Adaptive data parallel methods for ecosystem monitoring

Biological diversity is decreasing at an alarming rate worldwide. Improved ecosystem monitoring can help detect problems in time to intervene. Earth orbiting satellites collecting terabytes of imagery daily, can support effective monitoring of many habitats. Data parallelism is ideal for many automated image analysis algorithms, but less natural for the complex spatial structure of most ecosystems. This paper presents a coarse-to-fine processing framework: based on a set of spatial transformations, that compact disconnected regions to achieve more efficient nested data parallelism. Experiments with a montane island ecosystem in southeast Arizona use Landsat TM data to characterize the processing framework the spatial transformations, and the feature extraction algorithms