Abstract
Graph learning, when used as a semi-supervised learning (SSL) method, performs well for classification tasks with a low label rate. We provide a graph-based batch active learning pipeline for pixel/patch neighborhood multi- or hyperspectral image segmentation. Our batch active learning approach selects a collection of unlabeled pixels that satisfy a graph local maximum constraint for the active learning acquisition function that determines the relative importance of each pixel to the classification. This work builds on recent advances in the design of novel active learning acquisition functions (e.g., the Model Change approach in arXiv:2110.07739) while adding important further developments including patch-neighborhood image analysis and batch active learning methods to further increase the accuracy and greatly increase the computational efficiency of these methods. In addition to improvements in the accuracy, our approach can greatly reduce the number of labeled pixels needed to achieve the same level of the accuracy based on randomly selected labeled pixels.
Type
Publication
Communications on Applied Mathematics and Computation
This paper develops a graph-based pipeline for label-efficient multispectral and
hyperspectral image segmentation. Pixels or local image patches are embedded as nodes of
a similarity graph, and graph Laplace learning propagates the small number of queried
labels across the image.
The batch acquisition rule selects unlabeled nodes that are local maxima of a model-change
score on the graph. This prevents a batch from concentrating on redundant nearby samples,
while patch-neighborhood features provide spatial context beyond individual spectra. The
method therefore reduces both human labeling effort and the number of expensive learning
cycles.
Experiments across remote-sensing datasets show that carefully selected batches reach a
target segmentation accuracy with far fewer labels than random sampling and retain the
effectiveness of sequential active learning at substantially lower computational cost.