In recent years, deep learning has become a dominant research trend in the field of remote sensing. However, due to significant domain discrepancies among datasets collected from various platforms, models trained on a single domain often struggle to generalize to other domains. In domain-incremental learning scenarios, such discrepancies often lead to catastrophic forgetting, hindering the practical deployment of deep learning models. To address this, we propose ER-PASS, an experience replay-based continual learning algorithm that incorporates a performance-aware submodular sampling strategy. ER-PASS balances adaptability across domains and retention of knowledge by combining the strengths of joint learning and experience replay, while maintaining practical efficiency in terms of training time and memory usage. We validated our method on two remote sensing applications—building segmentation and land use/land cover (LULC) classification—using UNet and DeepLabV3+. Experimental results show that ER-PASS consistently outperforms existing continual learning methods in average incremental accuracy (AIA) and backward transfer (BWT), ensuring generalization across domains and mitigating catastrophic forgetting. While these results were obtained under restricted conditions, limited to a sequence of domains from high to low resolution and two applications, they underscore the potential of ER-PASS as a practical and general-purpose solution for continual learning in remote sensing.

In recent years, deep learning has become a dominant research trend in the field of remote sensing. However, due to significant domain discrepancies among datasets collected from various platforms, models trained on a single domain often struggle to generalize to other domains. In domain-incremental learning scenarios, such discrepancies often lead to catastrophic forgetting, hindering the practical deployment of deep learning models. To address this challenge, we propose ER-PASS, an experience replay-based continual learning algorithm that incorporates a performance-aware submodular sampling strategy. ER-PASS effectively balances adaptability across domains and retention of knowledge by integrating the strengths of joint learning and experience replay, while maintaining practical efficiency in terms of training time and memory usage. Additionally, by leveraging a performance-aware sample selection strategy that integrates submodular gain and task-specific evaluation scores, ER-PASS enables robust and stable learning. Experimental results on two distinct applications—building segmentation and land use/land cover (LULC) classification—demonstrate that ER-PASS outperforms existing continual learning methods in mitigating forgetting and generalizing across diverse application scenarios. Moreover, experiments conducted on both UNet and DeepLabV3+ validate the model-agnostic nature of ER-PASS, underscoring its potential as a practical and general-purpose solution for continual learning in remote sensing.

Taehong Kwak, Hyeonseong Choi, Yeseok Lee, Yongil Kim. Korean J. Remote Sens. 2025;41:1151-60. https://doi.org/10.7780/kjrs.2025.41.6.16

Change detection (CD) from remote sensing imagery is an essential task for monitoring various Earth dynamics. Recently, self-supervised learning (SSL) has emerged as a promising paradigm to alleviate the reliance on large-scale annotated datasets. Previous SSL methods for CD suffer from spatial misalignment caused by multi-view augmentation, inherent trade-offs among SSL paradigms, and the lack of explicit change semantics in representation. To overcome these challenges, we propose Difference-Paired Self-Supervised learning for Change Detection (DiPaSSCD), a novel pretraining framework that explicitly integrates change-oriented signals into SSL. The method consists of three key components: (i) Object-Based Local (OBL) Augmentation, which simulates realistic local variations while preserving spatial alignment; (ii) Mask-guided SSL Integration, which jointly optimizes global contrastive and dense prediction objectives to mitigate the limitations of each paradigm; and (iii) Difference-Paired Contrastive Learning (DiPaCL), which directly learns embeddings from paired inputs containing explicit differences, yielding inherently change-related feature representations. Extensive experiments on LEVIR-CD, S2Looking, and DSIFN datasets under both full and label-limited conditions demonstrate that DiPaSSCD consistently outperforms supervised and SSL-based baselines. Qualitative analyses further confirm its ability to suppress pseudo-changes induced by illumination variations while capturing subtle, fine-scale land cover changes. These results highlight the potential of DiPaSSCD as a versatile and architecture-agnostic pretraining paradigm for robust CD across diverse scenarios.

Hyeonseong Choi, Yeseok Lee, Taehong Kwak, Yongil Kim. Korean J. Remote Sens. 2025;41:995-1010. https://doi.org/10.7780/kjrs.2025.41.6.6