Existing mainstream hyperspectral image (HSI) classification frameworks, like multilayer perceptron (MLP) networks, combine learnable linear transformations with nonlearnable nonlinear activation functions. While these feature extraction algorithms achieve high classification accuracy, they require significant computational and memory resources, and many models lack adequate informative representations. Recently, fast Kolmogorov-Arnold networks (FKANs) have emerged as a lightweight alternative to MLPs, offering competitive performance with fewer parameters and lower computational costs. This paper presents a hierarchical progressive fusion network (HPFN) based on FKAN to address these challenges. The proposed network explores HSI information more comprehensively through branches at three scale levels (pixel, patch, global) while remaining lightweight. The orthogonality of information at different levels is preserved to some extent for diverse practical applications. We develop two FKAN based lightweight convolutional networks to extract pixel-level and patch-level features. In the global-level branch, a boundary-enhanced U-Net with an edge-constrained module based on Sobel operators mitigates over-smoothing. The cross-scale guided fusion mechanism allows for dynamic awareness and accurate pairing of pixel-level features with local and global features, reducing information redundancy. Experimental results show that the proposed network achieves competitive classification accuracy on four public datasets at low computational cost, validating its effectiveness.

Object detection plays a crucial role in remote sensing due to the urgent demands of various applications, such as urban planning and environmental monitoring. Despite notable progress, current methods still struggle with detecting challenging small objects. At the object level, the limited pixel representation, blurred details, and background interference of small objects impose greater demands on feature extractors. At the network level, resource bias fails to provide adequate learning signals for these objects. In this paper, we propose a Sparse Transformer-based detector (STDet) to tackle these challenges. Specifically, we design a Local-to-Global Transformer network (LGFormer) to explore essential feature representations. The Local Transformer Block establishes correlations between tokens and their surrounding data, while the Global Transformer Block captures long-distance dependencies related to the objects. Meanwhile, we introduce a Scale-Balanced Label Assignment (SBLA) strategy that considers more samples to small objects. SBLA dynamically shifts the learning focus to easily overlooked objects and alleviates the issue of sample imbalance. Extensive experiments on three large-scale remote sensing datasets demonstrate the effectiveness of STDet and its superiority in small object detection.

Recently, remote sensing image object detection based on convolutional neural networks (CNNs) has made significant advancements. However, small objects detection remains a major challenge in this field. Because the small size of the object makes it difficult to extract their features and these features are further weakened after downsampling in the network. In order to improve the detection accuracy of small objects in remote sensing images, this letter provides a feature super-resolution fusion framework based on cross-scale distillation. Specifically, we design a sub-pixel super-resolution feature pyramid network (SSRFPN) replacing the bilinear interpolation with sub-pixel super-resolution (SSR) modules to enhance the feature expression capability. Furthermore, we propose a cross-scale distillation (CSD) mechanism to guide the SSR modules in learning the features of small object regions more accurately. Finally, our method is applied to three detectors on two datasets for validation. We adopt YOLOv7 as the baseline model and achieve the best results, with the average precision at a threshold of 0.5 (AP0.5) of 95.0% and 82.3% on the NWPU VHR-10 dateset and DIOR dataset. And the mean average precision of small objects (mAPS) is improved by 8.5% and 2.5%.

Hyperspectral image super-resolution technology has made remarkable progress due to the development of deep learning. However, the technique still faces two challenges, i.e., the imbalance between spectral and spatial information extraction, and the parameter deviation and high computational effort associated with 3D convolution. In this article, we propose a super-resolution method for remote sensing hyperspectral images based on multi-domain spatial information and multi-scale spectral information fusion (MSSR). Specifically, inspired by the high degree of self-similarity of remote sensing hyperspectral images, a spatial-spectral attention module based on dilated convolution (DSSA) for capturing global spatial information is proposed. The extraction of local spatial information is then accomplished by residual blocks using small-size convolution kernels. Meanwhile, we propose the 3D Inception module to efficiently mine multi-scale spectral information. The module only retains the scale of the 3D convolution kernel in spectral dimension, which greatly reduces the high computational cost caused by 3D convolution. Comparative experimental results on four benchmark datasets demonstrate that compared with the current cutting-edge models, our method achieves state-of-the-art results and the model computation is greatly reduced.

Graph convolutional networks (GCNs) can extract features of samples in non-Euclidean space, which can be used for hyperspectral image (HSI) classification in collaboration with convolutional neural networks (CNNs). The features of GCNs and CNNs are incompatible to a certain extent, and traditional graph convolution methods use a single channel two-dimensional matrix to extract features. As a result, it is difficult to explore the relationships fully and flexibly between samples. To further exploit the potential of these two networks for collaborative extraction of HSI features, we propose a fusion framework based on multichannel GCNs and class-guided attention mechanism (MG2A). Specifically, a multichannel graph convolutional network (MGCN) module is designed for batchwise network training, where the adjacent matrix of each channel contains different information between samples. In addition, we develop a class-guided attention mechanism to adaptively fuse the features of multiple MGCN modules and learn the transformation process of the features. Finally, the features of CNNs and MGCN modules are fused at multiple layers through a fusion framework. Experimental results on four benchmark HSI datasets show that MG2A achieves better classification performance compared to other state-of-the-art methods.