Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Label-frugal satellite image change detection with generative virtual exemplar learning

Hichem Sahbi

TL;DR

This work tackles label-efficient satellite image change detection by introducing an interactive framework that queries an oracle on the most informative patch-pairs. It combines invertible graph convolutional networks with a bi-level optimization to design virtual exemplars that jointly optimize representativity, diversity, and ambiguity, enabling latent-space exemplar generation on a nonlinear data manifold. The approach yields stable exemplar synthesis and improved learning under frugal labeling, validated on the Jefferson dataset where it outperforms common display strategies and approaches the fully supervised upper bound. The proposed method offers a practical pathway to robust change detection in remote sensing when extensive labeled data are unavailable.

Abstract

Change detection is a major task in remote sensing which consists in finding all the occurrences of changes in multi-temporal satellite or aerial images. The success of existing methods, and particularly deep learning ones, is tributary to the availability of hand-labeled training data that capture the acquisition conditions and the subjectivity of the user (oracle). In this paper, we devise a novel change detection algorithm, based on active learning. The main contribution of our work resides in a new model that measures how important is each unlabeled sample, and provides an oracle with only the most critical samples (also referred to as virtual exemplars) for further labeling. These exemplars are generated, using an invertible graph convnet, as the optimum of an adversarial loss that (i) measures representativity, diversity and ambiguity of the data, and thereby (ii) challenges (the most) the current change detection criteria, leading to a better re-estimate of these criteria in the subsequent iterations of active learning. Extensive experiments show the positive impact of our label-efficient learning model against comparative methods.

Label-frugal satellite image change detection with generative virtual exemplar learning

TL;DR

This work tackles label-efficient satellite image change detection by introducing an interactive framework that queries an oracle on the most informative patch-pairs. It combines invertible graph convolutional networks with a bi-level optimization to design virtual exemplars that jointly optimize representativity, diversity, and ambiguity, enabling latent-space exemplar generation on a nonlinear data manifold. The approach yields stable exemplar synthesis and improved learning under frugal labeling, validated on the Jefferson dataset where it outperforms common display strategies and approaches the fully supervised upper bound. The proposed method offers a practical pathway to robust change detection in remote sensing when extensive labeled data are unavailable.

Abstract

Change detection is a major task in remote sensing which consists in finding all the occurrences of changes in multi-temporal satellite or aerial images. The success of existing methods, and particularly deep learning ones, is tributary to the availability of hand-labeled training data that capture the acquisition conditions and the subjectivity of the user (oracle). In this paper, we devise a novel change detection algorithm, based on active learning. The main contribution of our work resides in a new model that measures how important is each unlabeled sample, and provides an oracle with only the most critical samples (also referred to as virtual exemplars) for further labeling. These exemplars are generated, using an invertible graph convnet, as the optimum of an adversarial loss that (i) measures representativity, diversity and ambiguity of the data, and thereby (ii) challenges (the most) the current change detection criteria, leading to a better re-estimate of these criteria in the subsequent iterations of active learning. Extensive experiments show the positive impact of our label-efficient learning model against comparative methods.

Paper Structure

This paper contains 11 sections, 5 equations, 1 figure, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Proposed method
  3. A glimpse on interactive change detection
  4. Virtual exemplar design
  5. Graph convnets
  6. Invertible graph convnet design
  7. Regularization
  8. Experiments
  9. Ablation Study
  10. Comparison
  11. Conclusion

Figures (1)

  • Figure 1: This figure shows a comparison of different sampling strategies w.r.t. different iterations (Iter) and the underlying sampling rates in table \ref{['tab1']} (Samp). Here Uncer and Rand stand for uncertainty and random sampling respectively. Note that fully-supervised learning achieves an EER of $0.94 \%$. Related work stands for the method in refff33333; see again section \ref{['compare']} for more details.