 7 data to minimise missing data. Although not part of the mosaicing stage of the current Land Cover Change Program processing sequence, it is more efficient to cloud mask Landsat 7 SLC-off images before compositing /

16

mosaicing rather than as the first step in the thresholding process. In the individual image each cloud or patch of cloud is digitised as a single polygon. After compositing, many cloud-affected areas will appear as stripes in the composited image. Either separate polygons are required for every stripe or stripes of clear data will be masked. This will necessitate significant changes to the mosaicing and mosaicing QA specifications resulting in additional mosaicing effort, but this extra effort will be matched by a reduction of effort during the thresholding stage. Vector image date boundaries are created during the current mosaicing sequence. The vector polygons allow tracking of the image acquisition dates of any individual pixel / change area during the carbon modeling phase of the AGO program. They are also intersected with the stratification zone boundaries during thresholding to define sub-zones. Separate thresholds for forest / non-forest discrimination are calculated for each such sub-zone. Vector polygons provide a simple, efficient means of tracking image date boundaries when there is only a single image per path/row. However, vector boundaries are an inefficient means of tracking the areas of primary and fill images in a composite. Raster images are more efficient and are easily created during the mosaicing process (perhaps more simply than the current vectors). The attribute of the vector polygon currently includes both path/row and image date information, both of which are required to uniquely identify the specific source image. This information is too complex to efficiently store as a single number in a raster image. Further investigation is required to establish whether a coding scheme, a multi-band raster image or a combination of raster acquisition date image and vector path/row polygons is the most efficient strategy. 8. Thresholding Issues The region to which the thresholding processing is applied is restricted to the extent of the SPOT 4 imagery. 8.1 NSW Test Area Given the image registration issues identified for the ACRES composite products, it was not considered constructive to consider them further for this test area. The test area is shown again in Figure 11. Stratification zone 7 contains the black soils. Discriminating between forest and non-forest cover is very difficult in this zone. More omission and commission errors are made in stratification zones with this soil type than are typically observed in other zones. Some of these persist through the multi-temporal processing. The other stratification zones are relatively straightforward unless the imagery is particularly green or the sun angle is particularly low. The standard thresholding procedures were applied. For each sub-zone, image matching was applied to derive thresholds independently for the mosaiced image and the individual primary and fill images that form the mosaic. The 2005 single-date forest cover probability image was used as the base and the same starting threshold estimates were used for each image. Good thresholds were obtained for each sub-zone using matching. The actual final thresholds produced by the matching program for each input image varied slightly, particularly between the values estimated from the full ‘fill’ image compared to the primary image or the mosaic (about 80% primary image). However, the resulting probability images were virtually identical. The

17

probability image derived from the mosaiced image was used in the CPN run. All of the fill images are very similar to their primary image for this test area; hence it is not surprising that separate thresholds weren’t needed within a composite image. A similar situation exists for the Tasmanian and Western Australian test areas.

Figure 11: Landsat 7 SLC-off images for the New South Wales test area in map sheet SH56, bands 3, 5, 4 in BGR. The yellow lines and numbers show the stratification zones used in the Land Cover Change Program. The white lines show the boundaries of the two Landsat 7 scenes. The red lines indicate the extent of the SPOT 4 images obtained for the test area. 8.2 Land Monitor Example The wet / dry image pair from the Land Monitor project (shown in the discussion of the calibration of ACRES composites) provides the only opportunity to evaluate the thresholding process when the individual images forming the composite are very different. Figure 9 shows the four available images: the individual 3 March and 19 March 2006 images, the ACRES composite of the two images and the mosaic after the standard calibration processing. The latter clearly shows ‘stripes’ throughout the mosaic image. Again, thresholds were estimated separately for each sub-zone for the individual images and the mosaic image. Figure 12 shows the forest cover probability images produced as the result of automated matching with the same starting estimates. Stripes can be seen in the forest cover probability image derived from the mosaic. Most of the stripes are commission errors with the wetter stripes being assigned a higher probability of forest cover when the thresholds are derived from the mostly dry mosaic image. When thresholds are derived directly from the wet image, the commission errors are reduced and so fewer stripes appear when the individual probability images are mosaiced.

4 5

6

3

7

6

18

Figure 12: Left: Mosaiced Landsat 7 image, bands 3, 5, 4 in BGR. Centre: Forest cover probability image derived from the mosaic image. Right: Mosaiced forest cover probability images derived from the individual images. Image pairs (or triples) where such stripes occur in the composites will be identified at the calibration stage in the processing. It is recommended that the individual images are provided for thresholding in such situations and the probabilities are mosaiced. Operationally, this should be a rare occurrence. 8.3 WA Test Area The test area covers the eastern edge of the state forest, including some areas of new plantations and harvesting, and a typical wheatbelt area. The thresholding in these stratification zones is relatively straightforward unless the imagery is particularly green. The standard thresholding procedures were applied. After the experience in the NSW test area, where identical forest cover probability images were created by image matching using the mosaiced image and individual primary and fill images that form the mosaic, only the mosaiced image was processed. The 2005 single-date forest cover probability image was used as the base. Good thresholds were obtained for each sub-zone using matching. Differences between the primary and fill images are more pronounced than in NSW, although not as significant as in the Land Monitor example. The probabilities for one sub-zone showed some slight stripes in the probability image, however the difference in probability between the ‘stripes’ was slight. The striped probability image was used in the multi-temporal processing to investigate the effect of such striping on the products. The multi-temporal processing corrected the problem and the effect is not observed in the outputs. The spectral changes between the primary and fill images need to be reasonably large before separate thresholds are required. 8.4 Tasmanian Test Area The test area covers the central northern part f the state, including areas of agriculture, plantation forestry, wilderness and alpine vegetation. Difficulties are often encountered in separating green agricultural paddocks from young plantations where the agricultural and forest regions meet. Otherwise, the thresholding in these stratification zones is relatively straightforward unless the imagery is particularly green.

19

The standard thresholding procedures were applied. Some sub-zones required separate thresholds to be estimated from primary and fill images, particularly in the overlap between path 90 and 91 where January and April images overlap. The 2005 single-date forest cover probability image was used as the base for the agricultural and forest zones and the 1998 single-date forest cover probability image was used as the base for the wilderness and alpine zones. This is the usual practice for this map sheet. Good thresholds were obtained for each sub-zone using matching. The usual level of commission errors were encountered on green agricultural paddocks near forestry operations. 9. Comparison of Thresholding Results with SPOT 4 and Landsat 5 Image Data Forest cover probability images were created for the 2006 epoch using the Landsat 7 SLC-off data as described in this report. Forest cover probabilities were also created from Landsat 5 data using the standard methodology and from SPOT 4 data as described in Furby (2006). At least one of the Landsat indices for each stratification zone in the NSW test area includes Landsat TM bands 1 or 7, so new indices were derived for the SPOT 4 data. Landsat TM equivalent indices were used in the WA test area. The image dates for the SPOT 4 imagery are:

Table 3(a): SPOT 4 Image Dates for NSW Path/Row Date Cloud Inc. Angle 387/410 22/12/05 hazy -1.2 388/410 10/03/06 clear -30.4 389/410 01/02/06 Mostly clear +25.8 390/410 22/12/05 Mostly clear -24.4

Table 3(b): SPOT 4 Images for WA

Path/Row Date Cloud Inc. Angle 317/416 26/01/2006 Partly -13.8 318/416 05/02/2006 Clear +3.2 319/416 21/01/2006 Clear -13.8 319/415 10/02/2006 Clear +13.4

Table 3(c): SPOT 4 Images for Tas

Path/Row Date Cloud Inc. Angle 384/434 06/01/2006 Mostly clear +7.7 385/434 06/01/2006 Some cloud +10.7 386/434 17/02/2006 Some cloud -26.5 387/434 06/01/2006 clear -13.1

The image dates for the Landsat 5 imagery are:

20

Table 4(a): Landsat 5 TM Image Dates for NSW Path/Row Date Cloud

89/81 08/04/06 10% cloud 90/81 14/03/06 Mostly clear

Table 4(b): Landsat 5 TM Image Dates for WA

Path/Row Date Cloud 111/83 08/04/06 30% cloud 112/83 14/03/06 Mostly clear

Table 4(c): Landsat 5 TM Image Dates for Tas

Path/Row Date Cloud 90/89 26/02/06 40% cloud 91/89 02/03/06 70% cloud

Each of the forest cover probability images was added to the existing sequence of probability images from 1972 to 2005 for multi-temporal processing. Common error rates were applied for the 2006 epoch as there was no evidence to suggest that the probability images from any of the three data sources were more or less accurate than usual. Forest extent and change maps were created from the outputs of the multi-temporal processing according to the usual procedures. Figure 13 shows the forest cover products for a region in the black soil stratification zone in the NSW test area (zone 7 in figure 12). The Landsat 5 TM image for 2006 is displayed in the top left. The SPOT 4 image for 2006 is displayed in the bottom left. The three forest cover extent maps are displayed together in the top right of figure 13. The forest extent map derived from the Landsat 5 data is displayed in red. The forest extent map derived from the Landsat 7 data is displayed in green and the forest extent map derived from the SPOT 4 data is displayed in blue. Where these products coincide, the display appears white (forest) or black (non-forest). Colours indicate that the area is labelled as forest by only one or two of the products. In particular, yellow shows where the products from Landsat 5 and Landsat 7 are identical, but the product from SPOT 4 differs. Red, green and blue indicate that the area is labelled as forest in only the Landsat 5, Landsat 7 or SPOT 4 product respectively. The image in the bottom right of figure 13 is an overlay of the 2005-2006 clearing layers from the three products. Again the Landsat 5 derived product is in red, Landsat 7 in green and SPOT 4 in blue. The extensive yellow regions show that the results from the Landsat 5 and Landsat 7 images are consistent. This observation holds throughout the test area. These areas are mapped as ‘cleared’ in the products derived from SPOT 4 (blue regions in the bottom-right image in figure 13). Regions coloured green can be observed in the combined forest cover image in figure 13. These tend to be narrow linear features or along the edge of regions mapped as forest in the Landsat 5 product. As has been observed with the Landsat 7 data used in the 2000 and 2002 epochs, forest / non-forest discrimination is slightly better in Landsat 7 than Landsat 5 imagery allowing such features to be mapped. They cannot be excluded without under-estimating the extent of forest cover compared to Landsat 5. The same trends were observed in the WA test area. Tables 5 to 7 show the area of forest and change for the NSW test area calculated from the three products for the more recent epochs. The observation that the 2006 forest extent is similar for the two Landsat products is confirmed, with the forest extent slightly larger from the Landsat 7 product. The SPOT 4 product shows a different 2006 forest extent and both higher clearing and revegetation rates in the 2005-2006 interval. Most of the extra revegetation in the SPOT 4

21

product appears along the edges of forest areas, particularly in the more mountainous regions. Some of it seems due the registration not being quite as good when extreme satellite pointing angles are combined with mountainous terrain.

Table 5: Forest Extent Comparison for the NSW Test Area Epoch Landsat 5

Forest Extent (ha) Landsat 7 Forest Extent (ha)

SPOT 4 Forest Extent (ha)

2000 901612.9 901554.2 901303.5 2002 903661.7 903371.4 903129.6 2004 905869.4 905343.1 904437.8 2005 912433.2 912809.1 909352.4 2006 914660.2 916070.6 910981.1

Table 6: Area of Vegetation Loss (Clearing Layer) Comparison for the NSW Test Area Interval Landsat 5

Clearing (ha) Landsat 7 Clearing (ha)

SPOT 4 Clearing (ha)

2000 - 2002 11961.8 12230.8 11922.4 2002 - 2004 9223.0 8992.2 9054.2 2004 - 2005 11597.5 12210.6 12388.1 2005 - 2006 7270.3 6633.9 9970.6

Table 7: Area of Vegetation Gain (Revegetation Layer) Comparison for the NSW Test Area

Interval Landsat 5 Regrowth (ha)

Landsat 7 Regrowth (ha)

SPOT 4 Regrowth (ha)

2000 - 2002 14010.6 14048.0 13748.4 2002 - 2004 11440.6 10964.0 10362.4 2004 - 2005 18161.3 19676.7 17302.8 2005 - 2006 9497.4 9895.4 11599.3

It should be noted that, as illustrated in figure 14 for the NSW test area, all regions of land cover change observed in the test areas were detected in all three products. The regions in which the SPOT and Landsat products differ tend to be edge pixels, narrow features and the least dense forest cover. As none of the high resolution Ikonos imagery overlaps this test site (and if it did it would show 2002/2003 cover rather than 2006 cover), it is only possible to say which is most consistent rather than most accurate for the areas that differ. For change detection, consistency is an important consideration.

22

Figure 13: Top Left: 2006 Landsat 5 TM image, bands 3, 5, 4 in BGR. Bottom Left: 2006 SPOT 4 image, bands 2, 4, 3 in BGR. Top Right: Overlay of forest extent maps for 2006 for Landsat 5 (red), Landsat 7(green) and SPOT 4 (blue). Bottom Right: Overlay of 2005-2006 clearing maps for Landsat 5 (red), Landsat 7(green) and SPOT 4 (blue).

23

March 2006 March 2006 January 2005 December 2005 Figure 14: Top Left: 2006 Landsat 7 TM image, bands 3, 5, 4 in BGR. Bottom Left: 2005 Landsat 5 TM image, bands 3, 5, 4 in BGR. Top Centre: 2006 Landsat 5 TM image, bands 3, 5, 4 in BGR. Bottom Centre: 2006 SPOT 4 image, bands 2, 4, 3 in BGR. Top Right: Overlay of forest extent maps for 2006 for Landsat 5 (red), Landsat 7(green) and SPOT 4 (blue). Bottom Right: Overlay of 2005-2006 clearing maps for Landsat 5 (red), Landsat 7(green) and SPOT 4 (blue). One consequence of using Landsat 7 SLC-off imagery should be noted. There was more extensive cloud cover in the Western Australian test area than in New South Wales. As cloud was masked before compositing the images, some stripes of missing data remain in the mosaiced image. The multi-temporal processing infers the cover in such gaps using the data from the surrounding epochs. In the case of the 2006 epoch, the cover is assumed to be unchanged from 2005. In the example from the WA test area shown in Figure 15, the stripes of missing data appear within areas of forest that were cleared between 2005 and 2006. Above and below the missing data, the cover is correctly mapped as non-forest and hence as cleared in the 2005-2006 clearing map. In the gap, the cover is extrapolated from the 2005 forest cover and is not mapped as change, resulting in stripes in the extent and change maps.

24

March 2006 April 2006

January 2005 February 2006 Figure 15: Top Left: 2006 Landsat 7 TM image, bands 3, 5, 4 in BGR. Bottom Left: 2005 Landsat 5 TM image, bands 3, 5, 4 in BGR. Top Centre: 2006 Landsat 5 TM image, bands 3, 5, 4 in BGR. Bottom Centre: 2006 SPOT 4 image, bands 2, 4, 3 in BGR. Top Right: Overlay of forest extent maps for 2006 for Landsat 5 (red), Landsat 7(green) and SPOT 4 (blue). Bottom Right: Overlay of 2005-2006 clearing maps for Landsat 5 (red), Landsat 7(green) and SPOT 4 (blue). 10. Conclusions The standard scene selection criteria should be applied to select the ‘primary’, or main, image for each path/row. The recommended criteria for selecting the ‘fill’ image are closeness to the primary image acquisition date and low cloud cover. A slightly increased cloud cover is considered better than a longer time between acquisitions. Two images are generally sufficient for each path/row, a primary image and a single fill image. As a rough guide, between 94% and 99.5% of the path/row extent is filled by compositing two individual images for these test areas. Operationally a third image may be required to achieve this level of coverage for a small proportion of path/rows, however the availability of reasonably cloud-free image data will be the limiting factor for the 2006 epoch.

25

Individual Landsat 7 SLC-off ‘path’ images can be adequately registered and calibrated to the AGO base images with minimal additional ‘per image’ effort. Further investigation of alternative gain and offset correction strategies between images from the same path/row from different seasons is required. The ACRES Landsat 7 SLC-off composite products are not always well registered to the AGO rectification base. The displacement appears regional with up to 2 pixels (50m) in the NSW test area, but good matches in Western Australia and Tasmania. In the NSW test area, the displacement was observed between the ACRES composite and the NSW state DEM, causing the terrain-illumination correction to be less successful. A thorough comparison of the AGO and ACRES bases would be required if this product is to be considered further. The local calibration performed as part of the ACRES compositing process is also of concern. The ACRES process can make quite different cover types (cloud / no cloud and dry / wet) look identical. While this makes a good ‘picture’ product, it is of concern when trying to map ground cover. Some modifications to the mosaicing methodology are required when compositing Landsat 7 SLC-off images for a given path/row in combination with mosaicing path/row images to form map sheet units. A new method for tracking image date boundaries within the mosaics needs to be developed. It is strongly recommended that cloud masking of the Landsat 7 SLC-off images be performed before they are composited to significantly reduce the effort involved in cloud masking. This will have implications for forming national picture mosaics. Such a change will increase the mosaicing effort, but a reduction in the thresholding effort will offset the change. Where there is no significant change between the images being composited, the thresholding can be performed using the standard methodology, treating each composite as a single path/row image. In the rare event that differences have been flagged between images to be composited, thresholds should be derived from the individual images and mosaiced. Assuming that cloud-masking is being performed as part of the mosaicing stage, there will be no overall increase in thresholding effort and generally the effort will decrease. The forest cover extent and change products from the Landsat 7 SLC-off data are consistent with those obtained from the Landsat 5 data with a slight increase in forest extent, consistent with those observed in the 2000 and 2002 epochs in which Landsat 7 data was also used. The forest extent and change products formed from SPOT 4 image data had the least consistency with the Landsat 5 products. If missing data remains in any Landsat 7 SLC-off composites, the products may include stripes in the forest extent and change data if missing data partially cover areas of land cover change. 11. References Furby, S. L. (2006), Documentation for the 2005 Update of the Forest Cover Mapping for the Australian Greenhouse Office Land Use Change Program, CSIRO Mathematical and Information Sciences Technical Report 06/43. Furby, S. L., Wu, X. and O’Connell, J. (2006), Evaluation of SPOT 4 Image Data for Forest Cover Mapping, CSIRO Mathematical and Information Sciences Technical Report 06/155.