|Abstract or Summary
- Spatial Supervised Learning seeks to learn how to assign a label to each pixel in a spatial grid
such as the pixels of remote-sensed images. The standard approach is to treat each grid cell
separately and to use only the measured features of the grid cell to determine the assigned label.
However, spatial data usually exhibits spatial patterns so that labels of nearby grid cells are
correlated. It should be possible to learn this correlation and exploit it to improve the accuracy
of the predicted labels. This project studies simple recurrent sliding window classifiers for this
task. Recurrent sliding window classifiers feed their predicted output labels back as inputs when
making predictions for adjacent grid cells. The project shows how to generalize sequential sliding
window classifiers to spatial data and also studies the effectiveness of three ensemble methods:
(a) voting among sliding windows that approach a given grid cell from multiple directions, (b)
bagging, and (c) boosting. These methods are applied to the C4.5 decision tree algorithm and
to the naive Bayes algorithm. The results show that for naive Bayes, an increase in the input
and output context results in a significant improvement over the no-context approach, and this
result is improved with boosting, while bagging has little effect. For C4.5, using context delivers
a 36.5% reduction in error, and when combined with boosting the reduction in error is 56.1%.
Bagging also resulted in improvement for C4.5 making the greatest reduction in error, 46.3%,
when only output context was included as additional features.