6 March 2023 in IEEE Transactions on Geoscience and Remote Sensing
Hyperspectral image (HSI) classification based on neural architecture search (NAS) is a currently attractive frontier as it not only automatically searches complex neural network architecture, but also avoids professional knowledge and experience design, and alleviates the lacking of generalization ability as well when dealing with a new classification task. However, the existing HSI classification based on NAS has some drawbacks: 1) A huge number of training parameters and high calculations are inductive to over-fitting and high complexity. 2) Efficient operators are lacking in the search space which can distinguish spatial locations and spectral features in different bands. Furthermore, as the category samples in HSI data show a serious long-tail distribution phenomenon, HSI classification remains challenging. To address these issues, we propose a lightweight HSI classification model LMSS-NAS integrating multi-scale spectral-spatial attention. The main work includes three-fold: 1) In order to reduce the number of model parameters and promote spectral-spatial feature fusion, a new lightweight efficient search space is designed, which consists of three equivalent lightweight convolution operators with multiple receptive fields. 2) To fully use the spectral-spatial correlation of HSI, a cube-to-pixel classification framework is designed to mine the local spatial and spectral context. 3) Focal loss and label smoothing loss in computer vision tasks are jointly migrated to LMSS-NAS to improve the unbalanced samples’ classification and model robustness. Experimental results on four public hyperspectral data sets show that the proposed method can achieve competitive classification performance as well as low computational cost. Code is available at: https://github.com/xh-captain/LMSS-NAS.
5 April 2024 in IEEE Transactions on Geoscience and Remote Sensing
Few-shot learning (FSL) has achieved promising performance in hyperspectral image classification (HSIC) with few labeled samples by designing a proper embedding feature extractor. However, the performance of embedding feature extractors relies on the design of efficient deep convolutional neural network architectures, which heavily depends on the expertise knowledge. Particularly, FSL requires extracting discriminative features effectively across different domains, which makes the construction even more challenging. In this paper, we propose a novel neural architecture search-based FSL model for HSI classification, called HCFSL-NAS. Three novel strategies are proposed in this work. First, a neural architecture search-based embedding feature extractor is developed to the FSL in HSIC, whose search space includes a group of proposed multi-scale convolutions with channel attention. Second, a multi-source learning framework is employed to aggregate abundant heterogeneous and homogeneous source data, which enables the powerful generalization of network to the HSIC with only few labeled samples. Finally, the pointwise-based cross-entropy loss and the pairwise-based adaptive sparse loss are jointly optimized to maximize inter-class distance and minimize the distance within a class simultaneously. Experimental results on four publicly hyperspectral data sets demonstrate that HCFSL-NAS outperforms both the exiting FSL methods and supervised learning methods for HSI classification with only few labeled samples. Code is available at: https://github.com/xh-captain/HCFSL-NAS.
23 February 2024 in IEEE Transactions on Computational Imaging
Deep learning-based hyperspectral unmixing is getting increasing attention in the field of remote sensing, aiming at endmember extraction and abundance estimation at pixel scale. However, many existing deep learning-based unmixing methods are generally based on linear mixing model which ignore the complex nonlinear interactions of the light scattered. Furthermore, these methods often treat all spectral bands indiscriminately, failing to capture the characteristic difference between inter-class endmembers which is very unfavorable for endmember separation. To address these issues, we present BU-Net, a novel approach for hyperspectral unmixing based on the generalized bilinear mixing model (GBM), which is a two-stream stacked autoencoder architecture designed to enhance inter-class separability. In the encoder, we employ 3D convolutions with multiple receptive field to extract multiscale spatial and spectral features simultaneously. Addtionally, we design a novel band selection based on inter-class separability (BSICS), which identifies bands with inter-class separability (BICS) and the obtained bands are taken as an additional stream for improving performance. In the decoder, BU-Net develops a two-stream structure consisting of linear and bilinear components. This design matches the components and constraints in GBM theoretically. To further enhance separability between endmembers during training, we use the spectral angle distance between BICS and its reconstruction as a loss regularization term. Moreover, we utilize materials' representative pixels obtained in the process of BSICS to initialize endmembers, which offers effective guidance for modeling the spectral properties, ultimately leading to improved unmixing results. Experimental results on synthetic and real hyperspectral datasets show that our method outperforms state-of-the-art methods. This novel approach not only addresses the limitations of linear mixing models but also leverages the power of deep learning to enhance the accuracy of endmember extraction and abundance estimation in remote sensing applications. The codes are available at the website: https://github.com/1songwei/BU-Net .
8 February 2024 in IEEE Transactions on Geoscience and Remote Sensing
Hyperspectral images (HSI) have spectral variability, which leads to spectral dependence in adjacent and non-adjacent regions, and this dependence is essential for the classification of regions with mixed pixels. Current neural architecture search (NAS) methods have achieved significant advantages in HSI classification, but these methods cannot capture spectral dependence in non-adjacent regions because only use feedforward connections. Meanwhile, the cost of the search process in NAS is proportional to the scale of the search space, which limits the expansion of the search space. To address these issues, we propose a sparse-coding-inspired feedback neural architecture search (SCIF-NAS) method for HSI classification. Firstly, we view HSI samples as sequences and introduce a feedback mechanism in NAS to model the spectral dependence of non-adjacent regions to mitigate the effects of spectral variation. Secondly, we design several feedforward operations according to the characteristics of HSI, to form the search space together with feedback operations. Meanwhile, a sparse-coding-inspired NAS accelerated strategy is introduced to alleviate the search time burden caused by the expansion of search space. Thirdly, we integrate center loss with cross-entropy loss to construct a hybrid loss function that helps to obtain a better classification boundary. Finally, we conduct experiments on three popular HSI benchmarks, which show that SCIF-NAS outperforms the state-of-the-art methods in HSI classification.
2 May 2026 in Pattern Recognition Letters
20 October 2025 in IEEE Geoscience and Remote Sensing Letters
Road extraction from remote sensing images is inherently challenging due to spatial-spectral dual complexities. Spatial complexity stems from irregular and diverse topological structures and large-scale variations, whereas spectral complexity arises from clouds, shadows, and surface reflections. Existing spatial-domain methods excel at modeling geometric structures but often fail under severe spectral interference, while emerging frequency-domain approaches provide enhanced discriminability yet insufficiently capture road geometry. To fully exploit their complementarity, this paper proposes a frequency domain enhancement network (FDENet), which improves road discrimination and geometric modeling through spatial-frequency joint design and complementary fusion. First, a CNN-Transformer hybrid backbone named TransXNet is adopted to capture rich multi-scale spatial features. Then, a frequency enhancement module (FEM) based on strip convolution is proposed to differentially process wavelet subbands, with large kernels modeling low-frequency geometry and direction-aware kernels sharpening high-frequency edges. Finally, an adaptive fusion module (AFM) is developed to bridge the spatial-frequency representation gaps through adjusting channel weights of fusion dynamically. Experiments on DeepGlobe and Jilin-1 show that FDENet outperforms six state-of-the-art methods in F1 and IoU with fewer parameters. Code is available at https://github.com/oybk8/FDENet.