Three methods for estimating subpixel cover fractions in coarse resolution imagery
challenge of estimating sub-pixel proportions of di erent land cover types. This problem is
di cult because of the variety and variability of vegetation within individual pixels. This
thesis describes and compares two existing algorithms for estimating sub-pixel fractional
land cover and introduces a new algorithm that estimates sub-pixel cover fractions more
accurately than those currently used. The two existing algorithms are the linear spectral
mixture model, which has previously been applied to coarse-resolution imagery, and arti-
cial neural networks, which have been very successful across a wide range of tasks. The
paper introduces a new regression tree algorithm that directly estimates land cover fractions.
Our implementations of these methods enforce the requirement that the predicted
sub-pixel fractions sum to 1. In addition, we employed internal cross-validation methods to
calibrate the design parameters of the algorithms to optimize their performance separately
on each data set. The methods were tested on real and simulated data from three di erent
studies. The results show that the linear mixture model is signi cantly less accurate in terms of mean squared error compared to the non-linear neural network and regression tree
methods. In addition, we applied an ensemble learning method, bagging, to construct multiple
classi ers and combine their predictions by voting. The experiments demonstrate that
the regression tree algorithm combined with bagging gives the most accurate predictions.
This method proved efficient and easy to implement.
Advisor:Dietterich, Thomas G.
School:Oregon State University
School Location:USA - Oregon
Source Type:Master's Thesis
Keywords:vegetation mapping remote sensing mathematical models image processing
Date of Publication:05/28/2003