A physics-informed spectral-spatial unfolding network with fusion perception for maize spectral recovery from RGB images

Abstract

Given the substantial agronomic and economic significance of maize, the development of real-time and high-precision disease detection methodologies is essential for ensuring yield stability. While hyperspectral imaging excels at capturing fine-grained spectral signatures of infection, its higher detection precision comes with considerable high hardware and temporal costs compared to RGB imaging, posing significant challenges for scalable field applications. To bridge this gap, this article proposes a fusion perception unfolding network (FPUF-Net) for high-fidelity maize spectral reconstruction from RGB images. Distinct from conventional deep learning models, FPUF-Net unfolds the optimization problem via a half-quadratic splitting algorithm, solving the data subproblem and prior subproblem alternately during iterations. Specifically, a fusion feature learning network and a spectral-spatial joint attention network are designed within the data subproblem to explicitly exploit RGB spatial priors and mitigate spatial smoothing. Moreover, a spectral-spatial Transformer is utilized as the denoiser to capture long-range spectral dependencies in the prior subproblem. Experiments performed on a maize spectral recovery dataset comprehensively demonstrate that FPUF-Net can effectively reconstruct maize hyperspectral images with superior precision. The structural characteristics enable the network to perceive long-term spectral-spatial fusion features, significantly reducing reconstruction errors, particularly in the biologically critical red-edge region (620-700 nm). In downstream disease detection tasks, the overall accuracy of reconstructed HSIs improves over RGB by margins of 0.51% to 8.1% across different scenarios, while the average accuracy increases by 2.45% to 18.86%. These results indicate that the proposed model offers a viable, cost-effective solution for applying hyperspectral imaging in field settings, enabling its scalable use in agricultural robots.

Plant phenotyping relevance

RGB画像からトウモロコシのハイパースペクトル情報を再構成する手法を開発し、データセットで性能評価している。植物のスペクトル状態および病害検出に直接関わる手法が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.