Estimating leaf area index from satellite imagery using Bayesian networks

In this study, we investigated the use of Bayesian networks for inferring tropical dry forest leaf area index (LAI) from satellite imagery in dry and wet seasons. LAI was chosen as the variable of interest because leaf area is the exchange surface between the photosynthetically active component of the canopy and the atmosphere. Initial network estimates were obtained from ground truth plot data with known forest structure, LAI, and satellite reflectance in the red and near-infrared bands (as observed by the Landsat 7 Enhanced Thematic Mapper Plus sensor). We tested the performance of the Bayesian networks with scoring rules and also with confidence and surprise scores. We evaluated the networks on a per-pixel basis and created both LAI maps of the study area as well predicted the probability maps for the highest LAI states. Results not only demonstrate the predictive power of a Bayesian network but also its explanatory power which is far beyond what is typically available with current pixel classifier approaches such as spectral