Weixing Liu

Accurate and efficient cloud detection is critical for maintaining the usability of optical remote sensing imagery, particularly in large-scale Earth observation systems. In this study, we propose FEMNet, a lightweight dual-branch network that combines state space modeling with convolutional encoding for multi-class cloud segmentation. The Mamba-based encoder captures long-range semantic dependencies with linear complexity, while a parallel CNN path preserves spatial detail. To address the semantic inconsistency across feature hierarchies and limited context perception in decoding, we introduce the following two targeted modules: a cross-stage semantic enhancement (CSSE) block that adaptively aligns low- and high-level features, and a multi-scale context aggregation (MSCA) block that integrates contextual cues at multiple resolutions. Extensive experiments on five benchmark datasets demonstrate that FEMNet achieves state-of-the-art performance across both binary and multi-class settings, while requiring only 4.4M parameters and 1.3G multiply–accumulate operations. These results highlight FEMNet’s suitability for resource-efficient deployment in real-world remote sensing applications.

Domain adaptive object detection methods for remote sensing images typically rely on large-scale target domain data and multi-epoch offline adaption training. However, the wide variation in remote sensing conditions makes it difficult to gather sufficient data for every potential target domain, especially for unexpected domains. To address this challenge, we propose Learnable Online Style Adaptation (LOSA), a method that enables the source model to adapt effectively to new target domain styles with low test-time latency. Specifically, LOSA captures target domain styles using shallow feature channel statistics and predicts style shifts based on channel dependencies to recalibrate target features. Through a coarse-to-fine alignment loss between online target features and pre-computed source domain statistics, LOSA autonomously learns domain-specific style adaptation strategies. By adopting a dynamically optimized high learning rate, LOSA only requires a small number of samples for test-time training, making it suitable for real-time applications. Experimental results across various scenarios, including normal-to-corrupted, cross-band, and sim-to-real adaptation, demonstrate that the proposed method significantly improves cross-domain object detection performance. Moreover, our method can be well generalized to cross-domain image classification tasks.

Existing cloud detection methods often rely on deep neural networks, leading to excessive computational overhead. To address this, we propose a shallow convolutional neural network (CNN)-Transformer hybrid architecture that limits the maximum downsampling rate to 8×. This design preserves local details while effectively capturing global context through a lightweight Transformer branch. To enhance adaptability across diverse cloud scenes, we introduce two novel statistics-driven modules: statistics-adaptive convolution (SAC) and statistical mixing augmentation (SMA). SAC dynamically generates convolutional kernels based on input feature statistics, enabling adaptive feature extraction for varying cloud patterns. SMA improves model generalization by interpolating channel-wise statistics across training samples, increasing feature diversity. Experiments on four datasets show that the proposed method achieves state-of-the-art performance with 732K parameters and 1G multiply-accumulate operations. Code will be available at https://weix-liu.github.io/ for further research.

There are large nonlinear radiance differences between optical and synthetic aperture radar (SAR) from different imaging sensors. When combined with rotational and scale variations, achieving equivariant matching becomes highly challenging. The existing single network models face two primary limitations, insufficient learning of modal invariant features and insufficient characterization of rotation and scale. After task decomposition, we propose a novel rotation and scale equivariant network (RSENet) for optical-SAR image matching, comprising two modules, SAR-to-optical image translation module and equivariant descriptors (ED). First, we propose an image translation network to obtain more beneficial pseudo-optical images for matching. Second, we achieve high-precision matching via equivariant descriptors designed through scale angle estimation (SAE) and descriptor colearning. The network initially corrects image blocks explicitly via scale and angle loss functions that are based on periodic encoding (PE) learning. Subsequently, it employs an attention module based on cross-scale perception and feature channel excitation to extract hierarchical information from the corrected image blocks and to enhance the local feature representation of the descriptors. We validate the effectiveness of RSENet by conducting numerous experiments under combined transformations of various angles and scales. Our proposed method achieves an approximately 28% improvement in the matching success rate compared with existing methods.

We present REMM, a rotation-equivariant framework for end-to-end multimodal image matching, which fully encodes rotational differences of descriptors in the whole matching pipeline. Previous learning-based methods mainly focus on extracting modal-invariant descriptors, while consistently ignoring the rotational invariance. In this paper, we demonstrate that our REMM is very useful for multimodal image matching, including multimodal feature learning module and cyclic shift module. We first learn modal-invariant features through the multimodal feature learning module. Then, we design the cyclic shift module to rotationally encode the descriptors, greatly improving the performance of rotation-equivariant matching, which makes them robust to any angle. To validate our method, we establish a comprehensive rotation and scale-matching benchmark for evaluating the anti-rotation performance of multimodal images, which contains a combination of multi-angle and multi-scale transformations from four publicly available datasets. Extensive experiments show that our method outperforms existing methods in benchmarking and generalizes well to independent datasets. Additionally, we conducted an in-depth analysis of the key components of the REMM to validate the improvements brought about by the cyclic shift module. Code and dataset at https://github.com/HanNieWHU/REMM.

Vehicle detection methods based on deep learning have achieved remarkable results on remote sensing images. However, the performance of the detector degrades when the test images are distinct from the training images. Domain adaptive vehicle detection is a promising approach to bridging the domain gap. Existing methods usually adopt fully shared networks, but ignore the problem that features from different domains may be incompatible within a single model. In this letter, we present a novel domain adaptive vehicle detection method based on patch-wise domain-specific channel recalibration (PDSCR). The PDSCR module routes the feature to the corresponding network branch and extracts the channel dependence using separate parameters. In this way, our method can explicitly capture domain-specific information for each domain. Furthermore, we propose a dynamic weighted prototype alignment (DWPA) to avoid the negative effects of false pseudo-labels, especially in the early stage of training. Experimental results of adaptation from our synthetic dataset to three real vehicle detection datasets demonstrate the effectiveness of our method. Code and our synthetic data will be available at https://weix-liu.github.io/.