Chengkun Zhang

Balancing long-range dependencies and efficiency is crucial for remote sensing segmentation. While State Space Models (SSMs) offer linear complexity, they often struggle with local details. To address this, we propose the Lightweight Attention-Mamba Network (LANet). It incorporates an Attention-Guided State Space Module (AG-SSM) to transform indiscriminate scanning into selective refinement of salient features, and a Difference-Aware Gated Fusion (DAGF) module for effective multi-scale integration. LANet achieves state-of-the-art performance on ISPRS and LoveDA datasets. Notably, it attains 51.24% mIoU on LoveDA with only 8.23 M parameters, significantly outperforming UNetFormer (11.72 M) and RS3Mamba (43.33 M) in efficiency.

In China, the advancement of the rural revitalisation strategy has increased the demand for precise spatial information on village buildings. However, the irregular architectural forms in the Hehuang Valley of Qinghai complicate this task, making traditional extraction methods inadequate. This study proposes an efficient building extraction method using an improved UNet3+ model. The approach includes a comprehensive optimisation scheme addressing the model's limitations in complex environments, consisting of “structural optimisation, mechanism enhancement, and training strategy improvement.” Structurally, a pruning strategy enhances computational efficiency while maintaining accuracy. Mechanically, the Convolutional Block Attention Module (CBAM) is introduced to improve feature identification through a dual attention mechanism. The training strategy integrates Focal Loss, MS‐SSIM Loss, and IoU Loss functions, along with early stopping and learning schedulers to boost efficiency and generalisation. On an aerial imagery dataset, this method reduces parameters from 30.18M to 7.71M and increases precision from 92.49% to 95.33%. The method was validated using a drone‐captured dataset of typical villages in the Hehuang Valley.

Change detection (CD) in high-resolution remote sensing images (RSIs) aims at locating and understanding surface change areas. Despite some models have been proposed to solve the intrinsic problems of CD in RSIs (e.g. scale variation and internal non-consistency), they resulted in less than ideal outcomes in specific scenes, such as objects with the same semantic concept but different spectrums and irrelevant change objects in the background. To this end, this paper proposes a progressive refinement network (PRNet) to explore changes in more complex scenes in a continual calibration way. First, we excavate focused interactive deep semantic information with a proposed semantic refinement (SR) module based on the vision transformer and graph representation, which understands more useful semantic relations in ground objects. Second, we design a self-refinement (Self-R) module based on the supervised filtering framework to refine the shallow decoded features progressively. In addition, to ensure the structural information of the ground objects to the maximum extent, we propose a local detail enhancement (LDE) module based on multi-scale convolutional architectures at the low-level encoding stage. Comprehensive experiments on the two instance RSIs CD datasets and two public CD datasets, results demonstrate that the proposed PRNet achieves competitive performance with fewer parameters (3.44M). Our code is available at https://github.com/lmetay/PRNet.

Capturing finer and discriminative difference features (DF) is key to obtaining a high-quality change detection (CD) map. However, there is still significant scope for further study on fine-grained detection, especially concerning terms of improving structural integrity and reducing internal holes or sticking in DF. To this end, we propose a progressive difference amplification network (PDANet) with edge sensitivity to detect changed areas in optical remote sensing images (RSI), where the key point is to amplify DF and reinforce edge detail to improve CD accuracy. The edge sensitivity encoder (ES) is designed to capture the long-distance dependency, which compensates for the limited receptive fields of the convolutional neural network with fixed kernels. Meanwhile, we introduce the prior edge in the network training stage, which collaborates with the ESE to improve the structural integrity of the changed areas. On the other hand, the difference amplification decoder is proposed to enhance the representation of the changed areas, and it is achieved by integrating multi-scale DF and reconstructing the original single RSI using DF as full-stage guidance. Finally, the CD map and edge map are predicted based on the reconstructed feature and the maximum scale DF. Extensive experiments on one instance dataset and three CD benchmark datasets demonstrate that PDANet outperforms the state-of-the-art CD competitors both qualitatively and quantitatively.

Optical Remote sensing images (RSIs) are used in surface observation, and one of the most interesting research topics is change detection (CD). The internal problem of RSIs, including multi-scales changed objects and cluttered background, still deserve attention. Existing methods make great efforts to solve this problem but inevitably miss detection, which affects the model performance. To address this dilemma, this article proposes a Relation-aware Semantic Reasoning Network (RaSRNet) in an end-to-end manner to pop-out change objects in RSIs, where the key point is to perceive contextual semantic information. The relation-aware module in RaSRNet combats the lack of contextual information caused by the limited receptive field of the general convolutional layer, which facilitates all-around changed object detection. The multi-level semantic reasoning encoder-decoder backbone in RaSRNet extracts and reconstructs pixel semantic information, alleviates the interference of background noise and improves the integrity recognition of changed objects. In addition, the decoder backend undertakes two semantic segmentation branches, and introduces a semantic reasoning loss between the two branches to infer pixel semantic categories, which provides more accurate semantic features for the CD. Extensive experiments are conducted on the three public RSIs CD datasets, and the results demonstrate that the proposed RaSR-Net can accurately locate changed objects, which consistently outperforms the state-of-the-art CD competitors.