Detecting Forest Cover Change between the 1980’s and 2000’s over the Lebialem Highlands region, Cameroon: Image processing methodology employed Produced for the African Conservation Foundation (ACF) Lauriane Cayet-Boisrobert, MSc. ([email protected] ) Magali Moreau, MSc. and PhD candidate ([email protected] ) Volunteers for ACF through the UN Online Volunteer Program March 18, 2007

1. Abstract The aim of this work was to produce a map of forest-cover change for the African Conservation Foundation’s Lebialem Highlands Great Apes Research and Conservation Program in South-West Cameroon. The data was created from subsets of GeoCover, orthorectified Landsat 4 TM, Landsat 5 TM and Landsat 7 ETM+ scenes from the 1980s and 2000s, using ERDAS Imagine, a digital image processing software package. Atmospheric correction (and contrast-stretching) was performed to enhance the images quality. Masks were created to remove clouds, large shadows and water bodies from the images. An ISODATA unsupervised classification was performed and clusters were classified into two main land-cover categories: forested and non-forested areas. The classified results were mosaicked for both time-periods, and an image difference map was created, taking into account the masks for all periods. Since no groundtruthing data was available, it was not possible to analyze this further, or to perform an accuracy assessment.

2. Rationale Land cover change, as an environment indicator, is essential to the Lebialem Highlands Great Apes Research and Conservation Program. The Central African Regional Program for the Environment (CARPE) did not produce any land cover data over the area under study. It was thus necessary to produce maps of forest-cover for two time periods for the area under study, to highlight areas of forest-cover change in the study area. Additionally, no training sites were provided. This prevented us from running a supervised classification, and also from performing an unsupervised classification followed by assigning clusters to classes corresponding to the field. Two rasters of pixels corresponding to either ‘forest’ or ‘nonforest’ were produced and used to create the final forest-cover change map.

3. Data overview Landsat sensors are the only remote means to obtain free historical data covering a long period of time. However, Landsat images corresponding to the time-period of interest originated from different platforms and sensors. To avoid co-registration problems and radiometric variations between sensors and platforms, the Landsat GeoCover dataset was used in this study. Landsat GeoCover images are a collection of 28.5mpixel spacing imagery provided in a standardized, orthorectified format (processed by EarthSat). The GeoCover dataset was downloaded free of charge from Global Land Cover Facility (GLCF) website1. Only two images per date were available. 1

http://www.landcover.org

Table 1: Overview of the data source Path/Row 186/56 186/56 187/56 187/56

Platform Landsat-4 Landsat-7 Landsat-5 Landsat-7

Sensor TM ETM+ TM ETM+

Acquisition date 1988-02-02 2001-02-05 1986-12-12 2000-12-10

Landsat sensors cover the broad range of the light’s spectrum from the Visible to the Near Infra Red channels relevant for forest detection. Table 2 shows the different Landsat bands used for this study. For the final output, bands 3, 4, and 7 were used in the unsupervised classification to detect areas of forest and non-forest, while Bands 1 and 4 were used to make the masks. Table 2: Landsat 4-5 TM and Landsat 7 ETM+ bands Bands Band 1 Band 2 Band 3 Band 4 Band 5 Band 7

Wavelength (micrometers) 0.450 – 0.515 (30m) Blue 0.525 – 0.605 (30m) Green 0.630 – 0.690 (30m) Red 0.760/0.780 – 0.900 (30m) NIR 1.550 – 1.750 (30m) NIR 2.080/2.090 – 2.350 (30m) (NIR)

Resolution (metres) 30 30 30 30 30 30

4. Image preparation 4.1. Creating Image Subsets Once the images were imported, subsets of each scene were created for space limitation reasons – each scene was approximately divided in half, taking into account the location of the areas under study. This was done by creating an Area of Interest using the AOI tool in the Viewer and using the coordinates of this Area of Interest to clip the images using the ERDAS Imagine Subset tool. 4.2. Pre-classification Image Processing 4.2.1. Radiometric enhancement Atmospheric scattering was very present in the dataset used for this study, so an atmospheric correction technique was used to mitigate the reduction in image contrast due to this factor. The compensation for scatter calls for subtracting brightness values from a scene. This technique consists of offsetting the band’s distribution histogram by its minimum value. The result is a better distributed histogram over the whole range of digital number (255) and an increase in visual contrast. This was done by looking at the histograms for each band used (for each scene), and determining the amount by which to offset it. The Offset Values tool in ERDAS Imagine was used, and each band was saved. The band statistics were then re-computed. Since this procedure removes the lookup tables associated with the bands, these needed to be re-created. This involved manipulating the image contrast and brightness until the display was satisfactory and saving the changes. 4.2.2. Haze reduction and stripes reduction The Red band (band 3) is crucial in forest detection, and is often an important component in vegetation studies based on remote sensing. In the dataset available for this study, the 1986 scene with path 187, row 56, contained important striping in the bands in the visible range: bands 1, 2 and 3. Haze was also present on the three bands, but less on band 3. A Principal Component Analysis (PCA) technique was used to

compress the most relevant decorrelated information of the three bands of interest in the first Principal Components (PCs), while excluding correlated information in the last PC. This technique can be used to remove noise. A PCA was applied to bands 3, 4, 7, with three outputs. The first Principal Component image (PCA1) showed mostly land-cover types and details. The second Principal Component image (PCA2) showed mostly vegetation (bright). The third Principal Component image (PCA3) showed haze and stripes clearly. The table of eigenvectors showed that band 3 and 4 (and thus vegetation) was well represented by PCA1 and PCA2. Since PCA3 comprised most of the noise, and most of the relevant information was well represented by PCA1 and PCA2, the classification process only used these two PCs instead of using band 3, 4, 7. This procedure was only performed for this particular scene.

Figure 1: Visual results of applying a Principal Component Analysis to bands 3, 4, and 7 of the 1986 Landsat scene with path 186, row 57.

4.3. Masking 4.3.1. Clouds Masks Clouds and haze are mainly visible on the channels within the visible part of the light spectrum, and haze mostly visible within bands 1 and 2. In order to detect clouds, the bands in the visible part of the light spectrum can be used. With the red channel (band 3), ground pixels with high reflectance (urban areas/populated places, sloped mountain faces) had a similar reflectance as clouds on our particular scenes. This issue is to be related to the high viewing angle of the sensor used to take the images. Therefore, band 3 could not be used to mask clouds. Band 1 allowed for a strong discrimination of clouds from high-reflectance ground pixels such as urban areas/populated places and illuminated slopes. Additionally, since clouds appear with a higher reflectance than haze, they can easily be discriminated from haze in band 1. A threshold brightness value in band 1 was determined in order to discriminate clouds from all other pixel types, for each scene. These threshold values were then used as inputs to a conditional function in ERDAS Imagine’s Model Maker tool to mask clouds from bands 3, 4, and 7.

Figure 2: Varying detection of clouds and haze in NW and SW subsets of scene p187r56 (2001) depending on the band used: band 1 (left) and band 3(right). The upper images show that band 1 is better to discriminate clouds from bare soil with high reflectance than band 3. The lower images show that haze is more prominent in band 1 than in band 3. 4.3.2. Water and Shadow Masks Both images were also affected by shadows and contained water bodies that could have an effect on the classification. In order to reduce chances of erroneous classification, masks were applied to each image to remove these areas. The water and shadow masks were created by looking at the near infra-red band (band 4) and determining thresholds for brightness values associated with each category. These threshold values were then used as inputs to a conditional function in ERDAS Imagine’s Model Maker tool and the mask was applied to bands 3, 4 and 7. Once all the masks were created, image arithmetic was used through the Model Maker to multiply these masks with the original images. This resulted in images free from clouds, water bodies and large shadows.

5. Unsupervised classification Since we had limited information about the study area, we classified the data using ISODATA unsupervised classification. Unsupervised classification is a method that examines a large number of unknown pixels and divides them into a number of classes (or clusters) based on natural groupings present in the image values. The basic premise is that pixels from the same cover type should be close together in the spectral measurement space (i.e. have similar digital numbers), whereas pixels from different cover types should be comparatively well separated in spectral space (i.e. have very different digital numbers). This process can simultaneously use data from multiple bands.

5.1. Creating Clusters In ERDAS Imagine, unsupervised classification is performed using an algorithm called the Iterative SelfOrganizing Data Analysis Technique (ISODATA). For this algorithm, the number of clusters desired and a confidence threshold are specified by the user. Clusters are built iteratively: with each new iteration, the clusters become more refined. The iterations stop when the confidence level (or a maximum number of iterations specified by the user) is reached. The ISODATA tool was used to do this for each scene, using bands 3, 4 and 7. The masked image was used as input, six clusters were created for each scene, nine iterations were performed with a confidence level of 95% and areas of zero value were not used in the clustering process. This resulted in 4 raster images: p186r56_88_MSK_unsup, p186r56_01_MSK_unsup, p187r56_86_MSK_unsup and p187r56_00_MSK_unsup. 5.2. Defining land-cover classes The result was then compared with a 7-4-3 color composite of the same area and the NDVI images, as well as with a GLC2000 image of the area, to associate the clusters with forested and non-forested areas. From this information, spectral classes were assigned to information classes by using the Raster Attribute editor in the viewer containing the classified image. Due to a lack of ground-truth data, it was only possible to accurately distinguish classes representing forested and non-forested areas, this latter class including other vegetation types, bare areas, and built areas. 5.3 Merging clusters This resulted in all images being classified in two main classes: forested and non-forested areas. This was done by opening the classified images in ArcGIS and using the Spatial Analyst Reclass tool to merge clusters representing forested areas together.

6. Post-classification processing 6.1 Creating forest-cover maps for the 1980s and 2000s For each time period (1986-1988 and 2000-2001), the classified scenes of path 186 and 187 and row 56 were joined using the Mosaic tool in ArcGIS. This resulted in images representing areas of forest and nonforest for 1986-88 (lh_forest_1986_88), and 2000-01 (lh_forest_2000_01). The output represents areas of non-forest (value=1) and forest (value=2). The background and masks have a value of 0. 6.2 Creating a forest-cover change map There are several methods to identify forest-cover change. Some of the most common methodologies are based on differences in the values of the red channel (Landsat band 3), or differences in NDVI, both of which are often used to detect the presence of vegetation. However, due to the poor quality of the available data from the 1980’s, a method based on changes in band 3 was not appropriate. It was determined that we would have a general idea of changes in forest-cover by creating an image difference map of the forest/non-forest classified images. In order to identify areas of change correctly, masks corresponding to a scene for one year were applied to the corresponding scene in the other year. For example, the masks originally applied to the Landsat 4 TM image p186r56 of 1988 were applied to the Landsat 7 ETM+ image of 2001, and vice-versa. This insured that no change would be erroneously detected in either of the masked areas. This was done on all classified images using the ERDAS Imagine Modeler tool. The final masked images were joined in ArcGIS using the method described in section 6.1. The Math tool in Spatial Analyst was then used to subtract the classified image of 1986-88 from the classified image of 2000-01. The output is called lh_forestdiff, and the classes represent deforestation (value = -1) and reforestation (value = 1), while areas of no change and masked areas have no value (value = NoData).