Search papers, labs, and topics across Lattice.
The paper introduces HyPyraMamba, a novel architecture for hyperspectral image (HSI) classification that addresses limitations of CNNs and Transformers by integrating pyramid spectral attention (PSA) and Mamba-based sequence modeling. HyPyraMamba uses PSA to capture multiscale spectral features, adaptive expert depthwise convolution (AEDC) to enhance spatial-spectral feature expression, and spatial/spectral Mamba branches for improved spatial structure and spectral correlation modeling. Experiments on four benchmark HSI datasets demonstrate that HyPyraMamba outperforms state-of-the-art methods, particularly in reducing confusion between spectrally similar land-cover categories, while providing a favorable accuracy-efficiency tradeoff.
HyPyraMamba slashes confusion between spectrally similar land-cover types in hyperspectral imagery, outperforming CNNs, Transformers, and even vanilla Mamba.
In hyperspectral image (HSI) classification, the high-dimensionality and the complex coupling of spatial–spectral features present severe challenges to existing deep learning methods in terms of accuracy, generalization, and computational efficiency. Researchers have recently explored convolutional neural network (CNN) and Transformer-based methods to overcome these limitations, but CNN’s limited receptive field prevents effective modeling of long-range dependencies, while Transformers suffer from high computational cost and inefficiency in high-dimensional data. Motivated by these limitations, the state-space model (SSM) Mamba shows great potential as an efficient alternative for sequence and dependence modeling. Building on this foundation, we propose HyPyraMamba, a novel architecture designed to effectively overcome the above challenges. It integrates the pyramid spectral attention (PSA) module to capture multiscale key spectral features, thereby reducing interference caused by spectral redundancy. We developed an adaptive expert depthwise convolution (AEDC) module that enhances the model’s ability to express multiscale spatial–spectral features, and a sequence modeling module, Mamba. In the Mamba module, we utilize the spatial Mamba and spectral Mamba branches to enhance spatial structure and spectral correlation modeling. Extensive experiments on four benchmark HSI datasets demonstrate that HyPyraMamba significantly outperforms several recent state-of-the-art methods and provides a favorable accuracy–efficiency tradeoff. In particular, classwise analyses on spectrally similar land-cover categories (e.g., different soybean and bareland types) show that HyPyraMamba markedly reduces mutual confusion compared with CNN-, Transformer-, and Mamba-based baselines. The code will be available at https://github.com/dekai-li/HyPyraMamba
