Jianjun Sha

Deep learning has shown strong performance in hyperspectral images change detection (HSIs-CD), but it typically requires a large amount of labeled data. Acquiring such labeled data is often expensive, time-consuming, and infeasible in many application scenarios. Few-shot learning (FSL) addresses this limitation by training a model on a large number of training tasks and then applying it to test tasks with only a few labeled samples. However, in HSIs-CD, variations in imaging conditions lead to significant distribution gaps across datasets. Consequently, FSL models trained on the training tasks typically generalize poorly to test tasks and may fail to extract meaningful features. To address this, we propose FSDA-Former, a few-shot domain adaptation transformer for HSIs-CD. First, it employs a generative pretraining strategy based on a masked autoencoder (FSDA-MAE) to learn domain-agnostic representations and improve feature extraction across diverse tasks. Second, an FSDA Transformer layer (FSDA-TL) is introduced to refine the extracted features in a meta-learning manner, thereby learning class-discriminative representations and constructing a task-specific metric space. Third, a two-stage domain adaptation strategy (FSDA-TDA) is developed to further adapt FSDA-Former to the feature domain of the test task. FSDA-TDA further enhances the generalization ability of FSDA-Former under distribution shifts between training and test tasks. Experimental results on three real-world HSIs-CD datasets demonstrate that FSDA-Former consistently outperforms existing few-shot baselines. It achieves up to 3.60% improvement in Kappa and 1.58% in overall accuracy. These results demonstrate the effectiveness and superiority of FSDA-Former in HSIs-CD, even when only a few labeled examples are available.

For the task of 3D building reconstruction from stereo satellite imagery, existing methods often treat semantic segmentation, contour regularization, and disparity estimation as independent or loosely coupled tasks, failing to fully exploit their geometric correlations. To address this, this letter proposes the Semantic-Contour-Disparity Triadic Collaborative Network (SCD-Net). This network establishes a deeply coupled computational framework that extracts multi-scale features through a shared encoder, subsequently routing them to dedicated branches. Specifically, we design a Multi-Scale Edge Aggregation (MEA) module that significantly enhances the refined representation of building edges, and also propose a Semantic-Contour-Disparity Up-sampling (SCDU) module, which leverages dual geometric priors from semantic categories and building contours to guide disparity optimization, effectively resolving the issue of blurred building boundaries in disparity map. Furthermore, a novel building contour loss function is introduced, strengthening multi-task constraints by quantifying the geometric consistency between disparity edges and semantic contours. Experiments on the US3D and WHU datasets demonstrate that the SCD-Net achieves state-of-the-art performance on key metrics for stereo matching and building semantic segmentation.

Hyperspectral image change detection (HSIs-CD) often faces the challenge of limited sample sizes, and labeling data is both time-consuming and labor-intensive. Foundation models leverage extensive unlabeled data for self-supervised generative pre-training, allowing the model to learn rich data representations. However, few models have been specifically designed for HSIs, and existing methods often rely on pre-training datasets that are limited to data from a small number of satellite sensors. This limitation affects generalization, especially when there are significant differences between data from different sensors. Moreover, the difference map (DMP) of bi-temporal HSIs is often used as input to the networks. While the DMP-based approach reduces FLOPs, it may lead to information loss compared to dual-branch networks. In this letter, we propose a mini-patch-based generative pre-trained spectral-spatial transformer (GPSST) for semi-supervised HSIs-CD. We begin by collecting public HSIs datasets and dividing them into thousands of patches. Each patch is then split into spectral-spatial tokens, with a portion of these tokens masked and used as input for the GPSST. We then design a spectral-spatial masked autoencoder as the backbone of GPSST for self-supervised generative learning. Finally, we fine-tune the GPSST encoder using a small number of labeled patches and design a principal component analysis (PCA) branch to compensate for the information loss caused by the DMP. Our experiments demonstrate that GPSST outperforms existing methods, achieving superior accuracy in HSIs-CD.

Point cloud registration is a fundamental challenge in the field of 3D vision, particularly when dealing with scans that have limited overlapping features. In this paper, a local-to-global point cloud registration algorithm based on multi-keypoints joint matching is proposed. The algorithm incorporates neighborhood keypoints distance constraint and multi-keypoints joint matching scheme to establish robust and weighted correspondences. Furthermore, a fast segmentation and registration method is designed for overlapping regions, utilizing the weighted correspondences. The proposal algorithm achieves the expected results and is superior to mainstream traditional registration algorithms in terms of efficiency and accuracy.

Deep learning plays an increasingly important role for change detection (CD) in aerial optics remote sensing images (AORSIs). Convolutional neural networks with superior performance usually require an increased number of convolutional layers. However, the expanded size of convolutional layers significantly increases memory consumption, posing a challenge for practical applications especially in scenarios with low computing power. A tiny-Siamese network with bilateral channels (TSNBC) is proposed for AORSIs-CD. We first design a bilateral feature extraction module with few parameters to extract both detailed and global features. Furthermore, a difference enhancement module is designed to reinforce the distinctions between detailed and global features of bi-temporal AORSIs. The experimental results on the public dataset show that the TSNBC proposed in this paper achieves the best performance across many measured indices.

Satellite-Drone Image Cross-View Geolocalization has wide applications. Due to the pronounced variations in the visual features of 3D objects under different angles, Satellite-Drone cross-view image geolocalization remains an unresolved challenge. The key to successful cross-view geolocalization lies in extracting crucial spatial structure information across different scales in the image. Recent studies improve image matching accuracy by introducing an attention mechanism to establish global associations among local features. However, existing methods primarily focus on using single-scale features and employ a single-channel attention mechanism to correlate local convolutional features from different locations. This approach inadequately explores and utilizes multi-scale spatial structure information within the image, particularly lacking in the extraction and utilization of locally valuable information. In this paper, we propose a cross-view image geolocalization method based on multi-scale information and a dual-channel attention mechanism. The multi-scale information includes features extracted from different scales using various convolutional slices, and it extensively utilizes shallow network features. The dual-channel attention mechanism, through successive local and global feature associations, effectively learns depth discriminative features across different scales. Experimental results were conducted using existing satellite and drone image datasets, with additional validation performed on an independent self-made dataset. The findings indicate that our approach exhibits superior performance compared to existing methods. The methodology presented in this paper exhibits enhanced capabilities, especially in the exploitation of multi-scale spatial structure information and the extraction of locally valuable information.