Curvelet Decomposition-Based Tri-Branch Coupling Network for Hyperspectral Unsound Maize Seeds Identification.

Abstract

The rapid and nondestructive classification of maize kernels is of great significance for seed screening and quality evaluation. Existing hyperspectral image classification methods based on the Mamba architecture can effectively represent spectral and spatial features; however, they still face limitations in time-frequency analysis and multimodal feature fusion. In addition, traditional approaches often rely heavily on spectral preprocessing, which may introduce additional errors and compromise the model's robustness and generalization ability. To address these challenges, this paper proposes a novel cross-modal classification framework named CD-TriMamba, which jointly leverages hyperspectral data and visible-light images for comprehensive feature extraction and deep fusion. Specifically, an innovative feature extraction module is designed, consisting of a Spectral Curvelet Convolution (SCC) module for hyperspectral data and a Curvelet-Decomposed Convolution (CDC) module for spatial modeling. A feature rearrangement mechanism is further introduced to mine critical information from both spectral and spatial modalities. Finally, a ConvNeXt-guided tri-branch cross-fusion structure (TriMamba) is constructed to achieve deep collaboration and efficient integration between spectral and spatial features. Experimental results demonstrate that the proposed model achieves outstanding performance in seed classification, with an accuracy (Acc) of 99.2% and a Kappa value of 99.1%. These results strongly confirm the effectiveness and broad application potential of cross-modal feature fusion in maize kernel classification.

Plant phenotyping relevance

マルチモーダル画像からトウモロコシ種子の健全性を推定する新規分類フレームワークを開発しており、種子状態の取得・抽出手法が研究の中心である。

Code and data availability

The supplied article blocks describe the CD-TriMamba hyperspectral maize seed classification method but contain no data availability statement, no public dataset, image, code, or model deposit, and no author-provided URLs. No paper-specific public asset can be identified.

No evidence-backed public reproduction asset is currently recorded.