Change detection in satellite images is an emerging research area as it has a wide range of applications in natural resource monitoring, geo-hazard detections, urban planning, etc. Identifying physical changes on the ground and avoiding spurious changes due to other reasons like co-registration issues, change in illumination conditions, sun angle, and presence of cloud and fog is a challenging task. This work proposes a multitask learning based change detection model where two parallel pipeline architectures predict change map and image difference. The proposed model takes two images and their difference as input and provides them to a Backbone Network (BN). The output of the BN is fed into the proposed Multi-Scale Attention Module (MSAM) for the effective identification of changes in multi-temporal and very high-resolution aerial images. In another parallel path, the output of the BN is down-sampled and passed to the proposed Deconvolution with a Sub-pixel Convolution Module (DSCM) to generate image difference. Two loss functions are utilized in two parallel paths to train the overall model in an end-to-end supervised setting. A comprehensive set of experiments have been carried out, and the results reveal that the proposed DRMNet model has achieved an F1 Score improvement of 1.66% in CDD, 1.61% in SYSU, and 0.14% in LEVIR-CD datasets. It achieved an F1 score of 86.11% for the BCDD dataset with the new test image. Our model implementation is available at https://github.com/chouhan-avinash/DRMNet.
Cloudy conditions affect the quality of captured data by optical satellites. Multimodal techniques rely on synthetic aperture radar (SAR) images to recover cloudy pixels in optical images. These techniques face challenges of noise, modality and temporal differences. In this work, we propose a Diffusion Guided Masked Reconstruction (DGMR) framework for multimodal cloud removal, which consists of a Masked Reconstruction Network (MRNet), Conditional Diffusion Guidance Model (CDGM) and Non-cloudy Difference Similarity (NDS) soft constraint. DGMR effectively extracts local-global relations and combines complementary information using MRNet with coupled feature fusion and decoupled masked reconstruction. CDGM guides the intermediate features of MRNet to reconstruct more refined, cloud-free images. NDS ensures that the reconstructed output is consistent with temporal changes. DGMR achieves state-of-the-art results on four widely used benchmarks of SEN12MS-CR, M3R-CR and SMILE-CR datasets. The code and trained models are available at https://github.com/chouhan-avinash/DGMR/.
In transfer learning, domain adaptation is one of the key research areas. For domain adaptation, domain shift is a known problem when the data distribution of the source domain, from which the training data is fetched, and the target domain, from which the test data is fetched, vary significantly. Aligning the source and target domains is a solution, but due to alignment, the intrinsic properties of the data may be altered. To address this issue of domain shift, we introduce a novel method, called “A Unified Deep-Domain Adaptation Framework: Advancing Feature Separability and Local Alignment” (DDASLA) that incorporates an attention mechanism into the ResNet18 model to improve its feature extraction capability. Apart from self-attention, a combined loss function consisting of angular loss, Local Maximum Mean Discrepancy (LMMD), and entropy minimization is used. Angular loss enhances feature discrimination through angular alignment, whereas LMMD equalizes local data distributions across domains, and entropy minimization refines the decision boundaries. A comprehensive experiment on the Office and remote sensing datasets shows that DDASLA outperforms several state-of-the-art methods. These findings show that DDASLA improves model generalization and robustness across domains, paving the way for future domain adaptation research.