Introduction Early-stage detection and classification of lettuce heat responses are essential for non-destructive phenotyping, yet conventional assessment mainly relies on visible symptoms and manual observation. Methods This study constructed a lettuce hyperspectral dataset comprising heat-sensitive and heat-tolerant varieties under control and high-temperature treatments, and proposed the Dynamic Selective Peak Transformer (DSPformer). DSPformer integrates edge-enhanced feature extraction, dynamic multi-scale spatial-spectral representation, Peak-k selective attention, and a confusion-aware dynamic focal loss to enhance discriminative features while reducing spectral redundancy, class imbalance, and inter-class confusion. Results Under the patch-level evaluation protocol, DSPformer achieved 96.22% accuracy, 95.55% recall, 96.35% precision, and 95.95% F1-score, outperforming the compared CNN- and Transformer-based models. Day-wise evaluation showed that DSPformer reached 82.61% accuracy on Day 1 and 96.55% on Day 3, before visible heat-stress symptoms appeared on Day 6. Under a plant-level partition protocol, DSPformer maintained robust performance with 93.76 +/- 0.49% accuracy. Additional evaluation on the Indian Pines benchmark further demonstrated the applicability of DSPformer to general hyperspectral image classification. Discussion These findings suggest that hyperspectral imaging can capture heat-stress-sensitive information beyond visual phenotypes, and that DSPformer provides a promising framework for early, non-destructive lettuce heat-response screening and hyperspectral phenotyping-assisted breeding.
Hyperspectral imaging (HSI) offers nondestructive assessment of fish freshness by detecting biochemical alterations across spectral bands. However, conventional deep learning approaches do not fully address the particular characteristics of HSI data, such as spectral dominance over spatial textures, ordinal label structure, and a small number of training samples. We propose SGNet (Spectral-Grouped Network), a lightweight architecture that separates spectral and spatial feature extraction using grouped convolutions and a depthwise spatial pathway. A dual attention mechanism that couples channel-wise squeeze-and-excitation with spatial gating adaptively highlights informative features. SGNet achieves 97.8% classification accuracy and 0.64 days mean absolute error (MAE) with just 4.75M parameters when tested on our newly developed 16-day refrigerator-stored salmon fillet dataset. Ablation studies validate the contribution of each component, while comparisons demonstrate a five- to eighteen-fold parameter reduction relative to ResNet-50 and Vision Transformers. Our findings indicate that domain-aware design supports precise, real-time freshness prediction for industrial implementation.
Kazi Nabiul Alam, P. B. Zadeh, Akbar Sheikh-Akbari· 0 citations
This study develops an online grading model based on hyperspectral imaging and a lightweight convolutional neural network that provides a feasible technical pathway for intercepting mycotoxin contamination prior to grain storage.
Lin Tang· International Conference on...· 0 citations
The newly introduced Mamba architecture exhibits superior performance in the hyperspectral image (HSI) classification domain, as it achieves long-range dependence modeling with linear complexity and is widely regarded as a promising alternative to the transformer. However, existing Mamba-based HSI classification methods still face two major limitations: first, fixed multidirectional scanning paths lack content adaptability and introduce computational redundancy; second, the sequential scanning process disrupts intrinsic 2-D local structures, impairing fine-grained spatial features. To overcome these issues, we propose a novel dynamic global–local selection network (DGLSN), which synergistically integrates Mamba and a convolutional neural network within an input-adaptive multibranch framework. Our approach introduces two core modules: a global dynamic selection module that employs a lightweight decision network to activate only the most relevant Mamba blocks for efficient long-range modeling, and a local dynamic selection module that dynamically selects among spatial-, spectral-, and frequency-domain convolutional branches to extract discriminative multiview features. Extensive experiments demonstrate that DGLSN achieves state-of-the-art classification accuracy while requiring competitive training and testing times and significantly lower floating point operations, highlighting its superiority in both performance and computational efficiency.
Zhe Meng, Taizheng Zhang, Feng Zhao et al.· IEEE Journal of Selected Top...· 0 citations
Abstract. Hyperspectral image (HSI) classification is one of the core tasks in the field of remote sensing, whose key lies in the effective fusion of spectral and spatial information. Among existing methods, convolutional neural networks (CNNs) are limited by their local receptive field, making it difficult to model long-range spectral dependencies, while Transformers, although capable of capturing global relationships, suffer from high quadratic computational complexity. To address these issues, this paper proposes a Multi-scale Cyclic Adaptive Mamba Network (MCAM) based on state-space models (SSM) for hyperspectral image classification. First, a multi-scale feature convolution block is introduced to extract spatial features from local to global levels in parallel, thereby enhancing feature representation. Subsequently, a cyclic adaptive scan module is incorporated to strengthen the modeling of long-range spectral–spatial dependencies. Furthermore, a combination of triplet loss and classification loss is adopted to improve the model’s discriminative ability in few-shot learning scenarios. Experiments conducted on the Indian Pines and Liao Ning-01 datasets demonstrate that MCAM outperforms existing mainstream methods in terms of overall accuracy (OA), average accuracy (AA), and Kappa coefficient, particularly excelling in class boundary clarity and spatial consistency. This study validates the efficiency and potential of the Mamba architecture in HSI classification and provides new insights for subsequent related research.
Yihang Zou, Lina Xu, Yanni Dong· The International Archives o...· 0 citations
Hyperspectral image (HSI) classification has been widely applied in numerous fields. Although deep learning-based methods have improved classification performance, existing approaches still struggle to balance accuracy and computational efficiency. Convolutional neural network (CNN)-based methods are limited by local receptive fields, whereas Transformer-based methods suffer from high computational complexity, restricting their application in large-scale scenarios. To achieve collaborative optimization of classification accuracy and computational efficiency, a dual-branch spectral–spatial Mamba (DBSSM) network is proposed, which employs a dual-branch architecture with an interactive feature fusion strategy for efficient joint spectral–spatial modeling. The main contributions are as follows: 1) a spatial dual-scan Mamba (SDSM) module is developed to extract rich sequential features via a dual-directional scanning strategy and shared mechanism, enabling long-range dependency modeling with relatively low computational complexity; 2) a spectral group attention (SGA) module is designed to reduce the number of parameters via a grouping strategy while enhancing spectral information interaction using a self-attention mechanism; and 3) an interactive attention fusion module (IAFM) is proposed to achieve bidirectional interactive fusion between spatial and spectral features, thereby further enhancing the feature learning capability. Experimental results on four benchmark hyperspectral datasets demonstrate that DBSSM outperforms state-of-the-art methods in classification performance while maintaining a low parameter count and computational cost, validating its effectiveness. The code is available at https://github.com/Present-Li/DBSSM
Hanzhong Li, Hua Huang, Hongfeng Li et al.· IEEE Transactions on Geoscie...· 0 citations
A lightweight deep learning model, the Dual Multi-Layer Perceptron Feature Fusion Network (DMLPFFN) as a hyperspectral-based approach to the four tomato disease stages, which includes healthy, asymptomatic, early and late infection is demonstrated.
Poonam Chaudhary, Sneha Kandacharam· Indian Journal of Agricultur...· 0 citations