Methods .. Landsat Classification

Landsat classification methodology

After an initial classification carried out with MODIS MOD43B4 imagery a higher resolution approach was obtained by classification of Landsat ETM+. The Enhanced Thematic Mapper Plus (ETM+) on board of Landsat-7 provides 15-meter resolution “panchromatic” data and six bands in the visible, near-IR and mid-IR spectral regions at a resolution of 30 metres.

	Map of the path and row for Landsat (WRS)

Using the Worldwide Reference System (WRS), the global notation system for Landsat data, Madagascar is located between: Path 157 to 161 and row 068 to 078, comprising 37 Landsat scenes. For each Landsat scene 3 to 5 images are available, dated between 1999 and 2003.

The classification of Landsat images was performed using to the following steps:

1. Pre-processing: The entire data set, acquired from different sources and available in different formats, was first imported and layer stacked. Single multilayer images were conformed including the Landsat bands to be used in the classification (1,2,3,4,5 and 7). Also images were unified (reprojecting or rectifying when needed) to a common projection. Standard Universal Transverse Mercator sectors 38-39 South is used as the uniform projection system for this project.
   
2. Prioritisation of images used for classification: For each Landsat scene two to four different images are available. In most cases they were acquired in different seasons, providing valuable seasonality information. The criteria used for selecting images was:
   a. Scene quality in terms of cloud freeness
   b. Season requirements. When Landsat images for both dry and wet season are available the decision of which one to use was made upon peculiarities of the vegetation in that certain area.
   c. Date. Trying to use the latest image available.
   
3. Image classification: A supervised maximum likelihood classification algorithm was used within Erdas Imagine 8.6. The training sample was obtained using a digitised version of Faramalala’s vegetation map, the initial Modis classification and the Conservation International deforestation map. Also Landsat images from a different season were used in order to provide seasonality information (NDVI analysis and visual interpretation). The scheme used for this classification tried to separate 11 broad classes to be subsequently refined in a GIS environment.
   
4. Cloud-cover analysis: Although Landsat images were selected to have a minimal cloud cover, Madagascar’s east coast has frequent cloud cover. Of the Landsat images 35% (mainly centred in north east part of the country) used for classification have cloud problems. Clouds and their associated shadows provide incorrect reflectance values of the features in the earth surface and disrupt the classification. Therefore pixels covered by clouds or shadows were excluded from classification.
   First a cloud mask was developed for each image by running an unsupervised classification looking for 250 classes (ISODATA algorithm, 6 itineration, 0.950 convergence threshold). Classes representing either clouds or shadows were identified and the classified images recoded to conform a binary mask. Manual modification was carried out wherever misclassifications were identified. These binary images were used to mask out clouds and cloud shadows from the images prior classification.
   In each scene with cloud problems successive Landsat images were classified to replace cloud holes until the scene was complete or no more data was available. As a result cloud problems were minimised using the entire set of Landsat images available.
   
5. Post classification processing: A 3 x 3 pixel majority filter was applied to the output images in order to smooth the classification. Finally, classified images were assembled to conform a classified mosaic.