14 January 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
In response to the challenge of image degradation caused by strong sea breezes during drone surveillance over tidal flats, conventional methodologies have predominantly employed iterative up-sampling and down-sampling techniques to augment the receptive field of the network. However, this approach is prone to the loss of critical texture data within the tidal flat imagery throughout the sampling process. To mitigate these issues and enhance the recovery of sharp imagery from blurred inputs, we introduce a novel deep learning architecture based on traditional physical models. Our network structure mainly consists of two parts: Initially, applying wavelet transform to the input images, the extracted high-frequency components are refined using a combination of Bayesian adaptive thresholding and hard thresholding. This process not only ensures high fidelity of the high-frequency information but also contributes to generating tidal flat images with clearer texture details. Subsequently, given the relatively large scope of images captured by drones, there is an increased emphasis on the importance of contextual information during the feature extraction process. To this end, we have applied dilated convolution modules with varying dilation rates to the low-frequency components. This design enables the network to capture image features at different scales, enhancing the model's understanding of the tidal flat scene context and improving its feature expression capability. This allows the model to more accurately identify and extract blurred areas, thereby improving the deblurring effect. Additionally, we have incorporated a loss function based on the wavelet transform. This function guides the model to recover clear details from blurred images by minimizing the differences between the high-frequency and low-frequency components of the original clear image and those of the deblurred image. In the quantitative assessment of the real tidal flat image dataset, we observed that the algorithm has a parameter volume of 7.8M and has achieved significant performance improvement: the Peak Signal-to-Noise Ratio (PSNR) reached 33.11, and the Structural Similarity Index (SSIM) reached 0.7909. The enhancement of these metrics indicates that the algorithm excels in the recovery of image texture details while maintaining a compact parameter count. The optimized parameter configuration not only improves the algorithm's operational efficiency but also simplifies the deployment and training process of the model, making it more suitable for tidal flat scenarios.
20 June 2025 in Neurocomputing
14 November 2025 in Measurement Science and Technology
The significant advancement of modern remote sensing technologies enables the collection of multispectral, hyperspectral, and spatio-temporal data. In earth observation, accurate delineation of features such as shorelines and intertidal zones is crucial for quantitative measurements, involving area estimation and long-term environmental monitoring. However, semantic segmentation (SS) of optical remote sensing images remains challenging because of high intra-class variability, subtle inter-class differences, and complex backgrounds. Thus, an enhanced SS method based on an optimized U-Net architecture was established in our study to address these issues. Specifically, a hybrid feature extraction module was introduced in the encoder to strengthen shallow feature representation. Meanwhile, a multiscale spatial attention module was embedded in the skip connections to adaptively capture spatial dependencies across scales and improve feature fusion. Experiments on the WHDLD and Potsdam datasets demonstrate the effectiveness of the method. This method achieved mIoU scores of 61.33% and 77.47%, with pixel accuracies of 88.02% and 86.72%. Furthermore, the results on the Waterseg dataset confirm its robustness and relevance for accurate geospatial measurements in diverse environments.
1 June 2026 in Measurement