CGAP: A Hybrid Contrastive and Graph-based Active Learning Pipeline to Detect Water and Sediment in Multispectral Images

Fri, 07 Jun 2024 00:00:00 +0000

The contrastive graph-based active learning pipeline (CGAP) combines a learned feature embedding with graph Laplace learning to classify land, surface water, and near-water sediment in Landsat imagery. Custom contrastive augmentations make the features robust to geometric transformations, changes in spatial resolution, and light cloud cover while reducing the dimension used for graph construction.

CGAP includes a basic version trained from the labeled RiverPIXELS dataset and an adaptive version that uses active learning to guide a human-in-the-loop labeling process. A compact representative set is joined with each test image to form a graph, allowing label propagation without constructing one prohibitively large graph over every training pixel.

The experiments show stronger boundary and overall accuracy than the preceding GAP, support-vector-machine, random-forest, and DeepWaterMap baselines, using orders of magnitude less labeled data than a CNN–U-Net pipeline. The accompanying GraphRiverClassifier connects the trained model to Google Earth Engine and Colab for rapid analysis of arbitrary Landsat scenes.

Batch Active Learning for Multispectral and Hyperspectral Image Segmentation Using Similarity Graphs

Sat, 01 Jun 2024 00:00:00 +0000

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.

Multispectral Imaging | Bohan Chen's Personal Webpage

CGAP: A Hybrid Contrastive and Graph-based Active Learning Pipeline to Detect Water and Sediment in Multispectral Images

Batch Active Learning for Multispectral and Hyperspectral Image Segmentation Using Similarity Graphs