1 October 2022 in Journal of Applied Remote Sensing
Remote sensing scene classification has received more and more attention as important fundamental research in recent years. However, the redundant background information and complex spatial scale variability of remote sensing scene images make the existing convolutional neural network models, which mainly concentrate on global features, perform poorly. To effectively alleviate these problems, we proposed an MSRes-SplitNet model based on multiscale features and attention mechanisms for remote sensing scene image classification. First, MSRes blocks are constructed for the extraction of multi-scale features. Then, the multi-channel local features are fused by the Split-Attention block. Finally, the global and local feature information is aggregated by convolution, thus obtaining multi-scale features while alleviating the small-sample learning problem. Experiments are conducted on three publicly available datasets and compared with other state-of-the-art methods, showing that the proposed method MSRes-SplitNet has better performance while effectively reducing a large number of parameters.
19 April 2024 in IET Image Processing
Remote sensing scene images contain numerous feature targets with unrelated semantic information, so how to extract to the local key information and semantic features of the image becomes the key to achieving accurate classification. Existing Convolutional Neural Networks (CNNs) mostly concentrate on the global representation of an image and lose the shallow features. To overcome these issues, this paper proposes LPNet for remote sensing scene image classification. First, LPNet employs LKConv to extract the semantic features in the image, while using standard convolution to extract local key information in the image. Additionally, the LPNet applies a shortcut residual concatenation branch to reuse features. Then, parameter fusion combines parameters from previous branches, improving the capacity of the model to obtain a more comprehensive and rich feature representation of the image. Finally, considering the relationship between the classification ability of the model and the depth of feature extraction, the Feature Mixture (FM) Block is used to deepen the model for feature extraction. Comparative experiments on four publicly available datasets show that LPNet provides comparable results to other state‐of‐the‐art methods. The effectiveness of LPNet is further demonstrated by visualizing the effective receptive fields (ERFs).
28 January 2026 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Remote sensing scene classification (RSSC) aims to assign semantic labels to remote sensing images (RSIs). However, existing convolutional neural network (CNN)-based methods often suffer from limited feature representation due to their bias toward global semantics while neglecting local textures, their sensitivity to illumination variations, and their reliance on cross-entropy loss, which fails to explicitly model sample- wise relationships in the feature space. To address these challenges, we propose LFNet, a frequency–texture fusion network equipped with a distance correlation loss for geometry-consistent feature optimization. Specifically, a Local Binary Pattern Frequency Convolution Module (LFCM) is introduced, which consists of two parallel branches: (1) an OctBranch that captures complementary high- and low-frequency semantic information via octave convolution and dual attention refinement, and (2) an LBPBranch that extracts illumination-robust texture cues in the feature domain. In addition, a Distance Correlation Loss (DC-Loss) is designed to explicitly regularize pairwise Euclidean distances among samples, thereby enhancing intra-class compactness and inter-class separability in the embedding space. Extensive experiments on four public RSSC datasets, including UCM, AID, NWPU, and NATG, demonstrate that LFNet achieves competitive or superior classification performance, with up to 1.2% improvement in overall accuracy and 12.4% reduction in FLOPs compared with the baseline ResNet50, while maintaining strong robustness under varying illumination conditions.
6 April 2026 in IEEE Geoscience and Remote Sensing Letters
Cross-domain scene classification (CDSC) aims to classify remote sensing images by transferring knowledge from a labeled source domain to an unlabeled target domain. Existing single-scale adaptation methods often fail to capture fine-grained local features, are sensitive to illumination changes, rely only on global distribution alignment, and suffer from error propagation. To address these, we propose KFTNet, a knowledge-guided multi-scale frequency-texture alignment network. KFTNet improves adaptation through two core components: the multi-scale frequency texture alignment module (MFTAM) and the knowledge-guided module (KGM). MFTAM utilizes frequency-texture convolution and fusion attention to integrate high-frequency details, global structures, and illumination-invariant textures. It then performs conditional distribution alignment using multi-kernel maximum mean discrepancy, complementing global marginal distribution alignment via maximum mean discrepancy to enhance cross-domain feature adaptability. Concurrently, KGM mitigates error accumulation by incorporating high-confidence target samples as prior knowledge during early training. Experimental results confirm that KFTNet significantly outperforms state-of-the-art domain adaptation methods on CDSC tasks.
15 April 2013 in International Journal of Digital Content Technology and its Applications