A Bayesian network algorithm for retrieving the characterization of land surface vegetation

A hybrid inversion technique based on Bayesian network is proposed for estimating the biochemical and biophysical parameters of land surface vegetation from remotely sensed data. A Bayesian network is a unified knowledge-inferring process that can incorporate information derived from multiple sources including remote sensing and information derived from a priori knowledge. Using this inversion approach, content of chlorophyll a and chlorophyll b (Cab) and leaf area index (LAI) of winter wheat were estimated from data derived from simulations as well as field measurements. Estimations from the simulated data proved accurate, with root mean square errors (RMSEs) of 0.54 m2/m2 in LAI and 4.5 μg/cm2 in Cab. In validating the estimates against field measurements, it was found that prior knowledge of target parameters improved the accuracy of estimates, in terms of RMSEs from 0.73 to 0.22 m2/m2 in LAI and 9.6 to 4.0 μg/cm2 in Cab. Bayesian inference in this hybrid inversion scheme produces a posterior probability distribution, which can reveal such properties of the inferred results as updated information contained in the inversion result. Using entropy, the revision of posterior information about the parameters of interest was calculated. Including more data may allow more information to be retrieved about parameters in general. Exceptions were also observed where data from some viewing angles slightly reduced the information on the parameters of interest. It was also found that data from these viewing angles were less sensitive to the parameters. The method proposed here was also validated using LandSat ETM+ imagery provided by the BigFoot project. When used for mapping LAI with ETM+ imagery, the proposed method with an RMSE of 0.70 and a correlation of 0.67 produced a slightly better result than that from empirical regression.