Hyperspectral–Region Aggregation Network for Maize Leaf Nitrogen Content Estimation via Spectral–Regional Joint Modeling

Abstract

Abstract Accurate estimation of maize leaf nitrogen content is important for improving nitrogen-use efficiency and supporting precision crop management. However, leaf-level hyperspectral modeling is challenged by high spectral redundancy and heterogeneous spectral responses among local leaf regions. This study proposes a Hyperspectral–Region Aggregation Network (HSRAN) for maize leaf nitrogen content estimation from region-level hyperspectral spectra. HSRAN consists of a Spectral Adaptive Recalibration Encoder (SARE) and a Context-Aware Gated Aggregation Module (CAGM). SARE performs band-wise residual recalibration and extracts regional spectral representations, whereas CAGM models contextual dependencies among regional features and performs gated attention-based aggregation for leaf-level prediction. Field experiments were conducted in 2024 and 2025 at the jointing, silking, and maturity stages. HSRAN was evaluated against PLSR, RF, XGBoost, SVR, 1D-CNN, MLP, and Transformer1D models. Across the stage-specific and pooled datasets, HSRAN achieved the highest R² and the lowest RMSE while maintaining competitive MAE values. On the pooled full-growth-period dataset, HSRAN achieved an R² of 0.84, an RMSE of 3.63 g kg⁻¹, and an MAE of 2.59 g kg⁻¹. At the jointing, silking, and maturity stages, the corresponding R² values were 0.56, 0.76, and 0.72, respectively. Ablation experiments indicated that integrating SARE and CAGM improved R² from 0.80 to 0.84. To interpret regional contributions, the learned attention weights were mapped back to the original leaf coordinates recorded during regional sampling. Regions near leaf veins, tips, and margins often received relatively higher attention weights, suggesting that their local spectra provided informative cues for model prediction. These findings indicate that spectral–regional joint modeling can improve leaf-level hyperspectral estimation of maize nitrogen content. HSRAN provides a practical framework for non-destructive nitrogen assessment in maize.

Plant phenotyping relevance

トウモロコシ葉の窒素含量という植物形質を非破壊推定するためのハイパースペクトル深層学習手法を開発し、複数手法との比較・アブレーション検証を行っており、フェノタイピング手法が中心である。

Code and data availability

The paper describes maize leaf hyperspectral imaging and nitrogen measurements and the HSRAN model, but no public dataset, code, or model deposit is provided. The data availability statement only offers access from the corresponding author on reasonable request, so any qualifying asset would require contacting the.

No evidence-backed public reproduction asset is currently recorded.