|
Tall Shrub Abundance on the North Slope of Alaska from MISR, 2000-2010
Mark
James
Chopping, Montclair State University, chopping@pegasus.montclair.edu
(Presenter)
Rocio
Duchesne, Montclair State University, duchesneonr1@mail.montclair.edu
Zhuosen
Wang, University of Massachusetts Boston, wangzhs@bu.edu
Crystal
Schaaf, University of Massachusetts Boston, crystal.schaaf@umb.edu
Ken
Tape, U. Alaska, Fairbanks, fnkdt@uaf.edu
Tian
Yao, Montclair State University, yaot@mail.montclair.edu
We present maps of tall shrub fractional cover for the North Slope of Alaska constructed using a machine learning algorithm that exploits the surface information provided by the NASA Multiangle Imaging Spectro-Radiometer (MISR). Red band MISR BRFs on a 250 m grid were used to invert the RossThick-LiSparse_Reciprocal BRDF model via the AMBRALS algorithm to obtain isotropic, volumetric, and geometric kernel weights, albedos, model-fitting RMSE, and weights of determination.
The bootstrap forest machine learning algorithm generates N re-sampling data sets (with replacement) from the training dataset using a bootstrapping method. Each resampling data set is used to grow one decision tree in the forest. At each node in the tree, a group of independent variables are randomly selected to find the best split. The final estimate is the average of the predicted values from each tree. Observations not used in constructing the trees are used in validation.
A reference database of 444 locations was developed using field measurements on twelve 250 x 250 m plots to calibrate mean shrub cover and height estimates from the CANAPI algorithm, supplied with panchromatic imagery. The 444 observations were randomly divided into two datasets for training and validation (N=215, 229). The algorithm used a set of 19 independent variables: isotropic, volumetric, and geometric scattering kernel weights, ratios and interaction terms; white and black sky albedos; and blue, green, red, and NIR nadir camera BRFs. These were used to grow a forest of 9 decision trees.
The R2 values for the training and validation datasets were 0.89 and 0.46, respectively. The model was applied with a large volume of MISR data and the resulting fractional cover estimates were combined into maps using a compositing algorithm that flags results affected by clouds, surface water, burnt areas, extreme outliers, and topographic shading. In a second phase cloud shadows will be screened using a cloud mask.
Presentation:
2013_Poster_Chopping_24_100.pdf (4878k)
Presentation Type: Poster
Session: Poster Session 1-A
(Tue 11:00 AM)
Associated Project(s):
Poster Location ID: 24
|