Modeling biomass of forests in the southwest Amazon by polar ordination of Landsat TM

Polar ordination is used to project all pixels in Landsat TM band space onto a single axis whose endpoints are the centroids of training areas in two spectrally distinct natural forest types - one with and the other without bamboo - which dominate the landscape in the southwest Amazon. Assuming the same wood density in both forests, biometric data collected in the training areas (Oliveira, 2000) indicated that biomass of forest with high bamboo density was 29% lower than forest without bamboo. Linear interpolation along the ordination axis models biomass of spectrally intermediate pixels. Thresholds based on band-space distance from the axis and ordinated distance beyond the axis endpoints are used to mask pixels unacceptable for modeling: water, beaches, pastures, roads, urban area, and deep topographic shade. The final product is a biomass image of an area covering 90x90 km, with 79% of all pixels being modeled. Because of spectral variability between bamboo forests in different recovery stages after their last monocarpic dieback, the model should only be applied to landscapes with mature, spectrally stable bamboo populations.