Deriving a Forest-cover product over the Lebialem Highlands region, Cameroon: Image processing methodology employed – Final Version Produced for the African Conservation Foundation (ACF) By Lauriane Cayet-Boisrobert, MSc. ([email protected] ) & Magali Moreau, MSc. and PhD candidate ([email protected] ) Volunteers for ACF, through the UN Online Volunteer Program February 24th, 2007 1. Abstract This aim of this work was to produce land cover classification data for the African Conservation Foundation’s Lebialem Highlands Great Apes Research and Conservation Program in South-West Cameroon The data was created from two subsets of Landsat 7 Enhanced Thematic Mapper (ETM+) scenes from 2003, using ERDAS Imagine, a digital image processing software package. Contrast-stretching and noise-removal were performed to enhance the image quality, and 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 landcover categories: forested and non-forested areas. Since no ground-truthing data was available, it was not possible to analyse this further, or to perform an accuracy assessment. However, comparison with the Global Land Cover 2000 dataset for this area revealed that the classification was relevant. 2. Rationale The Central African Regional Program for the Environment (CARPE) did not produce any land cover data over the area under study. The Global Land Cover 2000 (GLC2000) dataset, developed by the Joint Research Centre of the European Union, did not have an appropriate resolution for the area of study, but it was used here as a comparison product. It was thus necessary to produce a land-cover for the level of the area under study and suitable for the objective of the conservation project. 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. Therefore, we decided to produce a raster of pixels corresponding to either forest or non-forest only. 3. Data overview Two Level-1G Landsat 7 Enhanced Thematic Mapper Plus (ETM+) images were selected and downloaded from the GLCF website1. Landsat 7 ETM+ comprises 5 channels corresponding to the visible and short-wave infrared bands, as well as a panchromatic channel, which are of interest for the study (Table 1). The area of study required 2 scenes: the first scene corresponding to path 186 and row 56 and the second scene corresponding to path 157 and row 56. The images showing too many clouds over the area of study were discarded. Only one image (almost cloud-free) for each scene was available. The acquisition dates were 1st of October 2003 for p186r56 and the 1st of January 2003 for p187r56, both on the ascending passes of the satellite. The two illumination angles were very different and some haze was present in both images. Landsat 7 ETM+ data were processed by Earthdata to Level 1G. This means that the data was already radiometrically and geometrically corrected automatically. However, certain errors were still present, mostly due to atmospheric conditions in the study area. Finally, the data was not co-registered to any other Landsat imagery 1

The source for this dataset was the Global Land Cover Facility, http://www.landcover.org

products (Landsat TM, MSS) and terrain corrections have not been applied to orthorectify the images, due to scale differences between the DEM product created previously and the Landsat scenes used here. Table 1: Landsat 7 channels

B1 B2 B3 B4 B5 B6 B7 B8

0.450-0.515 (30m) Blue 0.525 – 0.605 (30m) Green 0.630 – 0.690 (30m) Red B4 : 0.750 – 0.900 (30m) NIR 1.550 – 1.750 (30m) NIR 10.400 – 12.500 (60m) TIR 2.090 – 2.350 (30m) (NIR) 0.520 – 0.900 (15m) Panchromatic

4. Image preparation 4.1. Importing into ERDAS Imagine The images downloaded from the GLCF website were in binary format (.bsq). The method described on the website (using the import tool and select type as “Generic Binary”) did not work. Therefore, it was necessary to transform the binary format into Landsat 7 Fast EROS data type. This format is supported by the Import tool in ERDAS Imagine. This was easily done by changing the naming of the files. The convention is: L7fppprrr_rrrYYYYMMDD_AAA.FST Where L7 = Landsat 7 mission; f = ETM+ format (1 or 2) (data not pertaining to a specific format defaults to 1); ppp = starting path of the product; rrr_rrr = starting and ending rows of the product; YYYYMMDD = acquisition date of the image; AAA = file type: HPN = panchromatic band header file, HRF = VNIR/ SWIR bands header file, HTM = thermal bands header file, B10 = band 1, B20 = band 2, B30 = band 3, B40 = band 4, B50 = band 5, B61 = band 6L, B62 = band 6H, B70 = band 7, B80 = band 8; FST = FAST file extension. The panchromatic and visible and short wave Infrared bands were imported. The import tool only asked for the header files of the panchromatic, as well as the visible and shortwave infrared channels. The import tool created a layer stack of the bands 1,2,3,4,5, and 7 (called a multi-spectral image) and a panchromatic image. The ephemeris geographic Lat/Long coordinates were obtained from the ETM+ image header and are utilized with the Landsat 7 geometric model available in ERDAS Imagine 8.7.

4.2. Creating an image subset Once the images were imported, subsets of each scene were created for space limitation reasons – each scene was divided in half. 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. 4.3. Pre-Classification Image Processing Before the processing and analysis of the Landsat data can occur, certain pre-processing routines were applied to the images and some were attempted. The aim was to improve the quality of the image data by enhancing the spectral and the radiometric information, correcting radiometric noise, as well as improving the resolution of the image. 4.3.1. Radiometric enhancement Radiometric enhancement functions enhance the image using the values of individual pixels within each band. The techniques used in this study were contrast stretching and noise reduction. Histogram equalization was done using the General Constrast tool in the Raster options in the ERDAS Imagine viewer. The noise reduction tool removes noise using an adaptive filter, and was done using the Image Interpreter – Radiometric Enhancement tool in ERDAS Imagine. Both of these enhancements were applied to the Landsat scenes. There were a few clouds on the images, mostly on scene p187r56, and atmospheric scattering was present, also mainly on scene p187r56. It was important to reduce the haze effect, to reduce chances of misclassification. It was thus essential to apply a haze reduction algorithm. Such an algorithm exists in ERDAS Imagine, however, this function is only valid to remove haze from Landsat 4 and 5 TM data and Panchromatic data, so we were unable to apply it to our Landsat 7 ETM+ data. 4.3.2. Spatial enhancement While radiometric enhancements operate on each pixel individually, spatial enhancement modifies pixel values based on the values of surrounding pixels. Spatial enhancement focuses on spatial frequency. The spatial enhancement technique that is often used in classification studies is resolution merge. Resolution merge (pan-sharpening) techniques improve the resolution, and thus, the visual result of classifications. According to Zhang and Wang (Zhang 2004) resolution merge is believed to increase the accuracy of classification in urban areas. The use of Landsat ETM+ pan-sharpened imagery has proved to be an effective tool for detecting new urban developments, since it allows the clear distinction of land cover classes (Finney 2004). Thus, resolution merge is important when the objective is to detect roads, as well as obtain clear shapes of parcels, buildings, and urban developments. For the purposes of this work, there was no need to achieve such a resolution improvement. Furthermore, it has been demonstrated that fused multi-spectral and panchromatic data (i.e. resolution merge or pan-sharpening) do not have a significant influence on the improvement of land-use classes during the classification process when using bands 2,3,4 (Lewinski 2006). Bands 3 and 4 were both used in the study. 4.3.3. Spectral enhancement Spectral enhancement is the changing of the values of each pixel in the original image by transforming the values of each pixel along a multi-band basis. Spectral Enhancement allows different features that have specific reflective characteristics in different bands to be compressed if data is similar. It also allows modifying of the pixels of an image independent of the values of surrounding pixels. Essentially, spectral enhancement creates new bands of data that are more interpretable to the eye. For technical reasons, we were unable to perform certain spectral enhancement functions such as Principal

Components Analysis or Tasseled Cap Transformation2. A Normalized Difference Vegetation Index (NDVI) was created using the Image Interpreter, Spectral Enhancement tool in ERDAS Imagine to have a better idea of vegetated areas on the images. This helped to interpret the classification results. 4.4. Masking Both images were affected by clouds and shadows, as well as 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 clouds, large shadows and water bodies. The masks were created by looking at certain bands 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. Band 1 allowed for the best discrimination of clouds, and was thus used to create the cloud mask. Band 4 was used to determine the threshold value for water bodies. Since water absorbs nearly all light at the wavelength associated with this band, water bodies appear very dark. This contrasts with bright reflectance for soil and vegetation so it is a good band for defining the water/land interface. Band 4 was also used to create a mask for large shadows. Once 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 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 utilize data from multiple bands. 5.1. Creating Clusters In ERDAS Imagine, unsupervised classification is performed using an algorithm called the Iterative Self-Organizing 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, four clusters were created for p186r56 and six were created for p187r56, 9 iterations were performed with a confidence level of 95% and areas of zero value were not used in the clustering process. 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, we were only able to accurately distinguish classes representing forested and non-forested areas, this latter class including other vegetation types, bare areas, and built areas. Comparing the classified image with the GLC2000 revealed that the ‘forest’ clusters corresponded to areas with Montane Forest, Sub-Montane Forest, and lowland Mosaic of Savannas and Forest (Mayaux 2003).

2

For more information on PCA, see Nikolakopoulos (2004)

5.3 Merging clusters This resulted in both images being classified in two main clusters: forested and non-forested areas 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 The result of the classification process may create a lot of isolated pixels surrounded by a mass of pixels belonging to the same class. For practical visualization, it is better to apply the modal adaptive filter available in ERDAS Imagine under Interpreter > radiometric enhancement. It must be noted that for image mosaics, it is desirable to have radiometrically matched images, which requires either using images taken at approximately the same time and date or using a Bidirectional Reflectance Distribution Function (BRDF) model. As we do not have any BRDF model and the images were taken under very different solar illumination conditions (at two different dates), it is preferable to mosaic them after classification. Hence, the two classified products were joined using the Mosaic tool in ArcGIS. To have a better understanding of the image, raster images of the clouds, water and shadow masks were this mosaic using the same Mosaic tool. The output represents areas of forest (value=1), non-forest or ‘other’ (value=2), clouds (value=10) and water/shadows (value=100), and was named lh_forestcover. There is an associated layer file to give the names of the classes, and to represent the final output with appropriate colours.

REFERENCES Finney, K. (2004). Utilizing Pansharpened LANDSAT 7 Imagery to Detect Urban Change in the Vancouver Census Metropolitan Area: 1999-2002. Graduate Program in Spatial Analysis Research Paper. Toronto, Ontario, Canada, Ryerson University. Lewinski, S. (2006). Applying fused multispectral and panchromatic data from Landsat ETM+ to object oriented classification. Proceedings of the 26th EARSeL Symposium, New Developments and Challenges in Remote Sensing, Warsaw, Poland. Mayaux, P., Bartholomé, E., Massart, M., Van Cutsem, C., Cabral, A., Nonguierma, A., Diallo, O., Pretorius, C., Thompson, M., Cherlet, M., Pekel, J-F., Defourny, P., Vasconcelos, M., Di Gregorio, A., Fritz, S., De Grandi, G., Elvidge, C., Vogt, P. and Belward, A. (2003). "A land-cover map of Africa, Carte de l'occupation du sol de l'Afrique." Publications of the European Communities. Nikolakopoulos, K. G. (2004). Pansharp vs. wavelet vs. PCA fusion technique for use with Landsat ETM panchromatic and multispectral data. Image and Signal Processing for Remote Sensing X. L. Bruzzone, SPIE Proceedings. 5573: 30-40. Zhang, Y., Wang, R. (2004). Multi-resolution and multi-spectral image fusion for urban object extraction. XXth ISPRS Congress, Istanbul, Turkey.