Comparison of regression and geostatistical methods to develop LAI surfaces for NPP modeling

Graduation date: 2002This study aims to compare different methods of obtaining maximum growing season leaf area index (LAI) maps using remote sensing data, LAI and tree cover field data in a boreal forest near Thompson, Manitoba, Canada. The comparison includes aspatial methods such as traditional regression, inverse regression and reduced major axis, and spatial methods such as kriging, cokriging, kriging with an external drift, and conditional simulation. The LAI maps will serve as input in process models to obtain maps of net primary production (NPP).\ud The present work was done in the context of the BigFoot project (http://www.fsl.orst.edu/_larse/bigfoot) which focuses on the validation of the MODIS (Moderate Resolution Imaging Spectrometer) land cover, LAI/fAPAR\ud (fraction of absorbed photosynthetically active radiation), and NPP products\ud (http://modarcKgsfc.nasa.gov/MODIS, with the main objective of scaling up from in situ ground measurements to the moderate spatial resolution of MODIS data products (250 - 1000 m spatial resolution). Due to the clumped structure of the boreal forest and the presence of a highly reflective understory, vegetation indices derived from remotely sensed data were not useful in explaining LAI variability. The use of mid-IR bands and tree cover data improved the performance of the models. Kriging with an external drift performed better in the presence of trends and anisotropy. An integrated aspatial (reduced major axis)/spatial (cokriging) method produced a useful compromise between local accuracy and pattern representation. Conditional simulation maintained global accuracy and spatial variability. Conditional simulation also provided a measure of spatial uncertainty useful to assess how LAI variability affects process models, and to evaluate how spatial variability influences the upscaling from Landsat ETM+ (25-30 m) to MODIS (250-1000 m) spatial resolutions. Our main conclusion is that the selection of the optimal mapping technique depends on user requirements, because not all the desired map characteristics can be achieved simultaneously