A Survey of Sample-Efficient Deep Learning for Change Detection in Remote Sensing: Tasks, Strategies, and Challenges

TL;DR

This survey addresses the data bottleneck in deep learning-based remote sensing change detection by organizing CD tasks (Binary CD, Multi-class/Semantic CD, Time-Series CD) and detailing sample-efficient strategies across four supervision regimes: semi-, weakly-, self-, and unsupervised CD. It surveys a spectrum of techniques, from pseudo-labeling, consistency regularization, and graph-based learning to contrastive learning, masked image modeling, generative representations, and the emerging role of visual foundation models and multi-temporal foundation models. Key findings show that semi-supervised CD often achieves strong performance with limited labeled data, while unsupervised and zero-shot approaches still lag behind supervised methods on high-resolution data. The paper highlights challenges in domain transfer, spatiotemporal complexity, and unseen changes, and it points to future directions such as multi-temporal foundation models, few-shot/zero-shot CD, and interactive, user-guided CD to enable broader, real-world deployment in diverse RS contexts.

Abstract

In the last decade, the rapid development of deep learning (DL) has made it possible to perform automatic, accurate, and robust Change Detection (CD) on large volumes of Remote Sensing Images (RSIs). However, despite advances in CD methods, their practical application in real-world contexts remains limited due to the diverse input data and the applicational context. For example, the collected RSIs can be time-series observations, and more informative results are required to indicate the time of change or the specific change category. Moreover, training a Deep Neural Network (DNN) requires a massive amount of training samples, whereas in many cases these samples are difficult to collect. To address these challenges, various specific CD methods have been developed considering different application scenarios and training resources. Additionally, recent advancements in image generation, self-supervision, and visual foundation models (VFMs) have opened up new approaches to address the 'data-hungry' issue of DL-based CD. The development of these methods in broader application scenarios requires further investigation and discussion. Therefore, this article summarizes the literature methods for different CD tasks and the available strategies and techniques to train and deploy DL-based CD methods in sample-limited scenarios. We expect that this survey can provide new insights and inspiration for researchers in this field to develop more effective CD methods that can be applied in a wider range of contexts.

Figures (8)

• Figure 1: The number of literature publications associated with different CD topics over the past 10 years. Solid lines present different CD tasks, while the dashed lines indicate different supervision strategies.
• Figure 2: A comparison between (a) BCD, (b) MCD/SCD, and (c) TSCD. The color regions in $Y_1, Y_2, Y_3$ and $Y_c^{1 \rightarrow t}$ indicate the pre-defined LCLU/change categories.
• Figure 3: Comparison of annotation and data volume in different CD learning paradigms.
• Figure 4: Consistency regularization for WSCD bandara2022revisiting. Random perturbations are applied to the change representations, and a consistency loss is calculated between the origninal and perturbed CD results to improve the robustness of CD models.
• Figure 5: Refining CAM for SMCD within a teacher-student framework lu2024weakly. A CAM is obtained with image-level supervision (class loss), and is refined through knowledge distillation.