Multi-site evaluation of the MOD17 model over homogeneous agricultural vegetation

Agricultural lands play important role in the ecosystem-atmosphere interactions due to their large extent. Cultivation patterns in agricultural landscapes are frequently complex, making efforts to scale up local, patch scale flux tower measurements to regional or continental scale quite challenging. In particular, it is important to evaluate the spatial representativeness of flux measurement sites in agricultural lands. Remote sensing offers spatially representative radiance measurements that can be used to study ecosystem processes using sophisticated models, and can also be used for upscaling purposes.

We present here a multi-site evaluation of the performance of the well-known MOD17 product in case of croplands. One major drawback of MOD17 is that only a single agricultural category ('Cropland') is used globally. The generalized cropland parameterization is unable to capture the special features of croplands (different light use efficiency, drought resistance, etc.) which cause biased estimates. To provide critical evaluation of the model our first aim was to evaluate the model results with different crop types. Our basic consideration for choosing the sites was to cover the most typical C3 and C4 crops as well as various geographical locations. In order to avoid footprint problems, only short towers over homogeneous vegetation are included in our examinations. As a result, several EUROFLUX and AMERIFLUX sites participate in the study.

The mismatch between the spatial representativeness of the short flux tower and the 1 km spatial resolution MOD17 pixel (determined by the MOD15 FPAR product's pixel size) can lead to misinterpretation of the results. To overcome this problem, we applied a simple downscaling technique based on NDVI – FPAR relationship. MOD17 results were compared to Gross Primary Production (GPP) derived from flux tower measurements. In order to derive a standardized GPP dataset, we applied the same gap filling and flux partitioning methods on non-gap filled data from all sites. Flux partitioning was carried out using four methods in all sites in order to avoid methodological inconsistencies. The results also provide information on the uncertainty of the GPP estimations.