Physically‐based models of vegetation reflectance serve as a basis for extracting vegetation variables using directional and spectral data from modern‐borne sensors (e.g., MODIS, MISR, POLDER, SeaWiFS). Although many models have been inverted, only recently have significant efforts been made to provide operational algorithms. These efforts have exposed a need to significantly improve efficient and accurate methods for inverting these physically‐based models. The characteristics of the traditional inversion, table look‐up, neural network and other methods are discussed as well as the major achievements, advantages/disadvantages, and research issues for each method. The traditional inversion methods using repeated model runs are computationally intensive and are not appropriate for operational application on a per pixel basis for regional and global data. Thus, for larger data sets, simplified (reduced number of variables and/or physical processes) physically‐based models are generally used. The table look‐up and neural network methods are computationally efficient and can be applied on a per pixel basis. Moreover, they can be applied to the most sophisticated models without any simplifications. Finally, they do not require initial guesses to model variables as do the traditional inversion methods. However, traditional inversion and table look‐up methods are inherently designed to handle any arbitrary set of Sun‐view angles. Neural networks have not been generalized, as of yet, to handle arbitrary angles. We believe the most pressing research priority is to perform a rigorous comparison of the various inversion methods in terms of accuracy and stability, computational efficiency, general applicability, and number of variables obtainable.

Leaf area index (LAI) is an important surface variable for monitoring the status of vegetation and as input in a number of ecosystem process models. There are currently several coarse-resolution LAI maps over Canada, including a Canada Centre for Remote Sensing ten-day, 1-km resolution, Canada-wide product based on SPOT-4 VEGETATION (VGT), a MODIS eight-day, 1-km resolution, global product and a monthly, 7-km resolution, global map produced using POLDER-1. These products are difficult to validate because of their large spatial extent and coarse resolution. In this study we use in situ LAI measurements collected over a wide range of forest types and ecological zones in Canada to derive 30-m resolution reference LAI maps based on robust error-in-measurement regressions to Landsat Enhanced Thematic Mapper Plus vegetation indices. The reference maps and LAI products were aggregated to a coarser resolution (3 km for MODIS and VGT and 7 km for POLDER) before comparison to account for registration errors, and variability in sensor projected point spread functions. Spatially corresponding aggregated pixels with both high-quality reference and coarse scale LAI retrievals were compared. The comparison shows reasonable agreement (biases less than 25% or one LAI) between the VGT and reference LAI. The MODIS LAI product showed weak correlations (R²<0.25) over all sites at the scale of comparison and typically overestimated reference LAI in mixed forests by approximately 200%. The POLDER LAI product, only available in June 1997, showed almost no correlation to the summer reference LAI datasets. It underestimated reference LAI for an early growing season with an extent, in some cases, greater than the seasonal differences in LAI. This independent validation of three large area LAI products suggests that there may be substantial biases due to the lack of regional tuning of retrieval algorithms. These biases are far larger than the uncertainties in the reference-based LAI scenes in the case of the MODIS product. This suggests that reliable LAI maps may require regional calibration to meet the Global Terrestrial Observing System mapping requirements of plusmn15% uncertainties