DG-FDD: Mitigating Catastrophic Forgetting in Remote Sensing Change Detection

What Changed

Remote sensing change detection (RSCD) is a critical task in various applications, from environmental monitoring to urban development. However, deploying deep learning models for RSCD in real-world scenarios presents a significant challenge: catastrophic forgetting. This phenomenon occurs when a model, incrementally adapted to a new data domain (e.g., different sensors, atmospheric conditions, or geographical regions), loses its ability to accurately perform tasks on previously learned domains. Existing domain-incremental learning (DIL) methods have largely focused on preserving general image-level representations, often overlooking the crucial bitemporal discrepancy cues that are fundamental for robust change detection under shifting domain conditions.

Researchers have introduced DG-FDD, a novel domain-incremental change detection framework designed to overcome these limitations. DG-FDD integrates two key strategies: Difference-Guided Adaptation and Frequency-Decoupled Distillation. This framework specifically targets the preservation of change-relevant information and facilitates stable knowledge transfer across domains without the need for historical data, thereby mitigating catastrophic forgetting and ensuring consistent performance as models encounter new environments.

Technical Details

The DG-FDD framework is built upon two interconnected components: the Difference-Guided Dynamic Adapter (DGDA) and the Frequency-Decoupled Knowledge Distillation strategy with Cross-domain Synthesis (FDKD-CS).

The Difference-Guided Dynamic Adapter (DGDA) is designed to enhance change-aware feature adaptation. In RSCD, the core task is to identify differences between two temporally distinct images of the same geographical area. These "bitemporal discrepancy cues" are paramount. Traditional DIL methods often treat each image independently or process them in ways that dilute these critical differences. The DGDA explicitly models these bitemporal feature discrepancies. By focusing on the differences between features extracted from the two input images, the adapter promotes the learning of representations that are highly sensitive to actual changes while simultaneously reducing interference from domain-specific variations that are irrelevant to the change detection task. This targeted adaptation ensures that the model's capacity for identifying changes remains robust even as the underlying data domain shifts. The dynamic nature of the adapter allows it to adjust its parameters based on the characteristics of the incoming domain, further enhancing its adaptability.

The Frequency-Decoupled Knowledge Distillation strategy with Cross-domain Synthesis (FDKD-CS) addresses the challenge of stable knowledge transfer without access to historical data, a common constraint in real-world incremental learning scenarios. Catastrophic forgetting often arises because new learning overwrites knowledge acquired from previous domains. Knowledge distillation is a technique where a smaller "student" model learns from a larger "teacher" model. In DIL, the "teacher" often represents the model trained on previous domains. FDKD-CS innovates by operating in the frequency domain. It separates structural information, which is essential for change detection, from domain style, which varies significantly across different datasets. By decoupling these two aspects, the strategy enables the stable transfer of structural knowledge from the teacher model (representing historical knowledge) to the student model (adapting to the new domain) without being confounded by stylistic differences. The Cross-domain Synthesis aspect further enhances this by generating synthetic data that bridges the gap between domains, facilitating more effective knowledge transfer even in the absence of actual historical data. This dual approach ensures that the model retains its core change detection capabilities while effectively learning from new, unseen domains.

Benchmark Analysis

Extensive experiments were conducted on three public high-resolution RSCD datasets, utilizing both two-domain and three-domain incremental protocols to evaluate DG-FDD's effectiveness in mitigating catastrophic forgetting. The framework's performance was benchmarked against independently trained single-task models, focusing on mean relative changes in F1 score and Intersection over Union (IoU), two standard metrics for change detection accuracy.

Across six two-domain incremental sequences, DG-FDD demonstrated remarkable stability, recording mean relative changes in F1 and IoU of only -0.23% and -0.45%, respectively. This indicates a minimal degradation in performance when adapting to a second new domain. For the more challenging three-domain incremental sequences, evaluated across three distinct protocols, DG-FDD maintained its robust performance. The mean relative changes observed were -0.69% for F1 and -1.31% for IoU. These results highlight DG-FDD's ability to preserve historical knowledge while effectively adapting to multiple successive new domains.

The consistently low negative relative changes across both two- and three-domain scenarios underscore a favorable stability-plasticity balance achieved by DG-FDD. This balance is critical for continual cross-domain change detection, ensuring that the model remains plastic enough to learn new domain characteristics without catastrophically forgetting previously acquired knowledge.

Developer Implications

For developers working on remote sensing applications, DG-FDD offers a significant advancement in building more robust and adaptable change detection systems. The ability to incrementally adapt models to new domains without substantial performance degradation directly translates to more efficient model deployment and maintenance cycles. Instead of retraining models from scratch or managing complex historical data archives for each new domain encountered, developers can leverage DG-FDD to continually update their models. This reduces computational costs, accelerates deployment, and allows for more dynamic adaptation to evolving data sources and environmental conditions.

The framework's focus on bitemporal discrepancy cues ensures that the core change detection capability remains sharp, even as the underlying imagery characteristics shift. This is particularly valuable in real-world scenarios such as disaster response, urban expansion monitoring, or agricultural assessment, where data streams are diverse and constantly changing. Furthermore, the FDKD-CS component's ability to facilitate knowledge transfer without requiring historical data is a critical practical advantage, simplifying data management and privacy concerns often associated with retaining past datasets. Developers can deploy models that learn on the fly, maintaining high accuracy and relevance across a wide spectrum of operational environments.

Bottom Line

DG-FDD represents a substantial step forward in addressing catastrophic forgetting in remote sensing change detection. By explicitly modeling bitemporal discrepancies and employing a frequency-decoupled knowledge distillation strategy, the framework achieves a critical balance between retaining historical knowledge and adapting to new domains. The demonstrated minimal performance degradation across incremental learning scenarios provides a robust solution for developers seeking to deploy highly adaptable and stable RSCD models in dynamic real-world applications.

Pneumetron