Revisiting Consistency Regularization for Semi-supervised Change Detection in Remote Sensing Images (2204.08454v3)

Abstract: Remote-sensing (RS) Change Detection (CD) aims to detect "changes of interest" from co-registered bi-temporal images. The performance of existing deep supervised CD methods is attributed to the large amounts of annotated data used to train the networks. However, annotating large amounts of remote sensing images is labor-intensive and expensive, particularly with bi-temporal images, as it requires pixel-wise comparisons by a human expert. On the other hand, we often have access to unlimited unlabeled multi-temporal RS imagery thanks to ever-increasing earth observation programs. In this paper, we propose a simple yet effective way to leverage the information from unlabeled bi-temporal images to improve the performance of CD approaches. More specifically, we propose a semi-supervised CD model in which we formulate an unsupervised CD loss in addition to the supervised Cross-Entropy (CE) loss by constraining the output change probability map of a given unlabeled bi-temporal image pair to be consistent under the small random perturbations applied on the deep feature difference map that is obtained by subtracting their latent feature representations. Experiments conducted on two publicly available CD datasets show that the proposed semi-supervised CD method can reach closer to the performance of supervised CD even with access to as little as 10% of the annotated training data. Code available at https://github.com/wgcban/SemiCD

Overview of "Revisiting Consistency Regularization for Semi-supervised Change Detection in Remote Sensing Images"

This paper addresses the challenge of semi-supervised change detection (CD) in remote sensing imagery, focusing on leveraging unlabeled data effectively to improve the performance of CD models. Given the burgeoning volume of unlabeled remote sensing data from global earth observation missions and the substantial costs associated with obtaining detailed annotations for this data, the authors propose a novel semi-supervised framework that utilizes consistency regularization. This approach constrains the predicted change probability maps to remain stable under small random perturbations applied to the deep feature difference map, aligning well with the challenges posed by bi-temporal imaging.

Key Contributions

1. Semi-supervised CD Framework: The authors propose a semi-supervised learning paradigm grounded in consistency regularization specifically tailored for change detection. By doing this, they aim to utilize substantial volumes of available unlabeled remote sensing images, substantially reducing the dependence on large, annotated datasets.
2. Reformulation of the Cluster Assumption: The paper redefines the traditional cluster assumption in semi-supervised learning for the CD task domain, establishing that it is more applicable in the deep feature difference domain than in the direct bi-temporal image domain. This insight allows the use of perturbation-based regularization effectively.
3. Simple and Innovative Perturbation Techniques: The research details the application of varied small perturbations on feature maps to enforce consistent transfer function behavior in supervised learning scenarios. This ensures the model learns robust feature representations less sensitive to noise and thus generalizes better across domains.
4. Comprehensive Evaluation: Experiments illustrate the strong performance of the proposed framework on two public datasets, LEVIR and WHU, demonstrating proximity in performance to fully supervised models even using significantly less labeled data (as little as 10%). Cross-dataset generalizability assessments further underscore its robustness.

Theoretical and Practical Implications

The paper's findings have several notable implications. Practically, the technique reduces the cost and time constraints of data annotation, enabling more efficient use of satellite imagery for various applications such as disaster management and resource monitoring. Theoretically, the paper advances the understanding of consistency regularization in the change detection domain, advocating the deep feature difference as a more suitable space for maintaining cluster assumptions, thereby broadening the scope of semi-supervised learning in visual tasks.

Future Perspectives

Building on the contributions of this work, future research could explore the applicability of this framework to other types of visual data beyond remote sensing. Moreover, additional work could focus on refining perturbation methods or integrating other types of unsupervised learning techniques like generative models to perhaps further enhance the model's ability to distinguish changes of interest from background noise or irrelevant changes.

Ultimately, this research provides a compelling leap in the field of semi-supervised remote sensing change detection, setting the stage for more scalable and efficient earth observation analytics, which aligns with pressing global needs for enhanced environmental monitoring capabilities.