Toward generalized peach firmness evaluation: An adaptive spectral-vibration fusion and cross-cultivar transfer...

Fruit firmness is a key indicator of peach quality. Visible/near-infrared spectroscopy and acoustic vibration characterize firmness from chemical and mechanical perspectives, respectively. This study proposes a multimodal fusion and constrained transfer learning approach to combine these modalities for cross-cultivar peach firmness prediction. A dual-input Zero-Init Attention Fusion Network (ZIAF-Net) based on a multilayer perceptron was developed. It incorporates a Zero-Init Residual Attention Block for adaptive feature fusion. Attention intensity is controlled by a learnable scalar, alleviating inter-modal interference during the early stages of model training with limited samples. For cross-cultivar adaptation, a Parameter-Anchored Consistency Transfer (PAC-Transfer) strategy was also introduced. It integrates Anchor Loss to constrain weight deviations from the source parameters and Consistency Loss to reduce the imbalance of prediction errors across different firmness levels, thereby mitigating catastrophic forgetting and overfitting. ZIAF-Net was trained on the 'Baifeng' dataset and transferred to the 'Wanhujing' and 'Jinqiuhong' datasets. On the source domain, ZIAF-Net achieved an R_P² and an RMSEP of 0.8332 and 0.3762 N/mm, outperforming baseline models. In cross-cultivar transfer, PAC-Transfer achieved an R_P² of 0.7510 on the 'Wanhujing' dataset and 0.7388 on the 'Jinqiuhong' dataset, improving upon standard full fine-tuning by 19.5% and 23.2%, respectively. Ablation study results demonstrated that combining Anchor Loss and Consistency Loss outperforms the use of either loss function alone. Gradient attribution analysis revealed the key spectral bands and vibration frequency bands underlying the model decisions, confirming the complementarity of the two modalities for peach firmness evaluation.